1 Introduction
2 Measures and tools
2.1 Measures of transport behavior
2.1.1 Tour
2.1.2 Trip
2.1.3 Leg
2.1.4 Purpose
2.1.5 Stop
2.1.6 Transport mode
2.1.7 Mode-chain-type
2.1.8 Travel-diary
2.1.9 Ground truth
2.2 Pioneering smartphone-based travel surveys
2.3 Smartphone capabilities
2.4 Physical limitations for data validation
2.4.1 Person-to-device validation
2.4.2 Device-to-device validation
3 Measuring transport behavior
References | No. | Classes | Score | Metric | Validation | Area |
---|---|---|---|---|---|---|
Zhou et al. [131] | 6 | Walk, Bike, Bus, Car, Rail, Plain | 86.5% | Accuracy | Hold-out | Beijing |
Bohte and Maat [21] | 6 | Car, Train, Bus-Tram-Metro, Foot, Bicycle, Other | 70.00% | Accuracy | n.p. | Netherlands |
Martin et al. [75] | 5 | Walk, Bike, Bus, Car, Rail | 96.8% | Accuracy | Manifold-cross-validation | Minnesota |
Jahangiri and Rakha [55] | 5 | Walk, Bike, Bus, Car, Run | 95.1% | F-Score | Manifold-cross-validation, Out-of-bag-estimate | Tennessee |
Semanjski et al. [96] | 5 | Walk, Bike, Bus, Car, Rail | 94.00% | Accuracy | Manifold-cross-validation | Leuven |
Zhou et al. [132] | 5 | Walk, Bike, Run, in-Vehicle, Stationary | 93.8% | Accuracy | Hold-out | Georgia (USA) |
Zhu et al. [134] | 5 | Walk, Bike, Bus, Car, Rail | 93.45% | F1-Score | Manifold-cross-validation | Beijing |
Xiao et al. [119] | 5 | Walk, Bike, el-Bike, Car, Bus | 92.74% | Accuracy | Manifold-cross-validation | Shanghai |
Rasmussen et al. [90] | 5 | Walk, Bike, Car, Bus, Rail | 92.4% | Accuracy | n.p. | Copenhagen |
Yazdizadeh et al. [122] | 5 | Walk, Bike, Public transit, Car, Car and Public transit | 88.00% | F1-Score weighted average | Manifold-cross-validation | Montreal |
Dabiri and Heaslip [31] | 5 | Walk, Bike, Bus, Car, Rail | 84.8% | F-Score | Manifold-cross-validation | Beijing |
Byon and Liang [22] | 5 | Auto, Bus, Streetcar, Bike, Walk | 82.00% | F1-Score weighted average | Hold-out | Toronto |
Thomas et al. [104] | 5 | Walk, Bike, Bus, Car, Rail | 82.00% | Accuracy | n.p. | Netherlands, [43] |
Dabiri et al. [32] | 5 | Walk, Bike, Bus, Drive, Train | 76.4% | F1-Score weighted average | Manifold-cross-validation | Beijing |
Jiang et al. [57] | 4 | Walk, Bike, Bus, Car | 98.00% | Accuracy | Hold-out | Beijing |
Assemi et al. [11] | 4 | Walk, Bike, Bus, Car | 94.7% | Accuracy | Hold-out | New-Zealand |
Yazdizadeh et al. [123] | 4 | Walk, Bike, Transit, Car | 91.8% | Accuracy | Manifold-cross-validation | Montreal |
Mäenpää et al. [72] | 4 | Walk, Bike, Bus, Car | 90.7% | F1-Score | Manifold-cross-validation, Out-of-bag-estimate | Beijing. 1 week BUS trajectories, 1000 trajectories from Open Street Map (OSM) |
Yazdizadeh et al. [124] | 4 | Walk, Bike, Transit, Car | 83.4% | Accuracy | Manifold-cross-validation | Montreal |
References | Person-day | Users | Ground truth | Observations | Time | Area | Smartphone App |
---|---|---|---|---|---|---|---|
Semanjski et al. [96] | 24,900 | 8303 | Validated-by-respondents | 30,000 trips 3,960,243 GPS points 340,000 km | n.p. | Leuven | Routecoach |
Yazdizadeh et al. [124] | 88,630 | 6846 | Validated-by-respondents (102,904 trips) | 623,718 trips | 2 months collection period | Montreal | MTL Traject App |
Yazdizadeh et al. [123] | 88,630 | 6846 | Validated-by-respondents (P2D) | 102,904 trips | 2 months collection period | Montreal | MTL Traject App |
Yazdizadeh et al. [122] | 88,630 | 6846 | Validated-by-respondents (P2D) | 131,777 trips 33 mln GPS points | 2 months collection period | Montreal | MTL Traject App |
Bohte and Maat [21] | 40,208 | 1104 | Validated-by-respondents (P2D) | n.p. | 7395 days | Netherlands | GPS logger and Web based validation |
