Research On Pedestrian Detection Technology Based On Multi-scale Feature Fusion

With the rapid development of video surveillance,intelligent robots,intelligent security and other applications,pedestrian detection technology has gradually entered the public field of vision.For a single pedestrian detection task,domestic and foreign scholars have carried out relevant technical research,but the general target detection algorithm is not high accuracy in pedestrian detection under complex background.In order to improve the accuracy and speed of pedestrian detection,a pedestrian detection model based on YOLOv3 algorithm was designed,and a pedestrian detection method based on multi-scale feature fusion was proposed and improved.To solve the problem of missed detection and false detection in pedestrian detection,the improved algorithm includes: on the one hand,the detection layer is added to fuse global and local multi-scale feature information and improve the feature extraction ability of pedestrian target.In the stage of network post-processing,GIo U-loss is introduced as the bounding box loss function.On the other hand,the feature extraction network structure is improved and the linear scaling method is used to optimize the candidate box.The specific work of this thesis is as follows:(1)Expatiate on the research background and significance of pedestrian detection,and study pedestrian detection technology based on traditional methods and deep learning methods;Focus on pedestrian detection technology advantages and disadvantages,the latest analysis of pedestrian detection technology research,key research based on the algorithm of pedestrian detection YOLOv3 model,and from the target positioning,multiple feature fusion,the maximum inhibition,K-means algorithm four process analysis in detail,finally in the INRIA pedestrian detection data set on the preliminary experiment results.(2)Aiming at the problem of missing detection in pedestrian detection,this thesis proposes an improved multi-feature fusion pedestrian detection algorithm.Firstly,the algorithm achieves sufficient shallow network and deep feature extraction of pedestrian targets by adding multi-scale feature fusion prediction.Then,feature fusion of the model is carried out through route Layer,and the improved network model of YOLOv3＿New Layer is designed.Secondly,in order to improve the detection accuracy of the model,GIo U-loss is designed as the bounding box loss function,and the improved network is YOLOv3＿GIo U.Finally,through the experimental comparison,the optimized network model has better detection performance.(3)In complex pedestrian detection scenarios,in order to further improve the accuracy of pedestrian detection,linear scaling K-means pedestrian detection algorithm is proposed in this thesis.On the one hand,the algorithm improves the network structure of feature extraction,integrates BN layer into convolution layer,and improves the forward inference speed of the model.The optimized model is YOLOv3＿BN.On the other hand,linear scaling method was proposed to optimize the candidate box after k-means clustering.The improved model was designed as YOLOv3＿Linear Scale,and training was carried out on INRIA standard pedestrian data set by optimizing network training parameters combined with multifeature fusion idea.Experimental results show that the pedestrian detection algorithm optimized in this thesis has better generalization ability and improves detection accuracy and speed.