RetinaMOT: rethinking anchor-free YOLOv5 for online multiple object tracking

We used the state-of-the-art target detection framework YOLOv5 and modified it to an anchor-free detection mode following the method of CenterNet [95], YOLOv5 is a work proposed by [31]. Moreover, as of November 22, their work still has not stopped, the repository is still kept up to date, the algorithm is structured by YOLOv4 [7], using a series of improved methods to bring the speed and accuracy of target detection to new heights, while the YOLOv5 wave has set off, there are also many variations based on YOLOv5 variants: YOLOX [24], YOLOv6 [35], YOLOZ [5], there are currently many trackers that work in conjunction with the YOLOX framework, and it can be said that this has become a mainstream, But we bucked the trend using the YOLOv5 architecture, and we were able to achieve excellent tracking performance by simply modifying the model skeleton and anchoring strategy.

For the motion model estimation, we use the Kalman filter algorithm [75]. The position of motion information in video frames is estimated by Kalman filtering. When the feature extraction network cannot extract the object information, the Kalman filter can also predict the tracked object’s position. The correlation matching between the ID and the object is then done by the Hungarian algorithm [63].

Our algorithm’s Kalman filtering algorithm is the most basic without optimization. There is no treatment of nonlinear interference or exogenous interference between matched and unmatched targets. Our work focuses on constructing an excellent feature extraction framework. Constructing robust nonlinear predictive filtering systems is our next goal. We carefully studied some excellent algorithms for solving real-world problems, such as [84] constructing an intelligent controller based on disturbance suppression based on constrained nonlinear systems, which fully considers internal and external disturbances and various constraints and compensates for system perturbations in the absence of accurate system modeling. The disturbances are compensated by tracking errors and state estimation errors, and then the network weights are updated, which improves functional performance in cases where the system modeling is not accurate. Moreover, excellent results were achieved in the experiments. Yang [82] proposed two integral robust control algorithms. They ensured tracking performance by cleverly introducing filtering errors and asymptotic filters for the same class of simultaneous matched and unmatched disturbance effects. They used three methods to integrate over the inversion framework and iterate the filtered signal inversion, extremely and cleverly ensuring the asymptotic tracking performance. Yang et al. [83] used a new inversion method without adding any additional information about the internal dynamics and other physical parameters, integrating a robust nonlinear controller based on a continuous control method with disturbance compensation. High-performance tracking control of nonlinear systems with matched and unmatched modeling uncertainties was obtained. Their excellent work gives us ideas for the subsequent filter optimization.

The performance of the tracker and detector is most tested in exceptionally crowded scenes. Many trackers add the ReID model to improve the reidentification ability of the detector [17, 37, 91] so that the network can be trained simultaneously. Inference can extract features and targets at the same time. However, the ReID network is not used in our model. ByteTrack’s study found that the feature extractor can replace ReID to handle long-term associations between objects as long as the feature detector progress is high enough. It was also found that adding ReID did not improve the tracking results [89].

The advantage of YOLOv5 is its one-stage structure, which can extract pedestrian target features quickly and accurately. Therefore, we cancel the anchor profile and change the neck configuration in the network to three upsampling and three C3 operations, combined into one and output to the head for the prediction of heatmap, width, height, and offset, which is consistent with CenterNet.

As seen in Fig. 3, the Focus module is used at the beginning of the network, which downsamples the image utilizing array slicing by four array slices. firstly, by four array slices, respectively, The top left, top right, bottom left, and bottom right corners of the image are cut to preserve the feature information, fully preserve the semantic information, and expand the perceptual field. The number of channels is quadrupled by stitching the four feature maps according to the dimension, and finally, the number of channels is restored using convolution.

After the Focus module, the network alternately uses the Cross-CPC Attention (described in detail in the “Cross-dimensional coordinate perception cross-attention” section) and DeSFPP modules (described in detail in the “Adaptive cascade feature alignment pooling pyramid” section), allowing the BackBone to expand the field of sensation and fully extract feature location information. In the last part of BackBone, DeSFPP is used for upsampling features from different global pools, thus enabling the network to adjust the sampling position of the network. The neural network learns the offset of the occluded pedestrian to locate the pedestrian’s moving position accurately.

The BackBone information is output to the feature pyramid to extract the multi-scale pedestrian information. Lower-level semantic features are integrated with higher-level semantic features through top-down horizontal connections so that each layer has equally rich semantics [71]. By establishing the high-level semantics for the occluded pedestrian and the near and far pedestrian targets, each scale feature can be extracted effectively to improve the tracking accuracy and reduce the number of target identity switching. Finally, the fused features are sent to Head, which outputs an array containing HeatMap information, coordinate offset information, and target width and height prediction information.