Thomas et al. [104] | n.p. | 600 | Validated-by-respondents | 60,000 trips | 3 batches per 1 month each | Netherlands, [43] | Move smarter |
Xiao et al. [119] | 1248 | 202 | Validated-by-respondents | 4685 Trip-legs | n.p. | Shanghai | Shangai City—Smartphone based travel survey |
Dabiri et al. [32] | 4000 | 189 | Partially validated-by-respondents (69 respondents) | 17,621 trajectories 1,292,951 km 50,176 h | 3 years collection period | Beijing | Geolife [129] |
Rasmussen et al. [90] | 644 | 101 | Validated-by-respondents (P2P) | 6,419,441 GPS points 1783 h of travel | 3–5 days per respondent | Copenhagen | GPS logger |
Assemi et al. [11] | 372 | 76 | Validated-by-respondents | 760,000 GPS observations, 530 h trajectories | 2 months per respondent | New-Zealand | Advanced Travel Logging Application for Smartphones II (ATLAS II) |
Mäenpää et al. [72] | 4000 | > 69 | Validated-by-respondents | n.p. | n.p. | Beijing. 1 week BUS trajectories, 1000 trajectories from Open Street Map (OSM) | Geolife [129], Journeys APIa, OpenStreetMapb |
Dabiri and Heaslip [31] | 4000 | 69 | Validated-by-respondents | n.p. | 3 years collection period | Beijing | Geolife [129] |
Jiang et al. [57] | 4000 | 69 | Validated-by-respondents | n.p. | 3 years collection period | Beijing | Geolife [129] |
Zhou et al. [131] | 4000 | 69 | Validated-by-respondents | n.p. | 3 years collection period | Beijing | Geolife [129] |
Zhu et al. [134] | 4000 | 69 | Validated-by-respondents | n.p. | 3 years collection period | Beijing | Geolife [129] |
Zhou et al. [132] | n.p. | 12 | Validated-by-respondents | n.p. | 6 days per respondent | Georgia (USA) | Self Developed App |
Martin et al. [75] | n.p. | 6 | Validated-by-respondents | 347,719 GPS points in 96.59 h (1 Hz) 1.7 mln points Acceleration in 98.62 h (5 Hz) | n.p. | Minnesota | Self Developed App |
Byon and Liang [22] | n.p. | n.p. | n.p. | n.p. | 50 h | Toronto | Self Developed App |
Jahangiri and Rakha [55] | n.p. | n.p. | Validated-by-respondents | n.p. | n.p. | Tennessee | Self Developed App |
References | Method | Main features | AGPS | INS | GIS |
---|---|---|---|---|---|
Assemi et al. [11] | Nested logit model, muiltinomial logistic regression, multiple discriminant analysis | Skewness of speed distribution, share of travel time with speed (m/s) \(\in [2, 8)\), share of travel time with speed (m/s) \(\in [8, 15)\), maximum speed, 95% percentile acceleration, maximum acceleration, acceleration variance, direct distance \(origin \rightarrow destination\), travelled distance \(origin \rightarrow destination\) | Yes | No | No |
Bohte and Maat [21] | Rule-based | Distance \(GPS \rightarrow \,{\text{Points-of-interest}}\), Distance \(GPS \rightarrow LandUse\) | Yes | No | Yes |
Byon and Liang [22] | Neural network | Speed, acceleration, magnetic field, satellites number | GPS | Accelerometer magnetometer | No |
Dabiri and Heaslip [31] | Convolutional neural network, random forest, key nearest neighbor, support vector machines, multi layer perceptron | Speed, acceleration, jerk, bearing rate | Yes | No | No |
Dabiri et al. [32] | SEmi-Supervised Convolutional Autoencoder | GPS points: relative distance, speed | Yes | No | No |
Jahangiri and Rakha [55] | Random forest, bagging model, support vector machines, key nearest neighbor, Max-dependency Min-redundancy | Acceleration spectral entropy, acceleration range, Max angular velocity, average absolute acceleration, average angular velocity | Yes | Accelerometer, gyro-scope, rotation vector | No |
Jiang et al. [57] | Recurrent neural network, Hampel filter | Speed, average speed, standard deviation speed | Yes | No | No |
