Abstract
Affective states are key determinants of animal welfare. Assessing such states under field conditions is thus an important goal in animal welfare science. The rapid Defence Cascade (DC) response (startle, freeze) to sudden unexpected stimuli is a potential indicator of animal affect; humans and rodents in negative affective states often show potentiated startle magnitude and freeze duration. To be a practical field welfare indicator, quick and easy measurement is necessary. Here we evaluate whether DC responses can be quantified in pigs using computer vision. 280 video clips of induced DC responses made by 12 pigs were analysed by eye to provide ‘ground truth’ measures of startle magnitude and freeze duration which were also estimated by (i) sparse feature tracking computer vision image analysis of 200 Hz video, (ii) load platform, (iii) Kinect depth camera, and (iv) Kinematic data. Image analysis data strongly predicted ground truth measures and were strongly positively correlated with these and all other estimates of DC responses. Characteristics of the DC-inducing stimulus, pig orientation relative to it, and ‘relaxed-tense’ pig behaviour prior to it moderated DC responses. Computer vision image analysis thus offers a practical approach to measuring pig DC responses, and potentially pig affect and welfare, under field conditions.
Similar content being viewed by others
Introduction
Objective measurement of animal welfare under field conditions is an important goal in animal welfare science. Progress has been particularly strong in on-farm welfare assessment which initially focused on cataloguing and quantifying the resources provided to animals (‘inputs’: e.g. trough space, drinker access, lying substrate, ventilation etc.1) and used this information to make inferences about welfare. More recently, attention has widened to include measurements of the animals themselves (‘outputs’: e.g. lameness, indicators of injury or disease, abnormal behaviour2,3,4) in order to get more direct evidence about their welfare, as in EU Welfare Quality on-farm welfare assessment protocols5,6,7.
Because concerns about animal welfare are, for many, based on an assumption that non-human animals can experience negative affective (emotional) states and therefore suffer8,9,10,11, there is a particular need to develop validated indicators of animal affect that can be implemented quickly and easily. To this end, a variety of behavioural measures have been developed for use under farm conditions (e.g. human approach tests, novelty tests, Qualitative Behavioural Assessment12,13,14,15). Some have been systematically validated against other indicators or by using manipulations of affective state grounded in a clearly-argued rationale, others less so12,16. In many cases, a major barrier to uptake is the time required to collect and interpret the relevant data17,18,19. Consequently, there is still a need for new validated and reliable indicators of animal affect and welfare that are easy to implement under field conditions.
One potentially promising measure is the ‘Defence Cascade’ (DC) response shown to sudden, unexpected stimuli20,21,22,23. The DC is an adaptive suite of responses evolved to ensure appropriate detection, evaluation and response to alerting stimuli, and conserved across species. It involves initial detection and immediate response (startle) to a stimulus such as an unexpected noise20,21,22, monitoring and evaluation of the stimulus accompanied by freezing/immobility22,24, and a final response of either defensive/escape behaviour or resumption of previous activities. Critically, components of the DC response are modulated by affective state in humans and rodents, making them potentially valuable indicators of these states.
The startle component, a chained series of rapid flexor reflex movements that cascade through the body and serve a protective function (e.g. eye blink), acts as a behavioural interrupt and enhances vigilance by switching attention towards potential threats20,21,25. In humans and rodents, negative states appear to potentiate startle magnitude. For the human eye blink, electromyography is usually used to quantify the strength of contraction of the orbicularis oculi muscle that closes the eye. For rodents, the magnitude of movements such as jumping are recorded using a load platform. For example, humans show larger startle responses when viewing or anticipating unpleasant pictures26,27,28,29,30 or negative emotional expressions31, anticipating shock32, in the presence of odours previously paired with a stressful social experience33, and even when exposed to sweat odours from anxious donors34. In rodents, many studies demonstrate potentiated startle in subjects exposed to conditioned stimuli predicting an aversive event25,28,35,36,37,38, an effect that may be mediated by corticotropin-releasing hormone39 and attenuated by anxiolytic drugs (e.g. diazepam37). Startle-potentiating effects of aversive conditioned stimuli are also observed in monkeys40.
Conversely, startle magnitude is often attenuated in people exposed to pleasant pictures27,29, sounds41, or conditioned visual stimuli paired with the removal of a painful stimulus and hence ‘relief’42, and in rodents exposed to conditioned visual or olfactory stimuli associated with rewards such as food43,44,45,46 and rewarding brain stimulation47.
The links between negative states and startle potentiation and positive states and startle attenuation are not always observed. For example, startle magnitude sometimes increases in people anticipating pleasurable pictures48,49 or monetary reward32,50 (although startle magnitude attenuates following monetary reward50). Bach51 provides a theoretical and computational model which reconciles these findings with those described previously by assuming that forgone opportunities due to a brief startle response are lower than those due to being injured (e.g. by a predator) if one does not startle. The model thus predicts an increased startle, resulting in lower overall costs, when beneficial opportunities are available such as during reward anticipation. Other exceptions include that startle was potentiated when viewing the names of loved ones compared to neutral or famous names52 and that anxiolytics had no effects on fear-potentiated startle53. Likewise, enrichment25 and preferred home cage odour54 did not attenuate startle in rats, and chronic mild stress failed to potentiate startle in mice55.
The freeze component of the DC response may function to allow effective processing of new information about potential danger, and to decrease the likelihood of detection by predators22,56,57. Negative affective states associated with increased anticipation of negative events58 may thus favour prolonged freezing responses that minimise chances of being detected by a predator and reflect a higher threshold of evidence required to judge that the environment is safe. For example, freezing in rodent aversive conditioning studies is frequently used as an indicator of a fear-like state59,60,61 and is potentiated by induced stress62 and attenuated by drugs and other treatments assumed to generate a less negative state63,64,65,66,67. Rats selected for an anxiety and depression-like phenotype show higher levels of freezing in aversive conditioning tests68 and dogs freeze for longer in response to negatively than positively valenced stimuli69. As with startle, there are exceptions to these findings such as the enhancement of unconditioned freezing in appetitive contexts70 and the failure of GABA agonists, assumed to have an anxiolytic function, to modify freezing in mice71.
