Target Tracking Algorithm In Railway Integrated Video Surveillance System Research On Its Application Technology

Target tracking is a hot topic in the field of computer vision.This algorithm utilizes contextual information from video or image sequences to model the appearance and motion information of the target,predict its motion state,and calibrate its position.Target tracking technology is widely used in military and civilian fields.Based on realtime railway surveillance videos,this article studies and proposes improvements to deep learning based target tracking algorithms,which have positive significance in improving the accuracy and model generalization ability of target tracking algorithms.The main research work of this article is as follows:(1)In this paper,a lightweight KN-YOLOC target detection algorithm based on attention mechanism and multi-scale feature fusion is proposed.This algorithm uses YOLOX network as the baseline network.In order to solve the problem of inaccurate personnel attitude recognition,vertical rotation data enhancement is introduced to enhance the algorithm’s personnel attitude recognition capability;Introducing an improved SENet attention mechanism module to address the small target problem of personnel and vehicles,enhancing the network’s feature extraction ability;By introducing Ghost convolution instead of general convolution,the model can be lightweight;Finally,in order to further reduce the parameter scale of the network and improve the accuracy of the model,the idea of Conv Ne Xt network is introduced to improve some modules in YOLOX and improve the loss function of the algorithm.Experiments have shown that the KN-YOLOC object detection algorithm outperforms the YOLOX algorithm in terms of average accuracy by 5.8%,inference speed by0.0114 seconds,and parameter size reduction by 0.95 Mb.(2)Deep Sort algorithm,a popular multi-target tracking algorithm in industry,has the problems of frequent target ID conversion and low tracking accuracy.This paper introduces the twin network method to improve Kalman filter,making it more suitable for nonlinear systems;By improving the matching process of the Hungarian algorithm,we aim to address the issue of ID loss in the algorithm;By optimizing the pedestrian recognition network in the algorithm,the algorithm is more suitable for the railway video environment and effectively solves the problems existing in the original algorithm.On the basis of the above research,a multi target tracking algorithm KNYOLOC+improved Deep Sort algorithm suitable for railway environments was proposed.The experiment showed that on the MOT16 dataset,the target tracking algorithm improved tracking accuracy and tracking accuracy by 16.8% and 1%,respectively,compared to the original algorithm.(3)In the field of multi-objective tracking,the Transformer framework has the advantages of flexible structure,high interpretability,and more suitable for multimodal tasks compared to neural network frameworks.Therefore,the Transformer framework based multi-objective tracking algorithm(MOTR)is improved.The improved MOTR algorithm uses a new network structure and a self attention module structure.Experiments have shown that the tracking performance has been improved to some extent on the MOT17 dataset,with tracking accuracy and target identity conversion times improved by 1.1% and 1.2%,respectively,meeting the requirements of the tracking task.(4)Based on the actual requirements of the railway video monitoring system project,an easy-to-use human-machine interaction interface was built using the Py Qt framework,and a railway comprehensive video monitoring system was designed...