Mäenpää et al. [72] | Bayesian classier, neural network, random forest, auto encoder | Maximum acceleration, maximum speed, minimum acceleration, minimum speed, average acceleration, average speed, acceleration variance, speed variance, speed skewness, speed kurtosis, acceleration skewness, acceleration kurtosis | Yes | No | No |
Martin et al. [75] | Random forest, key nearest neighbor, principal component analysis, recursive feature elimination | Average change in acceleration (\(\Delta T = 120\) s), 80% percentile speed (\(\Delta T = 120\) s), variance change in acceleration (\(\Delta T = 120\) s), maximum speed (\(\Delta T = 120\) s), average speed (\(\Delta T = 120\) s), average change in speed (\(\Delta T = 120\) s) | Yes | Accelerometer | No |
Rasmussen et al. [90] | Fuzzy logic | 95% percentile acceleration, 95% percentile speed, median speed, network segment | GPS | No | Yes |
Semanjski et al. [96] | Support vector machines | Distance from (DF) motorway, DF railway, DF bicycle lane, DF bus stop, DF railways station, DF car parking, DF bicycle parking, DF bus line | Yes | No | Yes |
Thomas et al. [104] | Bayesian classier | Personal trip history, speed, altitude, longitude, latitude, public transport time-table | Yes | Accelerometer | Yes |
Xiao et al. [119] | Bayesian network | Average speed, 95% percentile speed, average absolute acceleration, travel distance, average heading change, Low-speed-rate (as the ratio of points with speed < threshold) | Yes | No | No |
Yazdizadeh et al. [123] | Counvolutional neural network augmented with ensemble method, with random forest as meta learner | GPS points: relative distance, speed | Yes | No | No |
Yazdizadeh et al. [122] | Random forest | Measures between origin-destination: cumulative and direct distance (m), travel time (Min.), average and 85th percentile speed (km/h), maximum, minimum difference between Min. and Max. acceleration (\({\mathrm{km/h}}^2\)), minimum and maximum slope; Max time interval (min) and Max distance (m) between each consecutive pair of GPS point; time of day and time of week; age, gender, occupation; average value of residential buildings around each individual’s home (in 250 m radius); direct distance between the origin and nearest public transit stop; direct distance between the destination and nearest public transit stop; average value of residential buildings around each individual’s home (in 250 m radius) | Yes | No | Yes |
Yazdizadeh et al. [124] | Semi-supervised Generative Adversarial Networks | GPS points: relative distance, speed | Yes | No | No |
Zhou et al. [132] | Random forest with 3 layers | Speed, \(Acceleration-Gravity\), fast fourier transform (frequency domain), energy of the signals, sum of spectral coefficients | Yes | Accelerometer | No |
Zhou et al. [131] | Random forest | 85% percentile speed, average speed, median speed, medium velocity rate, high velocity rate, low velocity rate, travel distance | Yes | No | Yes |
Zhu et al. [134] | Auto encoder, deep neural network | Average speed, travel distance, average acceleration, head direction change, bus stop closeness, subway line closeness | Yes | No | Yes |
References | No. | Classes | Score | Metric | Validation |
---|---|---|---|---|---|
Kim et al. [60] | 15 | Work, Study, Shopping, Social Visit, Recreation, Home, Business Meeting, Change mode/Transfer, Pick-up, Drop-off, Meal/Eating break, Personal Errand/Task, Medical/Dental, Entertainment, Sport/Exercise | 98.68% | F1-Score | Out-of-bag-estimate |
Feng and Timmermans [40] | 10 | Study, Social Visit, Recreation, Home, Service, Paid Work, Daily Shopping, Non-daily Shopping, Help parents/cildren, Voluntary work | 96.8% | Accuracy | Out-of-bag-estimate |