It can be seen in Fig. 2. The RetinaMOT method proposed in this paper has promising results on the mainstream tracking methods, achieving the best tracking accuracy with limited and surpassing the mainstream methods at a reasonably fast inference speed.

In traditional 2D convolution, feature extraction is often performed using a symmetric rectangular window. A standard traditional convolution kernel for feature extraction is uneven in spatial distribution, especially in height occlusion and object pixel deformation, with greater weight at the center intersection of spatially high-resolution images and high similarity in the way height occlusion targets are learned [14]. To enhance the diversity of features and avoid the interference of image noise and uniform weight distribution, we extract uniform weight features using horizontal and vertical convolutional kernels of different shapes. After the convolutional superposition operation is applied to the correction module, it can effectively capture the global information of pedestrian spatial location. The diversity features extracted by the horizontal and vertical convolutional kernels are fused with the weighted features [66], thus allowing the sample features to expand the convolutional perception domain and establish long-term spatial dependence around the occluded target.

At the beginning of the design, we referred to a series of sampling features and working cells of the human retina. The first is the photoreceptor in the retina, which has two components: 1. Rod cells. 2. Cone cells.

Rod cells in the retina are a type of cell in the retina that is comparable to the cone cells, mainly distributed around the center of the retina, and more sensitive to light than the cone cells; its function is mainly to perceive changes in light and dark, but the ability to distinguish light is poor, and it can only perceive the general color and shape of objects for distinguishing contours [58]. According to the function of Rod cells. We use averaging pooling in the entry part of the Photoreceptor Module for averaging only the feature points in the neighborhood, filtering their foreground information, and keeping the background information. The gradient is averaged according to the neighborhood size and then passed to the index position [41] according to the cell’s characteristics. With an input image \( X_{1} \) in the network, using an average pooling filter size of \( r \times r \) and a step size of r, we get \( \mathcal {T}_{1}= \) AvgPool \( _{r}\left( {X}_{1}\right) \), after which the features are extracted using \( 3 \times 3 \) convolution to get \( {K}_{1} \), and then the pooled features \( \mathcal {T}_{1} \) and the convolution are feature mapped, \( {X}_{1}^{ \prime }={\text {Up}}\left( \mathcal {F}_{2}\left( \mathcal {T}_{1}\right) \right) ={\text {Upsample}}\left( \mathcal {T}_{1} * {K} _{1}\right) \), \( {\text {Up}}(\cdot ) \) represents the linear interpolation function that maps the small-scale space features to the original feature space. The design is as shown in Fig.  4.

The role of cone cells in the human retina is to recognize color and object details. The cone cells have conical ends with pigments that filter incident light, thus giving them different response curves [18, 20]. Better performance in better visual conditions (when the light is good and the line of sight is sufficient). Based on these cellular properties, the we first input the image \( X_{1} \) to the convolution of \( 3 \times 3\) to obtain \( {K}_{2}\), denoted as After this series of operations, it can be expressed as \( {Y}_{1}^{\prime }=\mathcal {F}_{3}\left( {X}_{1}\right) \cdot \sigma \left( {X}_{1}+{X}_{1}^{\prime }\right) \) where \(\mathcal {F}_{3}\left( \textrm{X}_{1}\right) =\textrm{X}_{1} *{~K}_{2} \), where \( \sigma \) is used to denote the Sigmoid function, and then the dot product operation is performed. The convolution output of the feature input \(3\times 3\) after the dot product operation is \( {K}_{3} \), \(\textrm{X}_{1}{ }^{\prime } \) is the residual established to establish the weight of the input features and the two parts of the output features, cone cell module is designed in the way as Fig.  4. Finally the whole Photoreceptor Module output is expressed as \( {Y}_{1}=\mathcal {F}_{4}\left( {Y}_{1}^{\prime }\right) ={Y}_{1}^{\prime }* {K}_{3} \).

Secondly, there are two essential cells in retinal cells: 1. Horizontal cell, 2. Bipolar cell. The role of the horizontal cell is to adjust the luminance of the visual signal output from the photoreceptor cells and to achieve luminance adaptation of vision.

