Research On Pedestrian Detection Algorithm Based On Deep Learning

Due to the continuous development and advancement of computer vision,pedestrian detection has also been given a wide range of applications.At present,pedestrian detection mainly involves traffic monitoring,intelligent assisted driving,and intelligent robots.With the further study of researchers,pedestrian detection technology has developed rapidly.However,there are many influencing factors in the pedestrian detection process of the actual scene,such as the particularity of the camera that collects the data(fisheye camera,panoramic camera),the influence of the light(natural lighting and artificial lighting),the occlusion of objects and pedestrians,and the multi-scale of pedestrians,which poses a huge challenge for pedestrians recognition.This paper focuses on the pedestrian detection of fisheye images and the multi-scale problem of pedestrians,the main work and innovations are as follows:(1)Pedestrian detection has become an important research topic in the field of intelligent video surveillance.Fisheye lens is an effective video monitoring tool,which is characterized by wide field angle.Due to the special imaging principle of fish-eye lens,the edge of fisheye image has serious distortion,which poses higher requirements and challenges to the pedestrian detection technology of fisheye image.To solve the problem of fish eye image edge distortion,an effective rotation cutting method is proposed to process training and test data.The rotation cutting method divides fish eye image into edge image and center image according to the specific cutting point and rotation angle.Pedestrians in the center of fish eye image are mostly based on head and shoulder imaging,and most of the edges are distributed in the pose and contour of the pedestrians.The feasibility and effectiveness of the cutting method are verified by the traditional pedestrian detection algorithm HOG+SVM and the deep learning target detection algorithms(SSD,YOLOv3 and Faster R-CNN).The experimental results show that the rotary cutting method proposed in this paper effectively improves the pedestrian detection effect based on fisheye images.(2)Pedestrian detection technology has been widely used in practical scenarios nowadays,however,whether for ordinary images or fisheye images,different scales of pedestrians appear in the same photo or video due to the factors,such as the focal length of the camera,the distance between pedestrians and cameras,and the height and shape of pedestrians.To solve the multi-scale pedestrian problem,the paper proposes a multi-scale pedestrian detection algorithm based on improved YOLOv3.The algorithm uses the K-Means clustering algorithm to modify the size of the original anchor box in YOLOv3 to match multi-scale pedestrians and use the appropriate anchor box for pedestrian detection in feature maps of different scales.In order to verify the feasibility and effectiveness of the algorithm,the paper conducts experiments on the INRIA dataset and the USC dataset.The experimental results show that the improved YOLOv3 algorithm is superior to the original YOLOv3 in detection performance,and the mean average precision of the improved YOLOv3 is increased by7.14% and 6.82% on INRIA and USC datasets,respectively.