Pedestrian Detection Method Based On YOLOV3 Network

In recent years,the field of computer vision has developed rapidly,and video surveillance applications have been widely promoted.Pedestrian detection technology is becoming more and more important.The pedestrian detection task uses feature extraction of pedestrians in videos or images to determine whether there are pedestrians in the target and achieve precise positioning.From the current main popular person detection algorithm,we can see that there are many problems,such as poor robustness,low pedestrian positioning accuracy,poor detection accuracy,and relatively high missed detection rate.This thesis takes the pedestrian detection task application as the research object,studies the pedestrian detection model based on the YOLOV3 network,and makes certain improvements in the image division method and the initial candidate frame size.The main work content and innovations are:The first is to study the unique characteristics of the pedestrian target in the detection task.The second is to study the problems faced in the process of pedestrian detection,and improve from the following aspects: First,from the perspective of image division,in view of the detection speed and detection Accuracy needs to be considered in place.This thesis has launched many comparative experiments to determine that the network image is divided according to the size of 10×10.Second,from the perspective of the initialization of the candidate frame,this thesis has made innovations in the clustering algorithm,and based on this,the height and width of the candidate frame are predicted.Third,from the detection point of view,on the basis of the previously adopted network model,the shallow feature extraction method is introduced to enhance the network's ability to recognize features of different sizes.This thesis uses PASCAL VOC2007+2012and INRIA two data sets to conduct a comparative analysis of the original and improved to verify the effectiveness of the improvement in this article.Finally,I used my own life photos for detection and testing.The experimental results show that this algorithm can achieve accurate pedestrian positioning without missing detection.