Research On Multi-Target Tracking Algorithm Based On Joint Detection And Embedding

Multi-target tracking is an important field of computer vision,which is widely used in many fields such as intelligent security,automatic driving,urban transportation,national defense and military,and intelligent robots.In recent years,with the vigorous development of deep learning,detection-based multi-target tracking algorithms have also developed rapidly.The current mainstream multi-target tracking algorithm uses a unified model for detection and embedding to output at the same time,which is called one-shot multi-target tracking(one-shot MOT),which often uses the same convolutional neural network to complete the target detection features.Extraction and target identity embedding feature extraction are two tasks,and tracking is completed on this basis,taking into account the tracking effect and tracking running speed.However,there are still many problems in the one-stage multi-target tracking algorithm,such as target occlusion,poor effect when the target size is small,target ID switching caused by insufficient extraction of ID information and position information,and so on.Aiming at the above problems,this paper uses the multi-target tracking algorithm JDE(Joint Detection and Embedding)as the basic framework to research and improve the multi-target tracking algorithm.The specific content is as follows:(1)In order to improve the detection accuracy of the unified network model in the JDE algorithm and improve the tracking effect,a spatial domain attention mechanism and a channel domain attention mechanism are added in front of the prediction head of the YOLOv3 unified model to improve the tracking of the model when targets overlap Effect,combined with the Spatial Pyramid Pooling(SPP)module to expand the receptive field,obtain more comprehensive semantic information,and improve the model’s ability to detect and track small-sized targets.(2)In order to alleviate the excessive competition of the JDE algorithm due to the difference between the two tasks of detection and embedding features,which will damage the learning of related task features and further hinder the tracking performance,this paper reduces the dimensionality of the output embedding features.Balance the two tasks.In addition,the reduction of dimensionality also reduces the parameter amount of the model,improves the inference speed and reduces the calculation time.(3)In order to solve the problems of missed detection and false detection caused by the performance limitation of the detector in the JDE algorithm and the poor tracking effect caused by the fact that some completely occluded or severely occluded targets cannot be captured by the detector alone,this paper will reduce the weight of the The Gaussian smoothing interpolation algorithm(GSI)that does not rely on additional models is embedded in the JDE algorithm,which can fully improve the various indicators of the JDE algorithm without additional time-consuming components.This paper follows the training data set originally used by JDE,conducts training and experiments based on the tracking target of pedestrians,and designs ablation experiments to verify the effectiveness of each improvement.The experimental results show that the improvement schemes proposed in this paper have improved the performance of the JDE algorithm to a certain extent,and the final model is better than the original JDE algorithm in multiple evaluation indicators.In addition,by comparing with some advanced multi-target tracking algorithms,the algorithm in this paper also has certain comparability and advantages.