Research And Application Of Multi-Object Joint Detection And Tracking Network Algorithm

Multi-target tracking algorithms are used to detect and locate targets of interest in a given video frame sequence and assign a globally unique ID value to each target object.Although the common joint detection and tracking algorithms share the same backbone feature extraction network,the detection and tracking tasks are still relatively independent and have different dependencies on feature depth.Therefore,this thesis improves the traditional joint detection and tracking algorithm with the following main research contents and contributions:To address the problem that detection task and tracking task depend on different feature depths,this thesis proposes an improved algorithm(R&D)based on image fusion module and feature screening enhancement module from a single frame image,which achieves the purpose of assisting tracking by enhancing detection effect.The image fusion module achieves multi-level and multi-scale fusion of features,which effectively avoids the extracted image features to be overly inclined to the detection task or tracking task.The feature screening enhancement module reprocesses the image features,considers the correlation of different positions on the feature map,and screens out the features from the output features of the image fusion module that are suitable for the detection task and the tracking task,respectively.Comparative experiments were conducted on the MOT17 dataset.Compared with the baseline CenterTrack algorithm,the R&D algorithm improved MOTA by 2.5%,IDF1 by 3.5%,MT by 5.1%,ML by 5.4%,and FPS by 5.6.However,in terms of the number of ID switches,R&D increased the number of ID switches compared to the CenterTrack algorithm by 943(36%).To address the problem that the detection task and tracking task of common joint detection tracking algorithms are still relatively independent,this thesis proposes the R&D-Track algorithm based on the R&D algorithm,which updates the input features of the detection branch of the current frame by calculating the tracking offset of consecutive frames,and based on this,the R&D-Track algorithm achieves the purpose of assisting detection by tracking cues.algorithm proposes a cost matrix construction module to calculate the matching similarity information between feature maps;a tracking offset calculation module to calculate the tracking offset between two frames;a feature update module to update the input features of the current frame detection branch by the tracking offset;and a two-stage data association algorithm based on the track pool to improve the efficiency of data association.Comparative experiments were conducted on the MOT17 dataset.Compared with the baseline CenterTrack algorithm,the R&D-Track algorithm improves MOTA by 3.7%,IDF1 by 3.8%,MT by 5.2%,ML by 5.5%,IDs by 358(13.9%),and FPS by 0.1.For the dense crowd gathering scene and escalator scene,this thesis implements a trampling and falling early warning system based on R&D-Track algorithm to warn of dangerous events such as falling,retrograde and crowd convection.The system consists of two parts: front-end and back-end.The back-end mainly analyzes the input video in real time to determine whether a dangerous event has occurred and make corresponding warning measures;the front-end visualizes the monitoring video and analysis results.