Research On Multi-Object Tracking Algorithm Based On Deep Learning

With the rapid development of science and technology,computer vision has more and more extensive applications in daily life.The multi-object tracking studied in this paper is a typical computer vision technology.It refers to analyzing and processing multiple frames of images,judging the position changes of multiple objects in the image in a continuous time,so as to draw the tracking trajectory of each object.This technology is widely used in autonomous driving,public safety and other fields.Due to the complex environment background,frequent occlusion and motion changes of the tracked object in practical applications,this research is quite challenging.This paper focuses on the research and analysis of the multi-object tracking problem based on deep learning.The specific research contents are as follows:Firstly,in order to solve the problem that the detection and re-identification tasks are carried out independently,which brings a lot of repeated computation and slows down the inference speed,this paper designs a multi-object tracking method with joint detection and re-identification.The anchor-free detection and re-identification branches share the deep layeraggregation backbone network,and the depthwise over-parameterized convolution is used to replace the deformable convolution in the upsampling process.These designs effectively improve the feature expression ability and parameter utilization.On this basis,the high and low-confidence detection boxes are divided,and the high-confidence detections and trajectories are preferentially matched,and then the low-confidence detections are matched with the remaining trajectories.These designs improve the real-time and accuracy of tracking.Secondly,for the problem that the detection does not use the tracking clues,this paper designs a multi-object tracking method that combines detection and tracking.The backbone network combines Convolutional Neural Network with self-attention mechanism to improve feature extraction capability,and uses involution instead of ordinary convolution for upsampling.The re-identification feature embeddings are extracted point by point in two frames,a cost matrix is constructed to store the embedding similarity,and the tracking offset is derived from it,and the tracking offset and embedding are used for data association.These designs propagate object features from the previous frame to the current frame,aggregate these features for detection,and effective ly utilize the tracking information.Finally,for the problem that the multi-object tracking based on graph matching does not consider the relationship between frames,this paper models the relationship between trajectories and intra-frame detection as an undirected graph,and transforms the association problem into a graph matching.By using linear interpolation on missing frames to fill in the gaps of the trajectory,and using different losses to optimize bounding box regression,these designs enable multi-object tracking to maintain high performance in complex scenes.