Variation in startle and freeze components of the DC response thus offer potential as new indicators of affective valence and hence welfare. Such variation is likely to reflect a combination of influences on an animal’s current affective state, including temperamental and environmental factors, and hence can provide a summary readout of how particular individuals with particular experiences and genetic predispositions are being affected by their current situation. Preliminary studies of startle responses have been carried out in some farm animal species (e.g. pigs72,73,74; cattle75; sheep76), and a recent study by Ross et al.77 demonstrated attenuated startle responses in hens living with, relative to without, preferred enrichments. However, a major challenge to the use of DC responses as indicators of affective state and welfare in the field is quick and accurate measurement of these rapid and subtle behavioural patterns. ‘Gold standard’ measures such as direct behavioural observations (e.g. detailed by-eye video analysis) may be manageable for individuals or small experimental groups under lab conditions, but are slow and/or impractical in the field. One possible solution to this challenge is to use computer vision techniques to implement real-time automated analysis of video-recorded DC responses78,79. Successful development of this approach would allow cheap, objective and rapid measurement of DC responses on farm, or in other contexts such as abattoirs, using just a video camera, a standardised eliciting stimulus, and the required software.
Here we explore this possibility in a commercially important farm animal – the pig – by evaluating the validity of computer vision image analysis against ‘ground truth’ data provided by human behavioural observation which has been used successfully in previous studies of pig DC behaviour72,73,74. We also collect force, kinematic, and depth-camera (Kinect) measures for comparison, as these are widely used to assess movement in other species. Finally, we evaluate whether contextual factors influence the DC response. We choose the pig because it shows characteristic DC behaviour when unexpectedly disturbed, usually involving a whole-body startle and movement to a tense standing position, sometimes accompanied by a bark vocalization72, followed by a period of immobility or freezing during which the animal appears to be monitoring or attempting to detect the source of the disturbance. These responses terminate when the pig flees or resumes ongoing behaviour. Successful development of an automated measure of DC responses will open the way for validating variation in startle and freeze behaviour as an indicator of affective valence in pigs and trialling the approach in groups of animals and on farm73,80.
Materials and Methods
Animals and housing
Twelve pigs (Large White x Landrace) were sourced from a commercial farm at approximately seven weeks of age. They were housed in two straw-bedded rooms (4.6 m × 4.6 m), naturally lit and supplemented with artificial light between 0700 and 1900, with a target temperature of 20 °C, and each holding three males and three females. Pigs were fed to appetite twice daily and water was provided ad libitum. They were weighed once a week. Studies were ethically reviewed by the Bristol University Animal Welfare and Ethics Review Body and carried out in accordance with the Animals (Scientific Procedures) Act 1986 (Project Licence: PPL 30/2867).
Test room and equipment
The Test room contained a force-measuring pen, custom-built by Solutions for Research (https://www.solutionsforresearch.co.uk/) and consisting of a 1.31 ×1.33 m load platform fitted with four load cells surrounded by 1 m high walls, three of which were made of 2” wire mesh whilst the fourth had a clear polycarbonate guillotine door through which pigs could enter. Fast capture video cameras (3 monochrome Point Grey Dragonfly Express cameras running at 200 Hz) were positioned on two sides of the force-measuring pen and overhead, with the overhead position normalised by considering the location of the four corners of the pen. A standard video camera was set up for filming from one side of the pen. A Microsoft Kinect v1 camera running at 30 Hz and collecting RGB and depth data was also positioned above the pen. During weeks 4 and 12 of the study, Kinematic data were collected using four infrared Qualisys cameras (ProReflex MCU240, Qualisys AB, Goteborg, Sweden) running at 200 Hz and positioned at the four corners of the pen. The set-up is shown in Fig. S1 in the Supplementary Information which also contains information on data capture and synchronisation.
Habituation to test room
On arrival (week 1 of the study) pigs were given individual ear tags and then left undisturbed (except for cleaning and feeding) for three days. During weeks 1–2 the pigs were gradually habituated to human contact, moving to the Test Room and spending time on the load platform. They were initially introduced to the room in a group of six and food treats were available throughout. Over a period of days, the group sizes were reduced and time periods increased until each pig was spending at least 5 minutes alone on the platform.
Testing
The aim of this study was to generate recordings of a large number of Defence Cascade events and then, using all these events, to compare metrics generated by different measurement techniques in order to establish the validity of computer vision measures of the DC response. To this end pigs were tested repeatedly to generate the dataset of recordings. Testing was carried out when the pigs were relatively young and light (weeks 3–5 of the study; 9–11 weeks of age; 20–40 kg) and subsequently when they approached slaughter weight (weeks 11–13 of the study; 17–19 weeks of age; 50–80 kg). In each phase, each pig received three Standard test sessions and one Kinematic test session. One pig had to be excluded from two sessions due to illness, thus giving 94 sessions in total.
During each Standard test session each pig was taken from its home pen to the Test Room and encouraged to enter the force-measuring pen with food treats. It was then given c.2.5 min to settle before the first test. During this time, the pig’s behaviour was observed live by PS and a ‘relaxed-tense’ score given using the rating scale shown in Table S1. A startling stimulus was then presented and the pig’s response was recorded for 30 s using the load platform, Kinect, fast-capture, and standard video cameras. Sessions comprised a maximum of 5 tests (stimulus presentations), each separated by c.5 min, after which the pig returned to its home pen.
During sessions 1 and 2, the startling stimulus was a remotely-activated bursting balloon next to the force pen. Pigs rapidly habituated to this stimulus and therefore, given the repeated testing required for this experiment in order to yield the dataset required for analysis (see above), it was deemed necessary to vary the stimulus for the remaining sessions to generate sufficient startle/freeze responses for a comparison of automated image analysis and ground-truth measures of these. In session 3 the bursting balloon was combined with a person either calling out or stepping out from behind a curtain. From session 4 onwards the stimulus was varied in each test, the stimuli used included; an unfamiliar researcher stepping out from behind a curtain, a bucket containing metal nuts being dropped, a model person being hoisted above the pig, an umbrella being opened towards them or a bin bag being waved.
During Kinematic test sessions in weeks 4 and 12, pigs were taken from their home pen to a small room where 10 mm passive reflective spherical markers were attached to their backs using double sided sticky tape. One was placed between the ears and three more were spaced down the spine. The final two markers were transverse to each other in the region of the sacroiliac joints; these identified the rear of the pig. Once markers were securely in place the pig was taken to the Test room and the test procedure continued in the same way as for Standard tests with all measures recorded as usual, plus kinematic data recorded for 15 s following each startling stimulus.
In total, 285 tests nested within 94 test sessions (each comprising up to 5 stimulus presentations) were carried out across both phases. Technical issues with some tests mean the final dataset for analysis was 280 tests.
Observations of behaviour from video recordings
Recordings from the standard video camera were coded by-eye using the Observer program81. Each video was examined in slow motion, multiple times to record the behaviour exhibited in response to the stimulus (Jump and fall, Jump away, Jump on spot, Spin to face, Side step, Head up, Head turn, Muscle ripple, Ear prick, Sniff at/approach stimulus, Freeze; see Table S2). Coding continued until an outcome behaviour (Flee, Slowly leave, Return to normal behaviour; see Table S2) occurred which always happened within 30 s of the startling stimulus. An Observer Startle Magnitude score of 0–4 (least to most intense) was generated from the coded behaviour according to a rating scale described in Table 1. A total Observer Freeze Duration (s) was also calculated. Any freezes that lasted less than a second were reviewed and those less than 0.4 s were re-classified as ‘no freeze’ because they did not show the tension/immobility required by our behavioural definition: ‘muscles tensed, whole body stationary, ears often pricked’.
