洗煤厂工人的安全问题一直以来都被人们所重视，为此，洗煤厂通过引入智能化视频监控，以便及时发现监控视频中的异常状况，减少安全性问题的发生。但洗煤厂中监控设备安装位置固定，大型设备较多，导致工人在工作中易被局部遮挡，增加了行人重识别与跟踪的难度。本文针对局部遮挡导致的行人跟踪精度差、跟踪速度低、行人无法识别和错误识别的问题，研究了一种基于YOLOv8s+DeepSORT+FastReID的多目标跨摄像机识别与跟踪的方法，并对DeepSORT算法及FastReID算法进行优化。具体的工作内容如下所示：

针对行人跟踪精度差、跟踪速度低的问题，对目标跟踪算法做出以下优化。首先使用SIOU(Smoothed Intersection over Union)算法代替DeepSORT算法中的IOU(Intersection over Union)算法，提高了算法的收敛度；其次将ShuffleNetV2网络融入到DeepSORT算法中，并在ShuffleNetV2网络中添加ECA-Net注意力机制，提高了目标跟踪的准确率和跟踪速度。实验结果显示，在行人跟踪过程中，行人跟踪准确率提高了6.10%，跟踪速度提高了27.41%。

针对行人重识别技术在实际运用中自动化率低、行人无法识别或错误识别的问题，对FastReID行人重识别算法进行以下优化。首先采用自动保存行人特征图的方法代替人工采集，提高了行人重识别的自动化率。之后采用校正行人身份信息、优化行人特征图像库的方法，减少了行人无法识别或错误识别的次数，提高了行人重识别的准确率。实验结果表明，经过优化后的算法提高了行人重识别技术的自动化率，行人无法识别或错误识别的次数减少了80.54%，行人重识别准确率提高了7.88%。

经过实验验证，改进后YOLOv8s+DeepSORT+FastReID算法提高了行人重识别与跟踪技术在洗煤厂的跟踪精度、跟踪速度以及行人重识别的准确率，能够实现洗煤厂场景下的行人跨摄像机识别与跟踪，在计算力高的设备中能够达到实时识别和跟踪的需求。

The safety of coal washing plant workers has always been paid attention to, therefore, the coal washing plant through the introduction of intelligent video surveillance, in order to find the abnormal situation in the surveillance video in time, reduce the occurrence of security problems. However, the installation position of monitoring equipment in coal washing plants is fixed, and there are many large equipment, which leads to the partial occlusion of workers in the work, and increases the difficulty of person identification and tracking. Aiming at the problems of poor tracking accuracy, low tracking speed, unrecognition and misidentification of persons caused by local occlusion, this paper studies a multi-target cross-camera recognition and tracking method based on YOLOv8s+DeepSORT+FastReID, and optimizes DeepSORT algorithm and FastReID algorithm. The specific work content is as follows:

Aiming at the problems of poor tracking accuracy and low tracking speed, the following optimization is made to the target tracking algorithm. Firstly, SIOU(Smoothed Intersection over Union) algorithm is used to replace IOU(Intersection over Union) algorithm in DeepSORT algorithm, which improves the convergence degree of the algorithm. Secondly, the ShuffleNetV2 network is integrated into the DeepSORT algorithm, and the ECA-Net attention mechanism is added to the ShuffleNetV2 network to improve the accuracy and tracking speed of the target. The experimental results show that the accuracy rate of person tracking is increased by 6.10%, and the tracking speed is increased by 27.41%.

In view of the low automation rate of person re-identification technology in practical application, persons can not be recognized or wrong recognition problems, the FastReID person re-identification algorithm is optimized as follows. Firstly, the automatic saving of person feature map is used instead of manual collection, which improves the automatic rate of

person re-identification. Then, the method of correcting person identity information and optimizing person feature image database is adopted to reduce the number of unrecognizable or wrong identification of persons, and improve the accuracy rate of person re-identification. The experimental results show that the optimized algorithm improves the automation rate of person re-identification technology, the number of unrecognized persons or wrong recognition is reduced by 80.54%, and the accuracy rate of person re-identification is increased by 7.88%.

After experimental verification, the improved YOLOv8s+DeepSORT+FastReID algorithm has improved the tracking accuracy, tracking speed and person re-identification accuracy of the person re-identification and tracking technology in the coal washing plant, and can realize person cross-camera recognition and tracking in the scene of coal washing plant. Real-time identification and tracking can be achieved in devices with high computing power.