Research On Pedestrian Multi-Object Tracking Algorithm Based On Deep Learning

In recent years,with the rise of deep learning theory,various fields of computer vision have been fully developed.As an important part of computer vision,multi-object tracking technology has made great progress in this wave of deep learning.Since multi-object tracking plays an important role in many fields such as military,criminal investigation,and business analysis,it has important value in its research.Compared with traditional methods,the current mainstream multi-object tracking algorithms based on deep learning have made significant progress,but people put forward higher requirements for multi-object tracking algorithms,such as real-time efficient tracking,crowded scene tracking,etc.Existing tracking algorithms still have deficiencies in the above two aspects due to redundant network structures and insufficient target representation.This paper studies two problems of real-time efficient tracking and crowded scene tracking for multi-object tracking.The main work is as follows:In order to maintain the real-time performance of the multi-object tracking algorithm and improve the tracking performance,a multi-object tracking algorithm with joint detection and tracking is proposed.Compared with the traditional two-stage multi-object tracking algorithm,this algorithm integrates the detection network and the tracking network,and realizes the end-to-end training of the detection task and the tracking task.Aiming at the inconsistency of the features required by the detection task and the tracking task,a "spacechannel" attention decoupling module is proposed,which can decouple the shared features of the model in the two dimensions of space and channel,so that the detection task and tracking can be separated.Tasks have exclusive feature representations.The experimental results show that the multi-object tracking algorithm of joint detection and tracking can effectively improve the tracking performance while ensuring real-time performance.In order to solve the problem of poor multi-object tracking performance in complex scenes,a multi-object tracking algorithm based on Transformer is designed.Use Transformer to enhance the model’s representation of the target.In order to speed up the convergence speed of model training,the algorithm adopts a sparse attention module,which reduces the problem of identity switching in the case of occlusion while ensuring the speed of training convergence.The channel of the input vector is divided into two,and the self-attention calculation and the convolution calculation are performed respectively,which not only increases the position vector information,but also introduces the local feature information.Experiments show that this algorithm can reduce the number of parameters of the model without reducing the performance of the algorithm,and effectively improve the inference speed of the model.To sum up,this paper proposes corresponding solutions for the two existing difficult problems in multi-object tracking,which can effectively improve the performance of the model in corresponding scenarios,and lay a foundation for future research work.