Research And Application Of Pedestrian Tracking Algorithm Based On Deep Learning

With the rapid development of computer vision technology,multi-target tracking algorithms are becoming more and more mature.As a research hotspot,pedestrian tracking has broad application prospects,so it has attracted the attention of industry and academia.In terms of practical applications,pedestrian tracking algorithms which analyse monitoring scenes in video surveillance systems should take into account both real-time and tracking accuracy.At present,the pedestrian tracking algorithm based on deep learning has shown good performance compared with the traditional algorithm,but the convolutional neural network has a complex structure and a large amount of calculation,and when faced with problems such as blurred pedestrian features and mutual occlusion,it is impossible to guarantee accurate association of pedestrian trajectories.This paper explores the pedestrian tracking paradigm and optimizes the overall design of the model for the purpose of reducing the weight of the algorithm model,enhancing the integrity of the trajectory,and improving the accuracy of trajectory association under occlusion.The main work is as follows:1)Optimize the lightweight feature extraction network,choose Mobile Vi T as the basic network to capture the global and local features of pedestrians,and improve it for the purpose of reducing the amount of parameters and calculations.After analyzing the advantages and disadvantages of the Mobile Vi T network in the feature extraction process,the feature sampling mechanism is used to screen representative global features,the global encoding dimension is changed to reduce the amount of calculation,and the parameter amount is reduced by reducing the intermediate transformation dimension of the linear projection structure.The experimental results show that the original network runs at20 FPS on the MOT16 public data set,and the model parameters are 18.4MB.After optimization,the running speed is increased by 15%,and the model parameters are reduced by 13%.2)To solve the problem of frequent identity conversion caused by the disappearance of most of the features of pedestrians when they are occluded,a pedestrian posture keypoint detection branch is added to output multiple discriminant features on the basis of the joint appearance feature detection and tracking mode.In the process of association matching,the similarity between the occluded pedestrian and the trajectory is measured by the keypoint of the visible pedestrian pose,and different weights are assigned to the similarity distance obtained by various features,aiming to reduce the number of identity switching under the constraints of various angles.Experiments on the MOT16 public data set show that the number of pedestrian identity switching in the joint appearance feature detection and tracking mode is 1202,and the number of identity switching is reduced by16.5% after using multiple features for joint detection,and the accuracy of multi-target tracking can reach 47.6%.3)In order to improve the integrity of pedestrian trajectories,the multi-frame tracking mode is adopted on the basis of Center Track’s modeling of pedestrian movement process to reduce the excessive dependence on instantaneous detection performance.On the one hand,a convolutional gated recurrent unit is used to learn the spatio-temporal characteristics of each pedestrian trajectory,enhancing the ability to locate the pedestrian’s position when the detector fails;on the other hand,a coordinate attention enhancement mechanism is used to improve the recognition of the pedestrian’s apparent characteristics,enhancing the ability to recover the identity when the trajectory disappears and reappears.During the training of the multi-frame tracking algorithm,a pedestrian trajectory data enhancement strategy is applied to the input video sequence to improve the robustness of the algorithm.Experimental results show that target tracking accuracy can reach 59% on the MOT16 public dataset,an improvement of 8.1% compared to the baseline,an increase of 91 in the number of high-probability tracking trajectories and a reduction of 90 in the number of identity switches.4)Deploy the optimized pedestrian tracking algorithm on the Jetson Nano hardware platform,and use the TensorRT framework for inference acceleration to test the performance of the algorithm in application scenarios.The results show that the method in this paper runs at 5.5FPS and accurately recognizes pedestrians.