Montini et al. [76] | 9 | Work, Shop, Service, Recreation, Home, Pick-up, Drop-off, Business Meeting, Other | 79.8% | Accuracy | Out-of-bag-estimate |
Xiao et al. [120] | 8 | Work, Study, Shop, Social Visit, Home, Eeating Out, Pick-up, Drop-off | 96.53% | Accuracy | Hold-out |
Bohte and Maat [21] | 7 | Work, Study, Shop, Social Visit, Recreation, Home, Other | 43% | Accuracy | n.p. |
Yazdizadeh et al. [122] | 6 | Education, Health, Leisure, Shopping/Errands, Home, Work | 72% | F1-Score weighted average | Manifold-cross-validation |
References | Person-day | Users | Ground truth | Observations | Time | Area | Smart-phone App |
---|---|---|---|---|---|---|---|
Yazdizadeh et al. [122] | 88,629 | 6845 | Validated-by-respondents (P2D) | 131,777 trips, 33 mln GPS points | 1 month collection period | Montreal | MTL Traject App |
Bohte and Maat [21] | 40,208 | 1104 | Validated-by-respondents | n.p. | 7395 days | Netherlands | GPS logger and Web based validation |
Kim et al. [60] | 7856 | 793 | Validated-by-respondents (P2D) | 22,170 days, 130 mln GPS points | 5–14 days per respondent | Singapore | Future Mobility Survey |
Feng and Timmermans [40] | n.p. | 329 | Validated-by-respondents (P2D) | 10,545 activities | 3 month per respondent | Netherlands (Rotterdam) | GPS logger and Web based validation |
Xiao et al. [120] | 2409 | 321 | Validated-by-respondents (P2P) | 7039 trips | 7–12 days per respondent | Shanghai | Shangai City - Smartphone Based Travel Survey |
Montini et al. [76] | n.p. | 156 | Validated-by-respondents | 6938 activities | 7 days | Zurich | Self Developed App |
References | Method | Main features | AGPS | INS | GIS |
---|---|---|---|---|---|
Bohte and Maat [21] | Rule-based | Distance \(GPS \rightarrow \,{\text {Points-of-interest}}\), Distance \(GPS \rightarrow LandUse\) | GPS | No | Yes |
Feng and Timmermans [40] | Random forest | Activity duration, activity start time, travel time to activity, distance \(GPS \rightarrow \,{\text {Points-of-interest}}\) | GPS | No | Yes |
Kim et al. [60] | Bagging decision tree, random forest | Activity probability, distance-based empirical probability, activity transition probability, activity duration | Yes | Accelerometer | Yes |
Montini et al. [76] | Clustering, random forest | start time, end time, GPS points density, age, education, income, mobility ownership, activity duration, walk percentage | Yes | Accelerometer | Yes |
Xiao et al. [120] | Multi layer perceptron, particle swarm optimisation, multinomial logit, support vector machines, Bayesian network | Age, gender, education, working hours, income, time of week, activity duration, time of day, transportation mode, distance \(GPS \rightarrow \,{\text {Points-of-interest}}\), distance \(GPS \rightarrow LandUse\) | Yes | No | Yes |
Yazdizadeh et al. [122] | Random forest | Features returned by Open Trip Plannera itinerary: GPS tracks average speed, time interval between the first and last GPS track of a trip, average distance between consecutive GPS point, attributes from, itinerary length, total transit time of each returned, total walking time of each itinerary, total waiting time of each itinerary, total travel time, number of transfers, walking distance, itinerary average speed attributes from GPS tracks, difference between GPS tracks length and itinerary length, overlapping percentage of itinerary and GPS tracks | Yes | No | Yes |
References | Mode | Category | Score | Metric | Validation |
---|---|---|---|---|---|
Chen and Bierlaire [26] | Walk, Bike, Car, Metro | Multimodal, global, shortest-path | [80%, 99%] | Path similarity indicator | n.p. |
Torre et al. [106] | Bicycle | Match when possible, build when needed | n.p. | n.p. | n.p. |