Image analysis (IA) data
All IA measures were derived from the overhead fast capture camera (200 Hz frame rate) as it gave the clearest unobstructed view of the pig. A view from above is also likely to avoid obstructions such as pen walls, posts and other pigs under field conditions. To measure the magnitude of the startle response to the stimulus, accelerations of highly textured image regions were estimated using sparse feature tracking. For freeze durations, both speed and acceleration were extracted. Sparse feature tracking considers only the easiest regions to track and hence improves accuracy compared to dense feature tracking82. Birchfield’s implementation of the Kanade Lucas Tomasi tracker was utilised (KLT; https://cecas.clemson.edu/~stb/klt/), and tracked points are referred to as KLT points. Justification for using points rather than centroids is provided in the Supplementary Information.
The magnitude of the initial response to the startling stimulus (KLT Acceleration Startle Magnitude; acceleration being closely related to force) was defined as the maximum acceleration of 50 KLT points (pixels/frame2) in a temporal window 0.7 s after the startle stimulus (see Supplementary Information). The tracker was always initialized with 150 points. If points were lost by the tracker, new ones were initialized. For our calculations, we only considered the 50 points with the highest acceleration that were successfully tracked for 3 consecutive frames. KLT Speed Freeze Duration was defined as the total time the speed of the 50 fastest KLT points (pixels/frame) was below an empirically determined threshold for a continuous period of at least 0.4 seconds (see above). Thresholds were calculated using 10-fold cross validation with 90% training and 10% testing data repeated 10 times for all data (see Supplementary Information). The same method was used to determine KLT Acceleration Freeze Duration, in case acceleration of the KLT points provided a better predictor of freeze duration than speed of the KLT points.
Load platform data
The four load cells of the platform each generated an analogue signal encoding change in force into change in voltage. This was captured at a rate of 400 Hz, converted into digital form and stored on computer to generate a time series synchronised with the other data and used in all subsequent analyses (see Supplementary Information). Signals from the four transducers were summed to give a measure of total instantaneous vertical force and normalised by the mass of each pig as this changed significantly over the course of the experiment and heavy pigs exert more force than lighter ones. The weight of the pig was determined from the average total force reading prior to the startle stimulus. All readings had this weight subtracted and were subsequently divided by this value to get vertical acceleration in units of gravities, g, where one g is 9.8 m/s2. Load Platform Startle Magnitudes were calculated as the peak absolute acceleration during a 0.7 s window after the startle stimulus. Load Platform Freeze Durations were calculated using load platform acceleration measures and thresholding in the same way as for Image Analysis (IA) data.
Kinect data
Depth maps were extracted from Kinect data collected at a 30 Hz frame rate (see Supplementary Information). For each depth map sequence, a time series recording the vertical displacement of the pigs’ centroids for the total duration of each test was constructed. These time series were used in all subsequent analyses. Vertical centroid speed was the magnitude of the first differential with respect to time in these series. Acceleration magnitude was the second differential. Startle magnitudes and freeze durations were extracted using both of these measures. Kinect Speed Startle Magnitude and Kinect Acceleration Startle Magnitude were calculated as, respectively, the peak absolute values of the vertical centre of mass velocity (mm/frame), and acceleration (mm/frame2) during a 0.7 s window following the startle. The method used to calculate Kinect Speed Freeze Duration and Kinect Acceleration Freeze Duration was the same as that used for the IA data.
Kinematic data
Kinematic data were recorded for 15 s following the startle stimulus and sub-sampled at 30 Hz. We extracted the magnitude of velocities and accelerations for each spherical marker on the pig individually and then took the average of the magnitude of those markers’ trajectories. We only considered motion in the vertical direction as this was most likely to correspond to load platform readings. Kinematic Velocity Startle Magnitude and Kinect Acceleration Startle Magnitude were, respectively, the peak absolute values of vertical velocity (mm/frame) and acceleration (mm/frame2) during a 0.7 s window after the startle stimulus. Freeze durations were not calculated due to the short 15 s recording window which failed to capture longer duration freezes.
Statistical analysis
Data extraction generated the following variables. Startle Magnitude estimates: Observer Startle Magnitude Score; KLT Acceleration Startle Magnitude; Kinect Acceleration Startle Magnitude; Kinect Speed Startle Magnitude; Kinematic Acceleration Startle Magnitude; Kinematic Velocity Startle Magnitude; Load Platform Startle Magnitude. Freeze Duration estimates: Observer Freeze Duration; KLT Acceleration Freeze Duration; KLT Speed Freeze Duration; Kinect Acceleration Freeze Duration; Kinect Speed Freeze Duration; Load Platform Freeze Duration.
Comparisons of automated readouts with observer ground truth measures – startle magnitude
Due to inequality of variances and a highly skewed distribution of Load Platform data which was resistant to transformation, non-parametric statistics were used. We compared ground truth Observer Startle Magnitude scores to other measures of startle magnitude using Spearman Rank correlations. Given the hierarchical nature of the data, data points were not independent rendering derived p-values inaccurate. We therefore constructed multilevel regression models in MLwiN83 with Test (n = 280) nested within Session (n = 8) and Session nested within Pig (n = 12). The response variable was Observer Startle Magnitude score (0–4), thus requiring use of an ordinal response multinomial model, with a reference category of 4 (Flee). In this model, negative estimated coefficients of predictor variables would therefore indicate a positive relationship with Observer Startle Magnitude. We carried out univariate analyses by adding each of the other measures of startle magnitude individually into the model as predictors of our response variable and used Wald tests to examine the significance of the term in the model and thus generate approximate p-values. The predictors were KLT, Kinect, Kinematic and Load Platform Acceleration, and Kinect and Kinematic Velocity estimates of startle magnitude.
Comparisons of automated readouts with observer ground truth measures – freeze duration
Due to the large number of zero values present in this dataset, Spearman Rank correlations were again used to compare the ground truth Observer Freeze Duration with other measures. A multilevel model was constructed in MLwiN to derive p-values that accounted for the nested structure of the data, with Observer Freeze Duration as the response variable. Although the data were not amenable to being transformed to normality, we used a normal rather than binary model to allow comparison to the correlation results. In this model, a positive estimated coefficient of a predictor variable indicated a positive relationship with the response variable. The predictors were Kinect, KLT and Load Platform Acceleration, Kinect Speed, and KLT Speed estimates of freeze duration. To check that significant relationships were not solely influenced by the large number of zero values, we calculated the sensitivity, specificity and positive/negative predictive values of each of the freeze or no-freeze categories. Finally, for robustness we excluded the zero values altogether and examined the correlation coefficients for those cases where both measures scored a freeze as occurring.
