Multi-target Detection and Tracking Based on Improved YOLOv5 Algorithm and DeepSort Algorithm
Aiming at the difficulty of target detection and tracking in swimming pool due to water ripples,reflections,and similar appearances,a multi-target detection and tracking method based on improved YOLOv5 algorithm and DeepSort algorithm was proposed.By introducing attention mechanism,YOLOv5 algorithm was improved to enhance the ability of extracting target features.The detection results were input into DeepSort algorithm,and K neighborhood restriction was introduced into the cascade matching to filter the target detection box,which reduced the identity switching problem caused by the non-obvious appearance characteristics of the targets.Hungarian algorithm was used to match the detection box and the prediction box,and the distance intersection over union was used to replace intersection over union for the second matching of the unmatched detection box,which improved the tracking performance of DeepSort algorithm.The performance of the proposed multi-target detection and tracking method was verified by comparison and ablation experiments.The results show that the average accuracy of the improved YOLOv5 algorithm is 2％higher than that of the original algo-rithm.Combined with DeepSort tracking algorithm,the number of identity switching is reduced by 58 times on average,and the multi-target tracking accuracy is 80.26％,which is 3.85％higher than that of the original YOLOv5 algorithm and Deepsort algorithm.
万方数据
维普
