Research On Dense Pedestrian Detection Algorithm Based On Overhead Fisheye Images

Fisheye cameras are a special type of camera used in video surveillance equipment,with advantages such as a wide imaging range and a low probability of pedestrian occlusion.As a result,fisheye cameras are widely used in both densely populated open areas and small enclosed spaces.However,due to the inherent radial distortion and radial arrangement of fisheye cameras,fisheye image object detection technology poses a greater challenge than traditional object detection technology.This thesis addresses the radial arrangement problem in fisheye images by introducing a rotation parameter to achieve end-to-end processing of fisheye images.Subsequently,based on the rotation and radial distortion properties of fisheye images,a dense pedestrian detection algorithm based on overhead fisheye images is proposed to solve the problem of fast and accurate identification by dense pedestrian detection algorithms for fisheye images.The main work includes:(1)In view of the natural radial arrangement problem in fisheye images,this thesis introduces a rotation parameter into the original YOLOv5 model and designs a fisheye image pedestrian detection algorithm based on rotation reconstruction through rotation representation,Smooth L1 loss function and PIo U.The YOLOv5 network model has gained the ability to perceive rotated box targets.Subsequently,this thesis compares and evaluates with related methods at home and abroad on the WEPDTOF dataset,and also conducts ablation experiments.The results show that the improved network model has certain advantages in various indicators such as P,R,m AP,indicating that the network model based on YOLOv5 rotation reconstruction has good results in detecting fisheye targets and basically meets the needs of fisheye image object detection.(2)In view of the problems existing in YOLOv5 when detecting fisheye distortion rotation targets,this thesis designs a fisheye image pedestrian detection algorithm based on correction processing by introducing a secondary classification strategy,rotating Mosaic data enhancement strategy,improved Bi FPN module,deformable convolution module and KLD loss function based on the rotation distortion characteristics of fisheye images.The algorithm achieved good results on the WEPDTOF dataset,with an average accuracy m AP@0.5 of 98.0% for all categories and an overall average accuracy m AP@0.5:0.95 of 81% for all categories,indicating that the network model based on YOLOv5 correction processing is suitable for object detection under fisheye images.Finally,through ablation experiments,the effectiveness of the above improvement points is proved.This thesis solves the need for fast and accurate identification in dense scenes of fisheye images.