Analysis of the effects of other factors on observer measures of startle and freeze
Multilevel models were used to investigate the possible effects of a range of factors (e.g. pig sex, weight, behaviour prior to test, orientation in apparatus, time of day, experience of test across time and within a day, startle stimulus, test session type; Table S3) on observer measures of startle and freeze responses. As above, an ordinal response multinomial model was used for startle magnitude data with Observer Startle Magnitude (0–4) as the response variable. However for freeze data, where the comparison to continuous measures required in the above analyses was now not essential, we converted the non-normal data into a binary freeze/no-freeze variable and used a binomial model with a logit link function. The hierarchical structure of the data was as described previously.
Univariable analysis was first completed by adding each of the possible influencing factors individually into the multilevel model and recording whether they were significantly related to the response variable. A multivariable model was then constructed by sequentially adding in each of the significant influencing factors (most significant first) and checking at each stage that all previously added factors in the model remained significant. Once this had been done, any non-significant factors were entered into the model one by one. If they were now significant, the model-building process was repeated until none of the remaining factors, when introduced into the model, were found to be significant. Wald tests were then used to check whether the levels of hierarchical structure within the model were significant. If not, they were removed from the final model.
Results
Comparisons of automated readouts with observer ground truth measures – startle magnitude
The ground truth Observer Startle Magnitude measure was significantly positively correlated with all automated measures of startle magnitude (Table 2). The strongest correlation was with the Kinematic measures, although these were only available for a subset of the data. Multilevel models showed that the Load Platform and then KLT Acceleration measures were the strongest predictors of the Observer Startle Magnitude scores. (Table 2). KLT Acceleration data were strongly positively correlated not just with Observer Startle Magnitude, but also with all the automated measures, which are widely used to measure movement (Load Platform, rs = 0.824: n = 280, P < 0.001; Kinematic Velocity: rs = 0.821, n = 70, P < 0.001; Kinematic Acceleration: rs = 0.816, n = 70, P < 0.001; Kinect Speed: rs = 0.772, n = 280, P < 0.001; Kinect Acceleration: rs = 0.693, n = 280, P < 0.001; Fig. 1).
Comparisons of automated readouts with observer ground truth measures – freeze duration
The ground truth Observer Freeze Duration measure was significantly and positively correlated with the Kinect, Load Platform and KLT measures. Multilevel models showed that the Kinect Speed and then KLT Speed measures were the strongest predictors of the Observer Freeze Duration scores. (Table 3).
Given the large number of Observer Freeze Duration zero values, we investigated how well the different measures predicted whether an Observer Freeze was detected or not (Table 4). Sensitivities were variable, with the Kinect measures being most sensitive (0.76–0.80), followed by KLT Speed (0.71) and then the Load Platform (0.59) and KLT Acceleration (0.55) measures. The same order was evident for Negative Predictive Value which varied from 0.80 to 0.68. Specificities and Positive Predictive Values were all greater than 0.9 apart from those of the Kinect Acceleration measure (0.89 and 0.88 respectively). Correlation coefficients for the sample of True Positives (i.e. where both the Observer and the other measure scored a freeze as occurring) indicated that all measures were strongly positively correlated with Observer Freeze Duration (KLT Speed: rs = 0.920, n = 104, P < 0.001; KLT Acceleration: rs = 0.855, n = 79, P < 0.001; Kinect Speed: rs = 0.889, n = 109, P < 0.001; Kinect Acceleration: rs = 0.867, n = 114, P < 0.001; Load Platform: rs = 0.877, n = 84, P < 0.001). Furthermore, KLT Speed estimates of freeze duration were strongly positively correlated with those derived from the other automated measures, when both measures detected a freeze (Kinect Speed: rs = 0.938, n = 96, P < 0.001; Kinect Acceleration: rs = 0.879, n = 98, P < 0.001; Load Platform: rs = 0.894, n = 80, P < 0.001 Fig. 2). We did not calculate corresponding KLT Acceleration correlations due to the low sensitivity of this measure.
Overall, computer vision KLT Acceleration estimates of startle magnitude, and KLT Speed estimates of freeze duration, correlated well with the corresponding Observer ground truth measures and with Load Platform, Kinematic and Kinect measures of pig movement.
Effects of other factors on the observer measure of startle magnitude
The multilevel ordinal regression model of the influence of factors on Observer Startle Magnitude scores initially included Tests nested within Sessions within Pig. However neither of the sets of random effects for Session or Pig were significant when the predictor variables were added, and therefore the final model did not require any random effects. Factors significant in the final model are shown in Table 5. The primary reason that Session was significant is due to stronger startle magnitudes occurring in Session 4, when the startling stimuli started to be varied, compared to Sessions 5–8. The use of new stimuli resulted in increased startle magnitudes in comparison to the original balloon-burst stimulus, in particular when the bin bag was used. When compared to pigs that were rated as being calm prior to presentation of the first startling stimulus (relaxed-tense score of 1), those that received a score of 2 or 3 displayed a stronger startle reaction, but this effect was not seen for those animals with scores of 4 and 5. Larger startle reactions were seen when the pig was orientated away from the stimulus.
Effects of other factors on the observer measure of whether a freeze response occurred
The multilevel model initially included Tests nested within Sessions within Pig. However again neither Session or Pig random effects were significant once predictors were added and so the final model did not include a hierarchical structure. The likelihood of a freeze response decreased from Session 1 to 2 and increased significantly in Session 4 when the startling stimuli started to be varied. There was also an increased likelihood of a freeze response when the pig was facing away from the stimulus at the point of testing, and for pigs with a ‘relaxed-tense’ score of 2 or 3 compared to a score of 1 prior to the first startling stimulus. However, pigs with a score of 4 or 5 did not show this increased likelihood of freezing (Table 6).
Discussion
In this study, we investigated the validity of computer vision image analysis (IA) as a measure of startle magnitude and freeze duration Defence Cascade (DC) responses in pigs, because of its potential as a practical tool for assessing these responses under field conditions. We used behavioural observation of video as our ground truth measure of startle and freeze responses (see72,73,74) and also collected load platform, kinematic, and Kinect depth-camera measures because these have been used to assess movement in other species.