Quddus et al. [89] | Car | Unimodal, incremental, point-based | 99.2% | \(A = \frac{\#(correctly \,matched \,GPS\, points)}{\#(Total \,GPS\, points)}\) | n.p. |
Li et al. [68] | Car | Unimodal, incremental, point-based | 99.8% (sub-urban), 97.8% (urban) | \(A = \frac{\#(correctly \,matched\,GPS\, points)}{\#(Total\, GPS\, points)}\) | n.p. |
Wei et al. [115] | Car | Unimodal, incremental, shortest-path | 98% | Accuracy | n.p. |
Bierlaire et al. [19] | n.p. | Unimodal, global, shortest-path | [80%, 99%] | Path similarity indicator | n.p. |
Wu et al. [116] | Taxi | Unimodal, incremental, point-based | 93.58% | Prediction accuracy of next road by the road having the maximum probability | Hold-out |
Hunter et al. [52] | Taxi | Unimodal, incremental, shortest-path, supervised, unsupervised | 100% (1 s resolution), \(>90\%\) (30 s resolution) | Accuracy | Manifold-cross-validation |
Li and Wu [67] | Taxi | Unimodal, incremental, point-based | 87.18% | \(A = \frac{\#(correctly \,matched \,GPS\, points)}{\#(Total \,GPS \,points)}\) | Hold-out |
Jagadeesh and Srikanthan [54] | Dataset 1: Taxi. Dataset 2: n.p. | Unimodal, global, shortest-path | 91.3% | Average F-Score with: \(Precision = \frac{Length_{correct}}{Length_{matched}}\), \(Recall = \frac{Length_{correct}}{Length_{truth}}\), Input-to-output latency (Timelines) | Hold-out |
Newson and Krumm [78] | Car | Unimodal, incremental, point-based | 100% (1 s resolution), \(>90\%\) (30 s resolution) | \(Accuracy = 1 - E_L\), where \(E_L = \frac{(d_-+d_+)}{(d_0)}\), \(d_- =\) erroneous subtracted length, \(d_+ =\) erroneous added length, \(d_0 =\) length of correct route | Hold-out |
Lou et al. [71] | n.p. | Unimodal, global, shortest-path | \(A_N >81\%\) , \(A_L >87\%\) | \(A_N = \frac{\#(correctly\, matched \,road \,segments)}{\#(all \,road \,segments\,of\, the \,trajectory)}\), \(A_L = \frac{(\Sigma \,length \,of \,matched\, road \,segments)}{(length \,of \,the \,trajectory)}\) | Hold-out |
References | Links | Users | Ground truth (GT) | Observations | Area | Device |
---|---|---|---|---|---|---|
Quddus et al. [89] | 4605 | n.p. | 24-channel dual-frequency geodetic receiver | 4 h trajectories, 1 s resolution | London, sub-urban areas | GPS logger, gyroscope, odometer |
Li and Wu [67] | 583 | 12,000 | Hand match supported by Rule Based Algorithm | Training-set: 8678 GPS points (traces + syntetic from GIS), Test-set: 1334 GPS points (traces only), 10 s resolution | Beijing, urban areas | GPS logger |
Wu et al. [116] | n.p. | 442 + 13,650 | No GT available. Hidden Markov Models map-matching results as benchmark with [78] | 859,195 Traces, 3,709,666 Traces | Porto, Shangai | GPS logger |
Lou et al. [71] | n.p. | 189 | Validated-by-respondents (69 users only) | Dataset 1: Syntetic generated from road network (error normally distributed 20 stdev, 0 mean). Dataset 2: 28 GPS Traces (Trips) | Beijing | Geolife, [130] |
Chen and Bierlaire [26] | n.p. | 180 | No GT available. Unimodal map-matching result as benchmark | 10 s resolution | Lausanne (CH) Urban and outskirt areas | Nokia EPFL Lausanne [61] |
Hunter et al. [52] | n.p. | Dataset 1: 10. Dataset 2: 600 | Dataset 1: 1 s resolution GPS considered as high accuracy GT. Dataset 2: no GT | Dataset 1: 700,000 GPS points, 1 s resolution. Dataset 2: 600,000 points, 1 min resolution | S. Francisco | Mobile Millennium system—GPS logger |
