Research On Multi-object Joint Detection And Tracking Algorithm Based On Temporal Correlation

Multi-object tracking refers to the process of locating and maintaining identity IDs for multiple objects in a video under the same camera,which is of great significance for maintaining social and public order and building a safe city.In recent years,with the rapid development of deep learning,research on multi-object tracking based on joint detection and tracking has made many progress,but there is still a gap between the performance of relevant research and actual implementation.The reason lies in the randomness of the number of multiple object tracking,occlusion between objects,and the complexity of the tracking background,which pose difficulties and challenges to the neural network feature extraction and data association process.Through in-depth analysis of joint detection and tracking of multiple objects,this thesis conducts in-depth research from the perspectives of feature extraction and data association,and proposes two joint detection and tracking algorithms for multiple objects,effectively improving the issues of false detection,missed detection,and frequent ID switching,and improving the accuracy of multiple object tracking.The main work of this thesis is as follows:(1)A multi-object joint detection and tracking algorithm based on temporal context is proposed to solve the problem of missing features extraction in occlusion situations in multi object tracking joint detection and appearance representation networks,which leads to false detection and missed detection of objects.Firstly,by concatenating the compressed temporal feature information of historical frames,the current frame feature information is supplemented,and then the calibration factor is obtained through convolution operation to dynamically calibrate the convolution weight and deviation,and temporal feature fusion is performed to achieve the supplementation of occlusion features;Then,in order to avoid the problem of false detection caused by fused features when individual objects suddenly disappear in the current frame,a feature smoothing and enhancement module based on the current frame features is further constructed.This module mainly utilizes a coding decoding structure,uses an inverted bottleneck network to smooth features,and uses the visual features of the current frame to provide semantic information for detailed positioning,thereby enabling the network to learn better feature expression capabilities;Finally,a combination of focus loss,loss,and label smoothing cross entropy loss is used to monitor network training to help network convergence,optimize feature layers,and improve network tracking performance.(2)A joint detection and tracking algorithm for multiple objects based on threestage data association is proposed to solve the problem of insufficient matching and association capabilities of multiple object tracking algorithms for occluded objects,resulting in frequent ID switching and inability to further improve algorithm performance.Firstly,a trajectory smoothing strategy is adopted to improve the problem of Kalman filter prediction errors caused by long-term occlusion of trajectories.By establishing a virtual trajectory centered on the observation results,the parameters of the Kalman filter are modified,thereby improving the prediction accuracy of the Kalman filter;Then,in order to fully utilize the results of all confidence levels obtained from the detection,a secondary association detection frame mechanism is executed.By dividing the detection frame into a high score frame and a low score frame,IOU(Intersection Over Union)calculations are used twice to match the detection frame with the trajectory,making full use of the similarity between the low score detection frame and the trajectory,and promoting the correct matching of the occluded object and the trajectory.Combined with cascade matching in one stage,a three stage data association strategy is formed.Finally,a large number of experiments on data sets demonstrate the effectiveness of the proposed algorithm.The thesis has 28 diagrams,15 tables,and 84 references...