Moreover, enhance the contrast of the edge of the visual image and highlight the contour lines of the scene [32]. Bipolar cells (bipolar cells) receive signal input from photoreceptors and transmit it to anaglyph cells and ganglion cells after integration. In retinal signaling, bipolar cells show an essential function: shunting visual signals into ON and OFF signals [23, 85]. We designed two filters named Horizontal Filter and Vertical Filter according to the characteristics of the horizontal cell and bipolar cell; respectively, we refer to  [15] their work as shown in  5 in which they are the convolution kernels of \(N \times 1\) and \(1 \times N \), respectively. Given the horizontal and vertical filters, the vertical and horizontal striped regions allow the encoding of features. They are establishing dependence on the discrete distribution of the area around the pedestrian target.

We take advantage of the superimposability of convolution [43] (the same input features are sampled at the same downsampling rate using a compatible 2D convolution kernel. Generate feature maps of the same size.) as in the Eq.  (1). Add the rich features extracted by the Horizontal Filter and Vertical Filter to the Photoreceptor. Module, a new equivalent convolution kernel can be added, that is, the additivity of the two-dimensional convolution in the Horizontal Filter, Vertical Filter, and Photoreceptor. The additivity of the two-dimensional convolution holds for the convolutional kernels of the Horizontal Filter, Vertical Filter, and Photoreceptor Module. We can reuse the smaller Horizontal and Vertical Filter features for the Photoreceptor Module. Formally, this transformation is feasible on layers p and q. If the bar convolution satisfies the formula  (2), it is possible to obtain an equivalent convolution with the same output size as the Photoreceptor Module with the same output size.In Eq. (1), \( \varvec{I}\) is a matrix, \( \varvec{K}^{(1)}\) and \( \varvec{K}^{(2)}\) are two convolution kernels with appropriate sizes, \( \oplus \) denotes the summation operation, and in Eq. (2), \( \varvec{M}^{(p)} \) denotes the Photoreceptor Module’s output graph, \( \varvec{M}^{(q)} \) denotes the output graph of two bar convolutions, and H, W, D denotes a convolution of size \( H \times W \) with D number of channels. Retina-Conv is a feature summation operation for the three branches according to the convolutional feature summation. In this way, we can implement feature reuse. Enhance deep semantic information of crowded pedestrian targets to greatly improve tracking accuracy.

To fully validate the effectiveness of Retina-Conv, we did not use Retina-Conv convolution directly on the trace dataset when we designed it because we were thinking of easily embedding Retina-Conv into other modules of the whole network. We analyzed its theoretical feasibility and experimented on two classic classification datasets, Cifar10 and Cifar100. Our experiments were compared without any tricks (such as hyperparameter adjustment, learning rate adjustment, data augmentation, etc. ), using an Nvidia RTX 3070 8 G GPU; the learning rate was always kept at 0.01, BatchSize was always kept at 12, the optimizer selected random gradient descent, and the loss function used the cross-entropy loss function. The specific experimental data are shown in Table 1.

The RetinaConv designed in this paper can improve the feature extraction ability for various classic networks and can be easily embedded into each network, laying the foundation for improving the tracking accuracy later.

For intensive pedestrian prediction tasks, locating individual pedestrian targets for accurate spatial localization requires rich details of spatial features. When dealing with intensive pedestrian detection tasks, one of the key obstacles to overcome is effectively utilizing feature layers at different scales.

Initially, the spatial pooling pyramid (SPP) was proposed by [26]. The spatial pooling pyramid maximizes the pooling of input feature maps at different scales using different pooling kernels. The advantage of this is that the images are pooled into a deeper network, and the robustness of the detection model is improved. Similar methods were proposed by [9], proposed atrous spatial pyramid pooling (ASPP). They used hole convolution to obtain a larger perceptual field and make the image feature map size without losing too much edge detail, increasing the perceptual ability of feature maps at different scales.

At present, SPP is mainly used, and the native backbone of YOLOv5 also uses SPP in the final stage of the network to make pooling kernels of size 5, 9, and 13 in each grid of the feature map. However, the size of the feature map is the same after the maximum pooling of different pooling kernels. However, the more significant the pooling kernel is after maximum pooling, the smaller the feature map is, and the size is the same. However, many areas are filled using this approach, which is equivalent to the most specific upsampling feature, where the feature map is directly summed with the local feature residuals, as shown in Fig. 2, where SPP uses the feature map of a smaller pooling kernel and directly catches the feature map of a larger pooling kernel, leading to the problem of contextual misalignment after feature mapping. Indirectly, the location localization fails in the dense pedestrian dataset, making the width-height prediction of the target biased [48]. In addition, the ability of SPP to extract deep high-level semantic information is powerful. However, it mainly relies on stacking the giant pooling kernel in a convolution operation. Hence, his perceptual field is minimal, and there is much room for improvement in scenarios with complex target-tracking datasets (Figs. 5, 6).

