Research On Pedestrian Detection And Re-identification Methods Based On Deep Learnin

With the continuous development of urbanization,the increasing public demand for maintaining social order and stability,the application of video surveillance technology plays a crucial role in maintaining social stability.In recent years,the number of deployed cameras in public places has rapidly increased,and a massive amount of data is stored in video surveillance systems.Pedestrians are the key targets of attention in video surveillance data,and pedestrian detection and re-identification techniques related to them have received widespread attention and research.Based on deep learning technology,this paper conducts research on pedestrian detection and re-identification methods,and the main research work is as follows:(1)Pedestrian detection algorithm based on hybrid attention and structural reparameterization(YOLOv5s-MR).To address the issues of false alarms and missed detections caused by pedestrian occlusion in surveillance scenarios,a hybrid attention module is designed.This module can guide the pedestrian detection algorithm to focus on the visible parts of occluded pedestrians in a weakly supervised manner,enhancing the feature extraction capability of the pedestrian detection algorithm for pedestrian targets.Inspired by the structural reparameterization concept in the Rep VGG algorithm,a Rep-CSP module is designed.During training,a multi-branch structure is employed to enhance the model’s feature extraction capability.During inference,the branches are fused without increasing additional computational cost,thereby improving the detection accuracy of the model.Additionally,a tiny object detection head is added to improve the detection capability for small pedestrian targets.Experimental results demonstrate that this method achieves high detection accuracy and effectively alleviates false alarms and missed detections caused by pedestrian occlusion in surveillance scenarios.(2)Pedestrian re-identification algorithm based on Attention Guided Weighted-MultiFeature Aggregation(AGW-MA).To address the issue of insufficient discriminative power in current pedestrian re-identification algorithms,a multi-feature cascaded structure is designed.In addition,the cosine cross-entropy loss is introduced to leverage the characteristics of the multi-feature cascaded structure,which supplements detailed information from deep feature maps and increases inter-class sample discrimination,thereby enhancing the feature discriminative power of the pedestrian re-identification algorithm.To mitigate the additional noise caused by bounding box offsets and occlusions in detection algorithms,an Attention Pyramid Network(F-APNet)is introduced.By utilizing the attention pyramid network’s ability to better focus on effective pedestrian features in cluttered backgrounds,the interference from additional noise is reduced.Experimental results demonstrate that this method achieves good recognition performance on multiple pedestrian re-identification datasets,effectively enhancing the feature discriminative power of pedestrian re-identification algorithms.(3)Design and Implementation of Pedestrian Re-identification System.By combining the pedestrian detection algorithm based on hybrid attention and structural reparameterization with the pedestrian re-identification algorithm based on attention pyramid and multi-feature cascaded structure,a pedestrian re-identification system was designed.The reliability and effectiveness of the proposed algorithm in practical application scenarios were demonstrated through scene testing.