Research On Person Detection Technology Based On Yolov3

Person detection is a branch of object detection.It is an association of character recognition and character positioning.In layman's terms,it refers to locate a person in an image and finding the position and range of the person,and marking the person with a suitable frame.Today,with the intelligent development of various industries,person detection has a wide range of research and applications in many fields,such as intelli-gent transportation,video surveillance,target tracking and other fields are inseparable from person detection as a necessary pre-processing,so person detection Algorithms are crucial today to advance intelligent construction.This article first analyzes and summarizes the current development and research results of object detection and person detection,focusing on the analysis of some results in the field of deep learning,such as R-CNN,Fast-R-CNN,Faster-R-CNN,and Yolo A series of algorithms.On this basis,we are also committed to solving the current research difficulties in person detection problems,such as low accuracy caused by incomplete feature extraction in the network,optimization of candidate frames,etc.Based on the previous research,we mainly do Work have:(1)In the process of person detection,one of the key links is the selection of the pre-selection box.The choice of the pre-selection box will directly affect the quality of the detection of the person.Combined with our specific problems,considering the uncertainty of the size characteristics of the characters,in the selection stage of the candidate box,we use K-means clustering algorithm to cluster the data,and through visual analysis and quantitative analysis of IOU values,comprehensive Considering the clustering complexity and clustering results,we select candidate frames that can better fit the character data as our preselection box size.The selection of candidate frames is optimized so that the average IOU value is from the original 72.26% increase to 90.97%,which enhances the overall detection performance of the network.(CHAPTER ?)(2)In this article,after analyzing the existing detection networks,we believe that the end-to-end network structure,that is,the Yolo series of networks,can have superior performance in speed and detection properties.Characteristics of the end-to-end net-work,improved the current most popular person detection algorithm Yolov3 to build our network framework,the first stage continues to use Darknet-53,which has been proven to have outstanding performance in classification features,as a feature extrac-tion network.The multi-scale image features are synthesized later to obtain a feature map that can better express the picture features for person detection.A new person de-tection network model PDnet was proposed.The accuracy rate was increased from 86%to 92% in the quantitative evaluation standard,and the visual performance also showed better results.(CHAPTER ?)(3)On the basis of our previous research,we consider that the unavoidable pol-lution in the process of image acquisition and transmission causes the image quality to get damaged.We explore ways to solve the person detection in this part of data.First,we use denoising + detection method to solve the problem,comprehensive analysis and comparison of the existing denoising algorithms to select a denoising network with ex-cellent performance,and evaluate the quality of the picture in the early stage of our detection.Denoising processing and detection are performed.The method's detection and evaluation index for contaminated pictures are 1.42 percentage points higher than before,but the detection process is complicated.Then,in order to simplify the detection process,we use the method of adding training pictures to extend the training to enhance the PDnet we proposed earlier to solve this problem,and found that the mAP value of the detection result increased from 66.73% to 71.65%,And the detection process is simple.As our first fusion algorithm,further,we combine the feature attributes of the contaminated pictures and the good structure in the denoising network,and fuse the de-noising network with our detection network to propose a ResPDnet network structure.The new network is trained on data from blending,and the obtained model framework further improves the detection effect of such pictures to 73.11%.So far,the applicable range of our human detection algorithm has been expanded,and a feasible solution has been explored for similar problems.(CHAPTER ?).