Based on the above limitations, we propose a shape-shifting feature alignment pooling pyramid with a detailed design as shown in Fig. 8, and extensive experimental results show that the pooling operation loses most of the information in the image while pooling, which degrades the performance of the whole network. To minimize the loss of information during the pooling operation  [64], we first compress the feature maps into multiple pooled feature maps of \(N \times N\) size using adaptive averaging pooling and then fuse the feature maps in both spatial directions with semantic information, capturing their information-dependent associations on a particular space. This preserves the location information of dense pedestrians, which helps the detection network to better localize the occluded targets.

We designed six structures to perform the SPP analysis step by step. In the data experiment on the MOT17-half dataset, A indicates that the maximum pooling of SPP is directly replaced with adaptive average pooling (pooling kernel sizes are 1, 2, 3, 6); B indicates that the maximum pooling of SPP is retained and adaptive average pooling is added after the final merge (pooling kernel sizes are 1, 2, 3, 6); C indicates that each maximum pooling of SPP followed by C denotes adding an adaptive average pooling (pooling kernel size 1, 2, 3, respectively) after each maximum pooling of SPP; D denotes replacing the maximum pooling of SPP directly with adaptive average pooling (pooling kernel size 1, 2, 3, 6, respectively) and adding the hole convolution for merging; E denotes replacing the maximum pooling of SPP directly with adaptive average pooling (pooling kernel size refer to the maximum pooling size of SPP: 5, 9, 13).

After our extensive experiments, we proved that the SPP-A module in Table 2, for dense pedestrian prediction replaces the maximum pooling with adaptive averaging pooling directly, which can improve the tracking accuracy slightly, and the number of identity switches is reduced by 22 times. Compared to SPP-B, the tracking accuracy has no change, and the IDs of these metrics are decreased, while in the SPP-C structure, the features after maximum pooling are then connected to an adaptive average pooling to improve the tracking accuracy, indicating that the maximum pooling in SPP has a negative impact on the tracking task. We remove the maximum pooling along this line of thought and only use adaptive average pooling to compress the feature map to \(N \times N \), and add in SPP-E, we design the size of the feature map output from adaptive averaging pooling according to the size of the maximum pooling kernel of SPP, removed the hole convolution (multiple hole convolutions increase the inference cost in experiments), and used only the simplest adaptive averaging pooling kernel to make the tracking metrics based on the SPP-E, we introduce the feature pyramid design and design the DeSFPP structure by merging the top-down compressed feature maps with adaptive average pooling outputs of 5, 9, and 13, using retinal convolution to extract features before using adaptive average pooling and then merging all layers afterward (Fig. 7).

When designing DeSFPP, the design of the feature pyramid refers to the FaPN feature aggregation module proposed by  [30]. When the input image enters the module, features are first extracted, and channel reduction is performed using a retinal convolution of \( 3 \times 3\). Then multiple adaptive pools are performed using a cascaded fusion of FaPN, a step we call cascaded AdaFPN. And then feed an adaptive averaging pooling block with different output sizes to output feature map sizes of \( C_{5} \in \mathbb {R}^{C \times 5 \times 5}\), \( C_{9} \in \mathbb {R}^{C \times 9 \times 9}\), \( C_{13} \in \mathbb {R}^{C \times 13 \times 13}\), arranging them from the bottom up, and each layer of the feature map in into the feature pyramid. It is first used in the feature selection module to extract the target details on the high-resolution spatial map. It is mapped to the location information of the feature map to extract meaningful information and self-calibrate. Equation  (4) shows the feature selection module process.

