Research On Video-based Pedestrian Multi-object Tracking Technology

Pedestrian multi-object tracking is a research hotspot in the field of computer vision,and is widely used in intelligent surveillance,intelligent transportation,human-computer interaction and other scenarios.The pedestrian multi-object tracking algorithm mainly consists of four components: pedestrian detection,pedestrian re-identification,motion estimation,and data association.For the above four key technologies,the detection-based pedestrian tracking algorithm,the joint detection and embedding pedestrian tracking algorithm,and the instance segmentation-based pedestrian tracking algorithm are improved respectively.The work done in this thesis is as follows:1)To address the problem of low tracking accuracy due to detector lost and false detections in pedestrian multi-object tracking,a pedestrian multi-object tracking algorithm fuses SSO-YOLOv3 and Deep Sort is proposed.Firstly,SSO-YOLOv3 uses a lightweight Darknet53 network to reduce the number of convolution filters used in each residual block and improve the residual structure to balance the processing speed while enhancing feature extraction.The SE attention mechanism is introduced to learn the importance of semantic information among different channels.The atrous pooling pyramid module is added to increase the perceptual field while maintaining the spatial resolution.Using a stepped aggregation approach to aggregate multi-scale features to obtain high-resolution single-scale output,which improves label rewriting and anchor distribution problems.Using DIOU-Loss to accelerate model convergence and improve detection accuracy.The re-identification network of Deep Sort is replaced with Resnet50,thus improving the original network’s weak perception of tiny appearance.Finally,testing on MOT16 and MOT17 datasets,the tracking accuracy reached 51.3% and 50.6%.2)To address the problems of obvious changes in object appearance,irregular movements that easily lead to trajectory interruption and frequent identity switching in complex scenes,a pedestrian multi-object tracking algorithm based on the improved CStrack association strategy is proposed.Firstly,the association between features is enhanced using the appearance feature update module.Secondly,the secondary association tracking method is proposed.The first association uses IOU distance fusion appearance features as the cost matrix of association,and then second uses extended IOU association to alleviate the problems of motion estimation bias and the metric failure caused by indistinguishable appearance.The Gaussian regression algorithm is used to consider motion information and compensate for missed detection by interpolation to compensate for the missed detection.Finally,tested on MOT17 and MOT20 datasets,the tracking accuracy reached 73.9% and 64.2% respectively.The algorithm has obvious advantages in tracking accuracy and can better adapt to complex scenes.3)In order to make fuller use of the state information of object and further improve the tracking performance,the Mask-Sort based pedestrian multi-object segmentation tracking algorithm is proposed by correlating detection,segmentation and tracking.The motion similarity of the object is obtained by the gaussian mixture probability hypothesis density filter,and the appearance similarity is calculated by the correlation response.The mask-based affinity fusion module is then used to fuse the metrics to complete the association of detection results with trajectories;the motion camera compensation is used to mitigate the effect of the nonlinear motion of the object brought by the camera motion and improve the robustness of the filter prediction.The association between trajectories is used to link short trajectories to complete trajectories to reduce trajectory interruptions.Finally combined with the YOLOv5m-seg object segmentation algorithm,Mask-Sort is tested on the MOTS dataset.s MOTSA and MOTSA reached 61.2% and 77.9%respectively.The performance of segmentation tracking is effectively improved.