Our KLT Acceleration image analysis estimate of startle magnitude, and estimates from all the automated measurements we made, were significantly positively correlated with the observer behavioural observation data. Kinematic measures generated the highest correlation coefficient which is unsurprising given that they are designed to capture detailed 3D movements and were derived using four carefully positioned cameras to measure precise movement of markers on the pigs. The Load Platform produced the next strongest correlation, followed by the image analysis KLT Acceleration data, with the Kinect measure generating the weakest correlations. Load Platform and KLT Acceleration were the best predictors of Observer Startle Magnitude in our multilevel model. The relatively weak predictive power of kinematic data in these analyses likely reflects the smaller sample size for this measure. Our image analysis estimate of startle magnitude was also strongly positively correlated with the load platform, kinematic and depth-camera measures.
Our KLT Speed and Acceleration image analysis estimates of freeze duration were strongly positively correlated with the observer measure, as were Kinect Acceleration and Speed, and Load Platform data. Because there were a high number of zeroes (non-freezes) during tests, we carried out sensitivity and specificity analyses of whether or not a freeze was detected against the human observer ground truth. All measures had high specificity – they rarely detected a freeze when the observer did not detect one – with computer vision KLT measures performing particularly strongly. Their positive predictive values – how well a detected freeze actually indicates that a freeze did occur – were therefore high. However sensitivity was lower, particularly for Load Platform and KLT Acceleration data, which frequently did not detect freezes when the observer did and hence generated a high proportion of false negatives. In many such cases, the observer recorded short freezes (<2 s) in which the body appeared tensed but isolated parts (e.g. ears) were moving. This short duration and residual movement may have accounted for the failure of other measures to detect a freeze response.
For true positive data (when both the observer and automated measures detected a freeze), automated measures of freeze duration were strongly positively correlated with those made by the observer, with KLT Speed and Kinect Speed performing best. Furthermore, KLT Speed estimates of freeze duration were also strongly positively correlated with the other automated measures.
In addition to analysing the relationship between different measures of startle and freeze behaviour, we also investigated factors that may influence expression of these behaviours in pigs tested under laboratory conditions. In our study, repeat-testing of individuals was necessary to generate sufficient examples of DC responses to establish associations between different measures of these responses, whilst also following 3Rs recommendations to minimise animal use. However, repeat-testing resulted in pigs habituating to the original startling stimulus (bursting balloon), as has been found in previous studies of pigs72,84. Consequently, we varied the nature of the stimulus within test sessions, starting in session 4, and startle response magnitude increased after this session. Similarly, the likelihood of a freeze response initially decreased from session 1 to session 2 when the bursting balloon was used repeatedly, but increased in session 4 when stimuli started to be varied within sessions. The type of stimulus used also affected DC responses with bin bag and umbrella stimuli, both involving rapid movement, being particularly effective at generating greater startle magnitudes than the balloon. Use of stimuli with a pronounced visual component may thus be more potent inducers of DC responses in pigs compared to purely auditory stimuli.
Larger startle magnitudes were observed when pigs were orientated away from the stimulus possibly because startling stimuli presented in this context induced a greater surprise reaction20,21,22. An alternative explanation is that those pigs orientated away from the stimulus were more likely to make a large jump with movement (Observer Startle Magnitude score of 3) in order to face the stimulus, whilst pigs already facing the stimulus may have been more likely to jump on the spot (score of 2). However, our finding that there was also an increased probability of freezing when pigs were orientated away from the stimulus lends support to the idea that the stimulus was indeed more startling and surprising in this context, including leading to increased post-startle processing (during freeze) to resolve the source of the event21,22.
A subjective rating of how ‘relaxed or tense’ the pigs were before the first stimulus was delivered during a test also influenced DC responses. For both startle magnitude and freeze occurrence an increase in reaction was seen for pigs scored as 2 or 3 when compared to the calmest pigs (scored as 1). However, this difference was not observed in pigs scored as 4 or 5 (least calm). In fact, the probability of a freeze reaction was decreased in pigs with a score of 4 compared to those with a score of 1. Although this seems counterintuitive, we observed informally that the least calm pigs appeared to pay least attention to their surroundings and were often attempting to leave the test pen. Consequently, delivery of the startling stimulus did not intrude into their already active behaviour. This re-emphasises the need for pigs to be settled and calm at the point of testing, for example in a home pen.
It is interesting to note that our models found corresponding effects of moderating factors on both startle magnitude and probability of freeze occurrence. Both increased when startling stimuli were altered, when the pig was facing away from them, and in pigs who were slightly or moderately tense prior to their presentation, although not for animals rated as tense or very tense who appeared to be focused on leaving the test pen. This coherence of effects indicates that startle magnitude and freeze probability may reflect a similar underlying construct, for example affective valence25,28,38,40.
In summary, we found that computer vision image analysis measures were comparable with force, kinematic and depth measurements at estimating ground truth observer measures of both startle magnitude and freeze duration/occurrence. These findings demonstrate that computerised image analysis of video recordings can be used to detect and quantify startle and freeze Defence Cascade responses in individual pigs, and hence has potential as a practical automated measure of these behaviours under field conditions. We also identified factors including startling stimulus characteristics, orientation relative to the stimulus, and ‘relaxed-tense’ behaviour that can moderate DC responses and hence should be borne in mind during further investigation of this behaviour. The similarity of these moderating effects across both startle and freeze behaviours indicates that these two forms of DC response may represent similar underlying constructs such as affective valence. If variation in startle and freeze behaviour does indeed reflect affective state in pigs, as it appears to in other species20,21,22,28,59,60,61, computer vision image analysis may thus provide a novel method for assessing pig welfare in the field. Further research on the link between affective state and pig startle and freeze behaviour, the capacity of image analysis to measure these behaviours in groups of pigs, and the utility of the approach under field conditions, including startle stimulus design and issues of habituation, is needed to realise this potential73,78,80.
Data availability
Data for all variables used in the statistical analyses will be made available on request. Extensive sets of uncompressed raw video RGB data, Kinect data, load platform voltage data and kinematic x,y,z data are also available on reasonable request.
References
Main, D. C. J., Webster, F. & Green, L. E. Animal welfare assessment in farm assurance schemes. Acta Agriculturae Scandinavica Section a-Animal Science 51, 108–113, https://doi.org/10.1080/090647001316923171 (2001).
Whay, H. R., Main, D. C. J., Green, L. E. & Webster, A. J. F. Assessment of the welfare of dairy cattle using animal-based measurements: direct observations and investigation of farm records. Veterinary Record 153, 197–202, https://doi.org/10.1136/vr.153.7.197 (2003).
Whay, H. R., Main, D. C. J., Green, L. E. & Webster, A. J. F. Animal-based measures for the assessment of welfare state of dairy cattle, pigs and laying hens: consensus of expert opinion. Animal Welfare 12, 205–217 (2003).