For the pooling layer size in DeSFPP, we did many experiments to show why we chose [5, 9, 13] as our final choice. In Fig. 7. (we scaled all the IDs by a factor of 10 to keep the metrics in the same magnitude), we can see that the larger the pooling layer is, the more significant the impact on the tracking metrics is, When the input pooling layer size in the pooling pyramid is [88, 94, 100], the indicator has dropped very low. Hen the size is smaller than the [5, 9, 13] group, the impact is small, but all things considered, the [5, 9, 13] group is the optimal choice.\( f_{s}(\cdot ) \) denotes the significant feature selection operation (using \( 1 \times 1\) convolution kernel a BatchNorm), representing the input features as \( \textbf{C}_{i}=\left[ \textbf{c}_{1}, \textbf{c}_{2}, \ldots \right. \left. \textbf{c}_{D}\right] \), \(\textbf{D}\) means it is the ith channel of the input feature, \( \textbf{u}=\left[ \textbf{u}_{1}, \textbf{u}_{2}, \ldots \right. \left. \textbf{ u}_{D}\right] \) denotes the feature weights. The more important features are weighted more. \( f_{m}(\cdot ) \) denotes the significant feature construction layer (using \( 1 \times 1\) convolution kernel connected to a Sigmoid), \( \hat{\textbf{C}}_{i} \) denotes the output feature, \(z = \textbf{AdaPooling}(\mathbf {{C}_{i}})\) [78], z is computed in Eq.  (4)

The salient features processed by the feature selection module are passed into the FAM module. Due to the frequent use of adaptive averaging pooling, the feature maps are all compressed tiny. At this point, the feature map sizes after different pooling are presented bottom-up, and the feature maps of different sizes are aggregated using FAM. As shown in Eq.  (5).where \( f_{o}(\cdot ) \) denotes the feature space difference function, \( f_{a}(\cdot ) \) denotes the learning offset information function from the feature, \( P_{i}^{u}\) is obtained from \( C_{i-1} \) after upsampling, \( \left[ \hat{C}_{i-1}, P_{i}^{u}\right] \) is concatenated to obtain spatial difference information \( \varvec{\Delta }_{i}\), and \( \varvec{\Delta }_{i}\) with \(P_{i}^{u} \) using deformable convolution [11] via bias learning, which is used to align the top-down feature map features, as specified in Figs.  8 and 9.

After going through the cascade AdaFPN module, the multi-scale features of feature maps with different pooling sizes have been fused. In the other branch, the feature map with the input \(C\times H \times W\), the position information in space is extracted separately on each channel to get the weight information of features in wide dimension and high dimension, corresponding to the feature position importance, and then the current position and adjacent features are adjusted by \( 3 \times 1 \) and \( 1 \times 3 \) convolution, and then upsampled to \(C \times H \times W\) size. After \(1 \times 1\) convolution and sigmoid to obtain the weights, the output feature map is obtained. Each location in the output tensor is related to the pedestrian location information in the input tensor. By re-fusing with cascade AdaFPN, the long-term dependency of the whole scene can be constructed. It not only captures rich scene information but also helps greatly in tracking accuracy cues.

In recent years, attention mechanisms have been receiving attention in the field of vision. However, universal channel attention tion such as SE Net [29] and ECANet [69] only consider channel information modeling to weigh the importance of each channel of the feature without considering the intersection of space, channel, and location, which results in limited feature semantic information being generated. Therefore, a large amount of rich information is ignored in the target tracking process, and as a result, a large number of tracked targets perish. Based on the above research, we propose cross-dimensional coordinate-aware cross-attention. We are inspired by the research of [50] and  [28], who designed three branches to capture the cross dimension. However, in the third branch, we obtain the tensor (C, H), (C, W), and (H, W) dimension dependencies between them and use the Z-Pool module to concatenate the average pooling and complete pooling information. This approach is challenging to map the global spatial information into the channel features, resulting in the location information not being sufficiently mapped with other dimensional information.

We design three branches based on the above characteristics and defects, passing the tensor \( X_{1} \in R^{C \times H \times W} \) into the three branches simultaneously. The first branch transforms the tensor X into \( \hat{X}_{1} \) by doing the rotation operation along H direction. Its shape is \( \hat{X}_{1} \in R^{W \times H \times C} \), the output tensor \( \hat{X^{*}}_{1} \in R^{2 \times H \times C} \) after class pooling by the Z-Pool module. Then the Retina-Conv using \( 7 \times 7 \), BatchNorm provides an intermediate output and passes in the SiLU activation function. A linear correction is performed. Finally, \( \hat{X^{*}}_{1} \) is recalibrated to \( X \in R^{C \times H \times W} \) using the Sigmoid activation function to generate the attention weights after the inverse rotation. Eventually, an interaction is established between the H and C dimensions. The detailed design is shown in Figs. 10 and 11.