Jagadeesh and Srikanthan [54] | n.p. | Dataset 1: 21,807 GPS points, 20 trips, 421 km. Dataset 2: 1000 trips, 13,139 km | Dataset 1: Manual Check on Map-matched GPS points from higher accuracy source (smartphone), leveraging on knowledge of taxi route. Dataset 2: User validation | Dataset 1: 21,807 GPS points, 20 trips (TAXI), 421 km, 1 s resolution. Dataset 2: 13,139 km, 1000 trips. Dataset 3: Syntetic Dataset adding noise to Dataset 1 | Singapore | Dataset 1: Custom Smartphone App (Android), Dataset 2: Commercial Smartphone App |
Newson and Krumm [78] | n.p. | 1 | Route planned before data collection and hand match | 7531 GPS points, 80 km, 1 s resolution, degraded data simulation | Seattle | GPS logger |
Bierlaire et al. [19] | n.p. | Dataset 1: 1 users. Dataset 2: 3 users. | Dataset 1: Known true path. Dataset 2: no ground truth. Dataset 3: high accuracy GPS device | Dataset 1: 10 points Dataset 2: 25 trips 1041 GPS points, 10 s resolution | Lausanne (CH), Urban and outskirt | Nokia EPFL Lausanne [61] |
Li et al. [68] | n.p. | n.p. | Tightly-coupled carrier phase GPS receivers integrated with a high-grade inertial navigation system | 3363 epochs (sub-urban), 2399 epochs (urban), resolution: 1 epoch/s | Nottingham rural sub-urban, Central London | GPS logger, digital elevation model |
Torre et al. [106] | n.p. | n.p. | n.p. | 128 GPS Traces, 185,000 GPS points, 360 km, 1088 min | Minneapolis (Twin Cities) | Cyclopath Android App |
Wei et al. [115] | n.p. | n.p. | n.p. | 14,436 GPS points (SIGSPATIAL Cup 2012 DS), 19,080 GPS points, 1 s resolution | Seattle Shanghai |
References | Method | Main features | AGPS | INS | GIS |
---|---|---|---|---|---|
Quddus et al. [89] | Fuzzy logic, extended Kalaman filter | Speed, heading error, perpendicular distance, horizontal dilution of precision | 12-channel single frequency high sensitivity GPS receiver | Dead-reckoning | Yes |
Li et al. [68] | Rule based, extended Kalaman filter, integrity check | Altitude, longitude, latitude, traffic flow directions, road curvature, grade separation, travel distance, heading | GPS | Dead-reckoning | Yes |
Bierlaire et al. [19] | Probabilistic | Timestamp, longitude, latitude, speed, heading, horizontal error Std. Dev., network error Std. Dev. | Yes | No | Yes |
Li and Wu [67] | Feed forward neural network | Longitude, latitude, timestamp, heading | GPS | No | Yes |
Lou et al. [71] | Mixed method: topological, geometric, probabilistic | Distance GPS(t) \(\rightarrow\) GPS(t + 1), distance GPS \(\rightarrow\) network, shortest path between candidate points on network, average speed | Yes | No | Yes |
Torre et al. [106] | Hidden Markov model, Viterbi | Distance GPS \(\rightarrow\) node, maximum out-degree of the transportation graph | Yes | No | Cyclo-path map |
Wei et al. [115] | Global Max-weight, hidden Markov model, Viterbi | Fréchet distance, shortest-path | GPS | No | Open Street Map |
Chen and Bierlaire [26] | Probabilistic | Transport mode, distance, speed, acceleration | Yes | Accelerometer, Bluetooth Low Energy | Yes |
Wu et al. [116] | Recurrent neural network, long short term memory | Longitude, latitude, timestamp, destination | GPS | No | Open Street Map |
Newson and Krumm [78] | Hidden Markov model, Viterbi | Distance GPS(t) \(\rightarrow\) GPS(t + 1), distance GPS(t) \(\rightarrow\) network (only in range < 200 m) | Yes | No | Yes |
Hunter et al. [52] | Undirected graph Bayesian network, Viterbi | Path length, distance point projection \(\rightarrow\) GPS, number of signals, number of turns, average speed, Max/Min Num. lanes | GPS | No | 560,000 links map |
Jagadeesh and Srikanthan [54] | DS 1: hidden Markov model, Viterbi, conditional random fields (CRF). DS 2: multinomial logit model, k-shortest path with link-penalty approach | Path choice: free-flow travel time (s), number of traffic signals, average road class, number of class changes | AGPS, with WiFi and GPS off | No | Yes |