Main, D. C. J., Whay, H. R., Lee, C. & Webster, A. J. F. Formal animal-based welfare assessment in UK certification schemes. Animal Welfare 16, 233–236 (2007).
Temple, D., Dalmau, A., de la Torre, J. L. R., Manteca, X. & Velarde, A. Application of the Welfare Quality protocol to assess growing pigs kept under intensive conditions in Spain. Journal of Veterinary Behavior-Clinical Applications and Research 6, 138–149, https://doi.org/10.1016/j.jveb.2010.10.003 (2011).
Friedrich, L., Krieter, J., Kemper, N. & Czycholl, I. Test-retest reliability of the ‘Welfare Quality animal welfare assessment protocol for sows and piglets’. Part 1. Assessment of the welfare principle of ‘appropriate behavior’. Animals 9, https://doi.org/10.3390/ani9070398 (2019).
Andreasen, S. N., Wemelsfelder, F., Sandoe, P. & Forkman, B. The correlation of Qualitative Behavior Assessments with Welfare Quality protocol outcomes in on-farm welfare assessment of dairy cattle. Applied Animal Behaviour Science 143, 9–17, https://doi.org/10.1016/j.applanim.2012.11.013 (2013).
Duncan, I. J. H. & Petherick, J. C. The implications of cognitive-processes for animal-welfare. Journal of Animal Science 69, 5017–5022 (1991).
Duncan, I. J. H. Animal welfare defined in terms of feelings. Acta Agriculturae Scandinavica Section a-Animal Science, 29-35 (1996).
Mendl, M. Animal husbandry - assessing the welfare state. Nature 410, 31–32, https://doi.org/10.1038/35065194 (2001).
Dawkins, M. Animal welfare and the paradox of animal consciousness. Advances in the Study of Behavior 47 47, 5–38, https://doi.org/10.1016/bs.asb.2014.11.001 (2015).
Knierim, U. & Winckler, C. On-farm welfare assessment in cattle: validity, reliability and feasibility issues and future perspectives with special regard to the Welfare Quality approach. Animal Welfare 18, 451–458 (2009).
Battini, M., Barbieri, S., Waiblinger, S. & Mattiello, S. Validity and feasibility of Human-Animal Relationship tests for on-farm welfare assessment in dairy goats. Applied Animal Behaviour Science 178, 32–39, https://doi.org/10.1016/j.applanim.2016.03.012 (2016).
Phythian, C. J., Michalopoulou, E., Cripps, P. J., Duncan, J. S. & Wemelsfelder, F. On-farm qualitative behaviour assessment in sheep: Repeated measurements across time, and association with physical indicators of flock health and welfare. Applied Animal Behaviour Science 175, 23–31, https://doi.org/10.1016/j.applanim.2015.11.013 (2016).
Czycholl, I., Beilage, E. G., Henning, C. & Krieter, J. Reliability of the qualitative behavior assessment as included in the Welfare Quality Assessment protocol for growing pigs. Journal of Animal Science 95, 3445–3454, https://doi.org/10.2527/jas2017.1525 (2017).
Richmond, S. E. et al. Evaluation of animal-based indicators to be used in a welfare assessment protocol for sheep. Frontiers in Veterinary Science 4, https://doi.org/10.3389/fvets.2017.00210 (2017).
de Vries, M. et al. Assessment time of the Welfare Quality protocol for dairy cattle. Animal Welfare 22, 85–93, https://doi.org/10.7120/09627286.22.1.085 (2013).
de Jong, I. C. et al. Simplifying the Welfare Quality assessment protocol for broiler chicken welfare. Animal 10, 117–127, https://doi.org/10.1017/s1751731115001706 (2016).
Buijs, S., Ampe, B. & Tuyttens, F. A. M. Sensitivity of the Welfare Quality broiler chicken protocol to differences between intensively reared indoor flocks: which factors explain overall classification? Animal 11, 244–253, https://doi.org/10.1017/s1751731116001476 (2017).
Lang, P. J., Bradley, M. M. & Cuthbert, B. N. Emotion, attention, and the startle reflex. Psychological Review 97, 377–395, https://doi.org/10.1037/0033-295x.97.3.377 (1990).
Lang, P. J., Davis, M. & Ohman, A. Fear and anxiety: animal models and human cognitive psychophysiology. J. Affect. Disord 61, 137–159 (2000).
Lang, P. J. & Bradley, M. M. Emotion and the motivational brain. Biological Psychology 84, 437–450, https://doi.org/10.1016/j.biopsycho.2009.10.007 (2010).
Kozlowska, K., Walker, P., McLean, L. & Carrive, P. Fear and the defense cascade: clinical implications and management. Harvard Review of Psychiatry 23, 263–287, https://doi.org/10.1097/hrp.0000000000000065 (2015).
Roelofs, K. Freeze for action: neurobiological mechanisms in animal and human freezing. Philosophical Transactions of the Royal Society B-Biological Sciences 372, https://doi.org/10.1098/rstb.2016.0206 (2017).
Koch, M. The neurobiology of startle. Progress in Neurobiology 59, 107–128, https://doi.org/10.1016/s0301-0082(98)00098-7 (1999).
Vrana, S. R. Startle reflex response during sensory modality-specific disgust, anger, and neutral imagery. J. Psychophysiol. 8, 211–218 (1994).
Bradley, M. M., Codispoti, M., Cuthbert, B. N. & Lang, P. J. Emotion and motivation I: defensive and appetitive reactions in picture processing. Emotion 1, 276–298, https://doi.org/10.1037/1528-3542.1.3.276 (2001).
Grillon, C. & Baas, J. A review of the modulation of the startle reflex by affective states and its application in psychiatry. Clinical Neurophysiology 114, 1557–1579, https://doi.org/10.1016/s1388-2457(03)00202-5 (2003).
Hurlemann, R. et al. Diminished appetitive startle modulation following targeted inhibition of prefrontal cortex. Scientific Reports 5, https://doi.org/10.1038/srep08954 (2015).
Nelson, B. D. & Hajcak, G. Defensive motivation and attention in anticipation of different types of predictable and unpredictable threat: a startle and event-related potential investigation. Psychophysiology 54, 1180–1194, https://doi.org/10.1111/psyp.12869 (2017).
Gantiva, C., Araujo, A., Calderon, L., Gomez, P. & Reina, F. Psychophysiological responses to facial expression of surprise, sadness, and disgust. Australian Journal of Psychology 71, 100–107, https://doi.org/10.1111/ajpy.12222 (2019).
Bradley, M. M., Zlatar, Z. Z. & Lang, P. J. Startle reflex modulation during threat of shock and “threat” of reward. Psychophysiology 55, https://doi.org/10.1111/psyp.12989 (2018).