Similarly, the tensor \( X_{2} \in R^{C \times H \times W} \) is passed to the second branch to transform the tensor X into \( \hat{X}_{2} \) by doing the rotation operation along W direction. Its shape is \( \hat{X}_{2} \in R^{H \times C \times W} \), which is processed by the Z-Pool module for The output tensor \( \hat{X^{*}}_{2} \in R^{2 \times C \times W} \) after the class pooling process, then the Retina-Conv using \( 7 \times 7 \), then the BatchNorm provides an intermediate output which is passed into the SiLU activation function [19] for linear correction. Finally, the \( \hat{X^{*}}_ {2} \) is recalibrated to \( X \in R^{C \times H \times W} \) using the Sigmoid activation function to generate attention weights and then inverse rotation. Eventually, an interaction is established between the C and W dimensions.

Finally, the third branch, passed into the tensor \( X_{3} \in R^{C \times H \times W} \), is first encoded for each channel using horizontal and vertical bar pooling, and after vertical pooling, the output is.Then perform the same operation for W and channel C.Aggregating the features in both spatial directions greatly helps the stability of target tracking. Convolution and BatchNorm and H-Swith linear correction operations [33] are performed afterward.After the subsequent channel reduction and dimensional decomposition, the final output is obtained as \( \hat{X}_{3} \in R^{W \times H \times C} = F(X_{3}) \). After getting three components, \( \hat{X}_{1}, \hat{X}_{2}, \hat{X}_{3} \) the three feature components are converged together, and then a weight balance is performed, and finally, the feature information across dimensions is interacted with features to get deep semantic features and significant attention.In our experiments, the good results of cross-dimensional coordinate aware cross-attention are because the detection model considers the intersectionality of space, channel, and location simultaneously, so it captures more features for complex and dense scenes, allowing the target tracking network to perform better. In Fig.  10, we used the Grad-CAM [60] technique on ResNet50 to visualize it, this technique converts the gradient of the input image into a color image, and the darker the color indicates the greater attention to the region. The visualized image can help the understanding of cross-dimensional coordinate aware cross-attention. We use ResNet50 [27] because we want to see how the network focuses region changes after adding our attention, and the simple and practical structure of ResNet50 facilitates our observation and training. Our proposed attention can be seen in Fig. 10, compared to SENet and coordinate attention cross-dimensional coordinate aware cross-attention can focus a larger area, especially for the MOT series pedestrian dataset, where the area to be focused is scattered and requires complete map focus for this particular case.

DataSets Our ablation experiments are trained and evaluated on MOT17-half, MIX-MOT17-half dataset, MOT17-half is the MOT17 [49] dataset half divided into training set half divided into test set, MIX-MOT17-half indicates that the training set is composed of half of the MOT17 sequence and Caltech [16], CityPersons [88], CUHK-SYSU [79], PRW [92], ETH [21] datasets, the comparison experiment used MOT16 [49], MOT20 [12], CrowdHuman [62]datasets The MOT17, and MOT20 datasets are composed of various street or mall scenes, consisting of many sequences with different lighting, pedestrian density, static and dynamic camera shot videos, and different styles of people [42]. The test set results must pass the official Multiple Object Tracking Benchmark provided by mot challenge, the website provides the corresponding evaluation videos, and various strategies are used to avoid overfitting.

Metrics When evaluating the dataset, the following metrics are of primary interest, starting with MOTA (Multi-Objective Tracking Accuracy), which has been used as the primary evaluation metric in the field of multi-objective tracking and can be obtained from the Eq. (10).where GT denotes the number of accurate labels in the dataset labels, IDS denotes the number of target identity switches, FP and FN denote the negative samples predicted by the model as positive class FN: positive samples predicted by the model as negative class [22], respectively.

The next is the IDF1 (identification accuracy) metric, as in Eq. (11), where IDFP, IDFN, and IDTP are identifying false positive matches, false negative matches, and true positive matches.Finally, the IDS indicator represents the number of identity switches of the target during the tracking process; the lower the number, the better the effect.

We used YOLOv5 as our backbone, and after all the modification work, We added all the above modules to the native Yolov5 framework. We retrain the COCO dataset in Anchor-base [38] to obtain pre-trained weights for the backbone. All hyperparameters are used in the default Yolov5 parameters except Batch-size, which is set to the explicit memory limit, and 300 epochs are trained on the COCO dataset. To prevent the use of pre-training weights from causing early convergence of the network resulting in inaccurate accuracy, no pre-training weights are used in the ablation experiments, and pre-training weights are used in the comparison experiments. The MIX dataset (including MOT16, MOT17, MOT20, Caltech, CityPersons, CUHK-SYSU, PRW, ETH, and CrowdHuman) is then uploaded to the mot challenge private evaluator. For the ablation experiments, we used Nvidia RTX 3060 12 G graphics cards, and for the comparison experiments, we used two Nvidia RTX A6000 48 G graphics cards for training.

