Design And Implementation Of Video Pedestrian Target Retrieval Platform Based On Improved YOLOv5

Video data is widely used in various fields such as public safety,traffic management,and commercial regulation.In these areas,it is particularly important to retrieve specific pedestrians in the video.However,due to the large and complex volume of video data,how to efficiently and accurately retrieve pedestrian targets from it has become a hot research topic.Traditional manual search methods are inefficient and error-prone.This paper proposes a video pedestrian search technology based on improved YOLOv5s and implements a video pedestrian search platform,considering the characteristics and requirements of pedestrian search.Firstly,to reduce redundant information and improve efficiency,this paper uses a clustering-based keyframe extraction algorithm.It converts the video to HSV color space and calculates the color histogram for each frame.The algorithm clusters all frames to retain effective information in the video.Experimental results show that the clustering-based keyframe extraction algorithm achieves higher fidelity and compression ratio for videos focusing on pedestrians.Secondly,this paper performs pedestrian detection on the key frames of the video,and then intercepts the pedestrian images from the key frames according to the predicted frames of the pedestrians according to the results of pedestrian detection.To solve the problem of difficulty in balancing the complexity of the detection network and the accuracy of the detection in the pedestrian detection task,this paper proposes a lightweight pedestrian detection network based on the YOLOv5s network,which is called YOLOv5s-G~2 network.The YOLOv5s-G~2 network uses lightweight Ghost C3 and Ghost modules,which can maximize the reduction of network complexity while maintaining network detection accuracy.At the same time,the global attention mechanism is applied to the YOLOv5s-G~2 network,which effectively improves the ability of YOLOv5s-G~2 to extract pedestrian feature information.In addition,theα-CIo U loss function is proposed to replace the GIo U loss function,which effectively enhances the positioning ability of the predicted boxes generated by the network.The YOLOv5s-G~2 network is evaluated on the Wider Person dataset to ensure its effectiveness.Compared with the existing YOLOv5s network,our proposed YOLOv5s-G~2 network not only improves the detection accuracy but also reduces the complexity of the algorithm.Compared with common lightweight pedestrian detection networks,the YOLOv5s-G~2 network can better balance the complexity of the detection network and the accuracy of the detection.Then this paper uses a lightweight Bo T network to calculate the similarity between the pedestrian image to be searched and the extracted pedestrian image in the video.It identifies pedestrians in the video with similarity above the threshold as the same person as the retrieved pedestrian.In the design of the pedestrian re-identification network,the pedestrian re-identification network replaces the original Res Net-50 backbone with the lightweight OSNet.Experiments on the Market-1501 dataset demonstrate that using OSNet as the Bo T network backbone achieves better performance with fewer parameters compared to other networks.Finally,based on the proposed algorithm,this paper designs and implements a video pedestrian search platform using Qt as a GUI library.Then,functional experiments are conducted on the video pedestrian search platform to verify that the platform can stably implement the pedestrian search function.