Herten, N. et al. Enhanced startle responsivity 24 hours after acute stress exposure. Behavioral Neuroscience 130, 521–530, https://doi.org/10.1037/bne0000156 (2016).
Prehn, A., Ohrt, A., Sojka, B., Ferstl, R. & Pause, B. M. Chemosensory anxiety signals augment the startle reflex in humans. Neuroscience Letters 394, 127–130, https://doi.org/10.1016/j.neulet.2005.10.012 (2006).
Brown, J. S., Kalish, H. I. & Farber, I. E. Conditioned fear as revealed by magnitude of startle response to an auditory stimulus. Journal of Experimental Psychology 41, 317–328, https://doi.org/10.1037/h0060166 (1951).
Davis, M. & Astrachan, D. I. Conditioned fear and startle magnitude - effects of different footshock or backshock intensities used in training. Journal of Experimental Psychology-Animal Behavior Processes 4, 95–103, https://doi.org/10.1037/0097-7403.4.2.95 (1978).
Davis, M., Falls, W. A., Campeau, S. & Kim, M. Fear-potentiated startle - a neural and pharmacological analysis. Behavioural Brain Research 58, 175–198, https://doi.org/10.1016/0166-4328(93)90102-v (1993).
Richardson, R., Vishney, A. & Lee, J. Conditioned odor potentiation of startle in rats. Behavioral Neuroscience 113, 787–794, https://doi.org/10.1037/0735-7044.113.4.787 (1999).
Fendt, M., Koch, M. & Schnitzler, H. U. Corticotropin-releasing factor in the caudal pontine reticular nucleus mediates the expression of fear-potentiated startle in the rat. European Journal of Neuroscience 9, 299–305, https://doi.org/10.1111/j.1460-9568.1997.tb01400.x (1997).
Winslow, J. T., Parr, L. A. & Davis, M. Acoustic startle, prepulse inhibition, and fear-potentiated startle measured in rhesus monkeys. Biological Psychiatry 51, 859–866, https://doi.org/10.1016/s0006-3223(02)01345-8 (2002).
Bradley, M. M. & Lang, P. J. Affective reactions to acoustic stimuli. Psychophysiology 37, 204–215, https://doi.org/10.1017/s0048577200990012 (2000).
Andreatta, M., Muhlberger, A. & Pauli, P. When does pleasure start after the end of pain? The time course of relief. Journal of Comparative Neurology 524, 1653–1667, https://doi.org/10.1002/cne.23872 (2016).
Schmid, A., Koch, M. & Schnitzler, H. U. Conditioned pleasure attenuates the startle response in ratS. Neurobiology of Learning and Memory 64, 1–3, https://doi.org/10.1006/nlme.1995.1037 (1995).
Koch, M., Schmid, A. & Schnitzler, H. U. Pleasure-attenuation of startle is disrupted by lesions of the nucleus accumbens. Neuroreport 7, 1442–1446, https://doi.org/10.1097/00001756-199605310-00024 (1996).
Schneider, M. & Spanagel, R. Appetitive odor-cue conditioning attenuates the acoustic startle response in rats. Behavioural Brain Research 189, 226–230, https://doi.org/10.1016/j.bbr.2007.12.017 (2008).
Friemel, C. M., Zimmer, A. & Schneider, M. The CBI receptor as an important mediator of hedonic reward processing. Neuropsychopharmacology 39, 2387–2396, https://doi.org/10.1038/npp.2014.86 (2014).
Steidl, S., Li, L. & Yeomans, J. S. Conditioned brain-stimulation reward attenuates the acoustic startle reflex in rats. Behavioral Neuroscience 115, 710–717, https://doi.org/10.1037/0735-7044.115.3.710 (2001).
Sabatinelli, D., Bradley, M. M. & Lang, P. J. Affective startle modulation in anticipation and perception. Psychophysiology 38, 719–722, https://doi.org/10.1111/1469-8986.3840719 (2001).
Lipp, O. V., Cox, D. & Siddle, D. A. T. Blink startle modulation during anticipation of pleasant and unpleasant stimuli. J. Psychophysiol. 15, 155–162, https://doi.org/10.1027/0269-8803.15.3.155 (2001).
Skolnick, A. J. & Davidson, R. J. Affective modulation of eyeblink startle with reward and threat. Psychophysiology 39, 835–850, https://doi.org/10.1017/s0048577202010302 (2002).
Bach, D. R. A cost minimisation and Bayesian inference model predicts startle reflex modulation across species. Journal of Theoretical Biology 370, 53–60, https://doi.org/10.1016/j.jtbi.2015.01.031 (2015).
Lucas, I., Sanchez-Adam, A., Vila, J. & Guerra, P. Positive emotional reactions to loved names. Psychophysiology 56, https://doi.org/10.1111/psyp.13363 (2019).
Baas, J. M. P. et al. Benzodiazepines have no effect on fear-potentiated startle in humans. Psychopharmacology 161, 233–247, https://doi.org/10.1007/s00213-002-1011-8 (2002).
Richardson, R. & Defina, M. Effects of home-nest odors on the startle response in preweanling rats. Physiology & Behavior 64, 621–624, https://doi.org/10.1016/s0031-9384(98)00100-0 (1998).
Mineur, Y. S., Belzung, C. & Crusio, W. E. Effects of unpredictable chronic mild stress on anxiety and depression-like behavior in mice. Behavioural Brain Research 175, 43–50 (2006).
Wallace, K. J. & Rosen, J. B. Predator odor as an unconditioned fear stimulus in rats: elicitation of freezing by trimethylthiazoline, a component of fox feces. Behavioral Neuroscience 114, 912–922, https://doi.org/10.1037/0735-7044.114.5.912 (2000).
Fadok, J. P. et al. A competitive inhibitory circuit for selection of active and passive fear responses. Nature 542, 96–+, https://doi.org/10.1038/nature21047 (2017).
Mendl, M., Burman, O. H. P. & Paul, E. S. An integrative and functional framework for the study of animal emotion and mood. Proceedings of the Royal Society B-Biological Sciences 277, 2895–2904, https://doi.org/10.1098/rspb.2010.0303 (2010).
Phillips, R. G. & Ledoux, J. E. Lesions of the dorsal hippocampal-formation interfere with background but not foreground contextual fear conditioning. Learn. Mem. 1, 34–44 (1994).
Daldrup, T. et al. Expression of freezing and fear-potentiated startle during sustained fear in mice. Genes Brain and Behavior 14, 281–291, https://doi.org/10.1111/gbb.12211 (2015).
Ross, J. M. & Fletcher, M. L. Assessing classical olfactory fear conditioning by behavioral freezing in mice. Bio-Protocol 8, https://doi.org/10.21769/BioProtoc.3013 (2018).