We used the designed RetinaYOLO as our feature extraction backbone. It was trained on the COCO dataset without loading any pretraining weights to obtain new pretraining weights. The data enhancement used typical hue, saturation, rotation angle adjustment as well as mosaic and mixup, the image input size was set to 640 \(\times \) 640, the initial learning rate after training the COCO dataset was used to obtain the pretraining weights of RetinaYOLO for the backbone, and the MIX dataset was used for training. The input dataset size was 1088 \(\times \) 608 during training, and the network learning rate was 0.0001. The inference parameters of the tracking algorithm are set as follows: the tracking confidence threshold is 0.4, the detection threshold is 0.3, the nonmaximum suppression threshold is 0.4, the number of unmatched tracking frames is initially 30, and it is set to 100 frames to slightly improve the IDF1 index. To identify the lost target again in more subsequent video frames, if the detection frame and tracking target frame IOU are less than 0.2, the match is rejected, and the tracking accuracy is unchanged if it is set to 0.1. However, the number of target identity switches will increase. After training the MIX dataset to fine-tune the MOT series dataset, the learning rate step is set to 15 epochs, adjusted once to reduce by ten times, and trained again for 20 epochs. It is found that MOT15 and MOT20 both have a slight increase, but there is no change in MOT17 and MOT16.

In this section, we scrutinize all the methods mentioned above and conduct rigorous experiments on four valid methods in RetinaMOT, including RetinaConv, Cross-CPC, DeSFPP, and ByteTrack (Table 4).

Componentwise analysis In this subsection, we explore the effectiveness of individual components in the network. To conduct the experiments fairly and objectively, we do not use pretrained models in the ablation experiments. However, we use half of the MOT17 dataset for training and half for evaluation. The 30th calendar hour of training is used as the evaluation criterion. The baseline component ablation experiments are shown in the upper part of Table 3. It can be seen that individual components are embedded directly into the baseline without using any tricks to obtain performance gains. Adding RetinaConv in \({}{\textcircled {2}}\) and adding DeSFPP in \({}{\textcircled {4}}\) both can get more than 3.0 MOTA value gains. In \({}{\textcircled {4}}\), the direct embedding of DeSFPP we propose in \({}{\textcircled {4}}\)can significantly reduce the value of IDs and significantly improve the IDF1 metric. Our proposed Cross-CPC attention in \({}{\textcircled {3}}\)can directly increase MOTA and IDF1 metrics and reduce the number of IDs by 220. In summary, all our proposed single modules can improve the tracking series metrics exceptionally well.

Self-relation and cross-relation analysis To verify the effectiveness of component self-relations and component cross-relations in our tracking framework, we used several experiments involving cross-components with each other to explore the changes in their tracking metrics. In \({}{\textcircled {14}}\), a mixed cross-evaluation of all components can obtain metric gains of 9.6% for MOTA, 8.9% for IDF1, and 245 reductions in IDs based on the baseline; while compared to single components it can be found that some components are limited individually and cross-evaluated, For example, in \({}{\textcircled {4}}\), adding the DeSFPP component alone can achieve a MOTA gain of 3, but in \({}{\textcircled {11}}\), the three components together can only achieve a MOTA gain of 3. Therefore, they must work together in RetinaConv to achieve maximum performance. After we ported ByteTrack to our framework, we found that in both the component self-relations and component cross-relations experiments, we can see that directly adding ByteTrack gives approximately 2% MOTA gain and approximately 2% ID1 gain and results in a significant reduction in IDs.

because we only use the 30th epoch as the standard epoch for evaluation. Therefore, some components do not converge. Therefore, we increase the training dataset to obtain fuller and richer features, and we use MIX-MOT17-half. In the results, we can see that the improvement of \({}{\textcircled {2}}\) and \({}{\textcircled {3}}\) is slight. Only a 0.5% MOTA gain is obtained. However, this leads to a reduction in IDs, which indicates that the joint action of these two components makes pedestrian tracking more stable. It can better discriminate highly occluded targets and retrack them, but the MOTA gain of the component that gains more feature weights is limited under the joint action. Experiments in the \({}{\textcircled {4}}\) and \({}{\textcircled {5}}\) groups showed that the continued addition of other components again led to an improvement in the MOTA index, resulting in a significant reduction in the number of pedestrian identity transactions.

