Research And Application Of Multi-target Tracking Technology Based On Improved Deep SORT

With the development of society and the advancement of urban intelligence,a large number of video surveillance devices have emerged in our lives.Relying on human resources to process huge monitoring information is extremely inefficient.and high misjudgment and high omission are inevitable.Multi-target tracking technology fills the defect of manual processing.and can realize video monitoring and analysis in a non-contact.labor-saving.material-saving,high-precision and high-speed way.to protect public safety.Multi-target tracking refers to the detection of multiple targets(such as pedestrians.cars.etc.)in the video,and the tracking and ID assignment of each target.The existing multi-target tracking algorithms are mainly divided into the tracking algorithm based on initial frame and the tracking algorithm based on detection.Among them,the multi-target tracking algorithm based on detection has become the mainstream of multi-target tracking technology due to its flexibility and the increasingly powerful performance of detection network.Although this kind of algorithm has achieved certain research results.there are still four problems:1)low detector accuracy and time-consuming;2)Inaccurate tracking of newly emerged targets.with misjudgment and missed judgment;3)The tracking of irregular moving targets is inaccurate and the adaptive ability is poor;4)The poor feature extraction ability of the target makes the ID switching problem in the tracking process.In response to the above issues,this thesis proposes a multi-target tracking algorithm based on the improved Deep SORT,named Deep SORT++,which aims to improve the tracking effect of targets in practical applications.The Deep SORT++algorithm proposed in this thesis mainly improves Deep SORT from four aspects:(1)A detector module based on data expansion is proposed.By combining the data expansion technology,four methods of image blurring,increasing noise,color interference and random channel pixel deactivation are designed to effectively expand the existing data set,maintain the integrity of the original data,improve the scale and diversity of the sample data set.enhance the robustness and generalization ability of the detection algorithm,and improve the detection accuracy.At the same time,the two-stage Fast R-CNN detector in Deep SORT was replaced by the first-stage YOLOv3 detector,which improved the detection speed.(2)The Kalman filter prediction module based on orientation supplement is proposed.The orientation network is introduced.When the target first appears,the initial state of the Kalman filter prediction is adjusted by using the speed information of all targets in the current picture(that is,the speed information at the same time),combined with the orientation information predicted by the orientation network,so as to improve the prediction accuracy of the Kalman filter for the new target and further alleviate the phenomenon of omission.(3)The update module of Kalman filter based on inflection point adaptation is proposed.In the case of turning,the weight of the detection frame and the prediction frame during the update of the Kalman filter is adjusted based on the included angle of the vector,so as to improve the fast adaptive ability of the Kalman filter update to the irregular moving target and obtain more accurate position information.(4)The appearance feature extraction network module based on attention mechanism and comparative learning is proposed.The multi-head convolution attention mechanism and comparative learning are introduced to optimize the original appearance feature extraction network,improve the feature differentiation of the network for different objects,and extract more effective appearance features.In this thesis,the Deep SORT++algorithm and other multi-target tracking algorithms are measured on the MOT 16 dataset.The experimental results show that the improved algorithm proposed in this thesis effectively improves the accuracy of multi-target tracking,and proves the superiority and effectiveness of the model proposed in this thesis.