Research On Pedestrian Retrieval Method In Mining Area Based On Multi-scale Prediction And Weight Constraint

With the frequent occurrence of safety accidents in mining areas,the safety of workers in mining areas has become the top priority of mining areas.Most of today's mining area has a relatively perfect video monitoring system,the use of multiple cameras no dead angle to cover the working area,however,the capacity of surveillance video is huge,and the manual screening of video content is not only time-consuming and laborious,but also inefficient,and it is impossible to timely and accurately feedback the location and life safety information of the staff in the mining area.Therefore,the realization of accurate retrieval and identification of workers in the mining area has become an urgent need for safe production in the mining area.On this basis,this thesis studies a pedestrian retrieval method in mining areas based on multi-scale prediction and weight constraints,the fusion improvement from the two aspects of pedestrian detection and pedestrian re-identification effectively solves the problem of pedestrian retrieval in complex mining scenes,and improves the retrieval and identification accuracy of mining workers in actual scenes.First of all,this thesis introduces the research background and significance,domestic and foreign research status,and analyze the problems of pedestrian retrieval in mining areas.Explain the relevant theoretical knowledge involved in the method proposed in this article,summarize the basic structure of convolutional neural networks,classic network models and common network optimization methods.Secondly,pedestrian detection algorithm based on multi-scale prediction YOLOV4 is proposed.First of all,aiming at the problem of small targets and scale changes in mining scenes,the multi-scale prediction is improved to make full use of the shallow layer features between different scales,so as to improve the representation learning of small target by the network model.Secondly,in view of the inherent attribute that the pedestrian height is greater than the width,K-means++ algorithm is used to cluster to generate the anchor box conforming to the sample characteristics.Then,a rectangular transformation and inference method is proposed to solve the large amount of redundant filling information in the post-processing stage of M-YOLOV4 pedestrian detection algorithm and improve the speed of pedestrian detection inference.Finally,the mixed-precision multi-scale training method is used to improve the model training speed.The experimental results in the VOC?COCO pedestrian detection dataset constructed in this thesis show that in complex mining scenarios,the MYOLOV4 pedestrian detection algorithm proposed not only has good detection accuracy and detection performance,but also can achieve real-time video stream processing.Then,channel attention pedestrian re-identification algorithm based on weight constraint is proposed.Firstly,aiming at the problem of pedestrians wearing different clothes in mining areas,channel attention feature extraction network based on appearance invariance was proposed to improve the self-adaptability of the model to the change of pedestrian appearance and realize the accurate identification of pedestrian identity in mining areas.Then,aiming at the difficulty of uniform work clothes for pedestrians in the mining area,a difficult sample sampling loss function based on weight constraints is proposed to ensure that the model obtains better discrimination and high performance.Finally,according to the characteristics of the mining scene,a Miner-Market pedestrian re-identification data set that is more suitable for engineering application value is constructed.Experimental results show that the proposed algorithm not only has higher matching recognition accuracy,but also greatly improves the generalization and robustness of the model in the real scene.In addition,five different scenes are selected from the inside and outside of the mining area to visualize the pedestrian retrieval,which proves the superiority of the pedestrian retrieval method in the mining area based on multi-scale prediction and weight constraint from multiple perspectives.Finally,summarizes the research contents and prospects the possible research directions in the future.The thesis has 32 figures,11 tables,and 95 references.