We found that our components can improve performance in self-relational verification and multicomponent cross-validation from ablation experiments.

For comparison experiments, we embed all the proposed modules into the YOLOv5 framework and train and validate the performance of our model on the MS COCO dataset, with training parameters following the default parameters of YOLOv5. For fairness and objectivity of the comparison, no pretraining weights are used for training, and all 300 epochs are retrained. In Fig. 12, the change of metric information during our training is shown. We use the YOLOv5-S model as our Baseline and compare all the state-of-the-art lightweight frameworks in the last two years, as seen in Table 5. Our RetinaYOLO adds only 1.7 M parameters and 1.2 G FLOPs to the YOLOv5-S model but increases the \( \textrm{mAP}^{test}_{0.5:0.95} \) metric by 6.5% and the \( \textrm{mAP}^{val}_{0.5:0.95} \) metric by 5%. If we use the YOLOv6 -T model for comparison, we have 6.1 M fewer RetinaYOLO parameters and more than two times less computation. But the \( \textrm{mAP}^{val}_{0.5:0.95} \) metric is going to gain 1.4% more. It can be seen in table 5 that we outperform all the lightweight frameworks of the YOLO family in terms of performance, and we are in the middle of the pack in terms of several parameters and complexity, balancing the speed and accuracy paradoxes extremely well.

After training the MS COCO dataset, we load the best round of pretrained weights in the tracking framework. To make the results fair and just, we use the same training data and evaluator to train the complete MIX dataset, which contains the training sequences of MOT20, MOT17, and MOT16. To make the controller results in the comparison experiments fair, the algorithm results of Table 6 are derived from the online evaluator of MOT Challenge, the most famous and unbiased evaluation site in the field of target tracking, where target tracking researchers around the world are eager to see new records. The algorithms we compared were all developed by MOT Challenge. The algorithms we compare are all published by the MOT Challenge private detector and are trained using an additional dataset (Crowd Human + MIX dataset). In the detector part of all compared algorithms, various large pretrained datasets are used, such as ImageNet 1K and ImageNet21K [13]), while our detector is trained only on the COCO dataset, which is much smaller in size than the other detectors, and the target tracking paradigm is based on the TBD and JDE paradigms. However, in terms of actual performance, we can compete with them with SOTA performance.

As seen in Table 6, we stand out among all state-of-the-art tracking methods on the MOT16 dataset. Our tracking accuracy is slightly higher than that of GSDTv2 [73], but more than 13 times faster, with 125 fewer IDs and a 5.1% higher IDF1 metric. Similarly, in the current latest publicly available dataset MOT20, compared to the [36] and [98] methods, all of our metrics are higher than the LCC tracking method. The RekTCL method, on the other hand, is faster than ours and has the highest IDF1 metric among all compared methods with a 2.3 times higher total number of target ID changes. After comparison, we found that we are not the highest in model speed, but the balance of model speed, accuracy, and other vital metrics shows that we should be the best.

We visualized our tracking results and analyzed the typical situation, as shown in Fig. 13. We analyzed MOT20-01, MOT20-02, and MOT20-03 in the MOT20 dataset, and the three images from top to bottom represent a segment in a video sequence. You can see that in the MOT20-01 sequence, our target of interest can be seen walking side by side with other targets at the beginning of the video. At this time, our target interest number is 141. In the middle segment, the target is obscured and loses its number at this time. After displacement, the target reappears and can correctly match number 141 again.

In the MOT20-02 sequence, the target we focus on at the beginning of the video clip is number 129, a partially obscured target that does not have complete pedestrian characteristics and has been in a semi-obscured state. In the middle part of the clip, the target is more or less displaced and more obscured, and at the end of the clip, although a long time has passed, the tracker can still accurately identify and number this target.

In the MOT20-03 sequence, the video clip has an obstacle blocking all the targets passing this obstacle; at the beginning of the clip, there is a pedestrian who is passing this obstacle; the number is 769 at this time, after being completely blocked, the excellent Kalman filter comes into play and accurately predicts the linear moving object, still retaining the tracking frame, at the end of the clip the pedestrian reappears. The detector can identify again that this is the previous target. The number is 769.

From the above visualization analysis, our method can perform accurate identification for target identity, extract deep semantic information about the target, and accurately predict and track objects that are obscured for a long time.