Yee, N., Schwarting, R. K. W., Fuchs, E. & Wohr, M. Juvenile stress potentiates aversive 22-kHz ultrasonic vocalizations and freezing during auditory fear conditioning in adult male rats. Stress-the International Journal on the Biology of Stress 15, 533–544, https://doi.org/10.3109/10253890.2011.646348 (2012).
Hashimoto, S., Inoue, T. & Koyama, T. Serotonin reuptake inhibitors reduce conditioned fear stress-induced freezing behavior in rats. Psychopharmacology 123, 182–186, https://doi.org/10.1007/bf02246175 (1996).
Ohyama, M., Kondo, M., Yamauchi, M., Imanishi, T. & Koyama, T. Asenapine reduces anxiety-related behaviours in rat conditioned fear stress model. Acta Neuropsychiatrica 28, 327–336, https://doi.org/10.1017/neu.2016.17 (2016).
Osada, K., Kurihara, K., Izumi, H. & Kashiwayanagi, M. Pyrazine analogues are active components of wolf urine that induce avoidance and freezing behaviours in mice. Plos One 8, https://doi.org/10.1371/journal.pone.0061753 (2013).
Sartori, S. B. et al. Enhanced fear expression in a psychopathological mouse model of trait anxiety: pharmacological interventions. Plos One 6, https://doi.org/10.1371/journal.pone.0016849 (2011).
Modi, M. E. et al. Peripheral administration of a long-acting peptide oxytocin receptor agonist inhibits fear-induced freezing. Journal of Pharmacology and Experimental Therapeutics 358, 164–172, https://doi.org/10.1124/jpet.116.232702 (2016).
Widman, A. J. et al. Rats bred for high anxiety exhibit distinct fear-related coping behavior, hippocampal physiology, and synaptic plasticity-related gene expression. Hippocampus 29, 939–956, https://doi.org/10.1002/hipo.23092 (2019).
Huber, A., Barber, A. L. A., Farago, T., Muller, C. A. & Huber, L. Investigating emotional contagion in dogs (Canis familiaris) to emotional sounds of humans and conspecifics. Animal Cognition 20, 703–715, https://doi.org/10.1007/s10071-017-1092-8 (2017).
Knox, D., Fitzpatrick, C. J., George, S. A., Abelson, J. L. & Liberzon, I. Unconditioned freezing is enhanced in an appetitive context: implications for the contextual dependency of unconditioned fear. Neurobiology of Learning and Memory 97, 386–392, https://doi.org/10.1016/j.nlm.2012.03.002 (2012).
Sweeney, F. F., O’Leary, O. F. & Cryan, J. F. GABA(B) receptor ligands do not modify conditioned fear responses in BALB/c mice. Behavioural Brain Research 256, 151–156, https://doi.org/10.1016/j.bbr.2013.07.035 (2013).
Blackshaw, J., Blackshaw, A. & McGlone, J. Startle-freeze behaviour in weaned pigs. International Journal of Comparative Psychology 11, 30–39 (1998).
Statham, P. et al. Starling pigs – a potential welfare measure? In Proceedings of the International Conference on Pig Welfare p.124 (Ministry of Food, Agriculture and Fisheries of Denmark, Copenhagen, 2015).
Carreras, R. et al. Can the way pigs are handled alter behavioural and physiological measures of affective state? Behavioural Processes 142, 91–98, https://doi.org/10.1016/j.beproc.2017.06.005 (2017).
Boissy, A., Veissier, I. & Roussel, S. Behavioural reactivity affected by chronic stress: An experimental approach in calves submitted to environmental instability. Animal Welfare 10, S175–S185 (2001).
Desire, L., Veissier, I., Despres, G. & Boissy, A. On the way to assess emotions in animals: Do lambs (Ovis aries) evaluate an event through its suddenness, novelty, or unpredictability? J. Comp. Psychol. 118, 363–374 (2004).
Ross, M., Garland, A., Harlander-Matauschek, A., Kitchenham, L. & Mason, G. Welfare-improving enrichments greatly reduce hens’ startle responses, despite little change in judgment bias. Scientific Reports 9, https://doi.org/10.1038/s41598-019-48351-6 (2019).
Blokhuis, H. J., Veissier, I., Miele, M. & Jones, B. The Welfare Quality project and beyond: Safeguarding farm animal well-being. Acta Agriculturae Scandinavica Section a-Animal Science 60, 129–140, https://doi.org/10.1080/09064702.2010.523480 (2010).
Amorim, F. E., Moulin, T. C. & Amaral, O. B. A Freely available, self-calibrating software for automatic measurement of freezing behavior. Frontiers in Behavioral Neuroscience 13, https://doi.org/10.3389/fnbeh.2019.00205 (2019).
Statham, P. et al. The Defence Cascade as an on-farm welfare measure. In Proceedings of the UK and Ireland Regional Meeting of the International Society for Applied Ethology.
Noldus. The Observer XT Reference Manual. Version 10.5. (Noldus Information Technology bv, 2011).
Shi, J. & Tomasi, C. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. (IEEE, 1994).
Charlton, C., Rasbash, J., Browne, W., Healy, M. & Cameron, B. MLwiN Version 3.03. (Centre for Multilevel Modelling, University of Bristol, 2019).
Lind, N. M., Arnfred, S. M., Hemmingsen, R. P. & Hansen, A. K. Prepulse inhibition of the acoustic startle reflex in pigs and its disruption by D-amphetamine. Behavioural Brain Research 155, 217–222, https://doi.org/10.1016/j.bbr.2004.04.014 (2004).
Acknowledgements
We are grateful to the UK Biotechnology and Biological Sciences Research Council (Grant BB/I005641/1) for funding this research, and to two anonymous reviewers for their helpful comments.
Author information
Authors and Affiliations
Contributions
M.M., P.S., S.J., S.H., N.C., G.R.C., W.J.B. and E.S.P. designed the experiments. P.S., S.J., S.H. and G.R.C. collected the data and extracted it for analysis. P.S. and S.J. generated behaviour observation data outputs, and S.H., N.C. and G.R.C. generated data outputs from computer vision, load platform, Kinect and kinematic records. P.S. analysed the data with support from W.J.B. and M.M. M.M. and P.S. wrote the paper and prepared the figures. S.H., G.R.C. and W.J.B. commented on the manuscript.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary Information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Statham, P., Hannuna, S., Jones, S. et al. Quantifying defence cascade responses as indicators of pig affect and welfare using computer vision methods. Sci Rep 10, 8933 (2020). https://doi.org/10.1038/s41598-020-65954-6
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41598-020-65954-6
This article is cited by
-
Going Deeper than Tracking: A Survey of Computer-Vision Based Recognition of Animal Pain and Emotions
International Journal of Computer Vision (2023)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.