Pedestrian Detection And Pose Estimation Based On Complex Environments

Pedestrian detection and pose estimation are important research contents of deep learning.Pedestrian detection is an essential and important task in any intelligent video surveillance system,because it provides basic information for the semantic understanding of video clips.As it is possible to improve the safety system,it has obvious scalability for automotive applications.Human posture estimation is also a computer vision-based technology that can detect and analyze human posture.As a relatively basic task in computer vision,human pose estimation is the pre-task of human behavior analysis,action recognition,and human-computer interaction.In recent years,with the development of deep learning,pedestrian detection and human pose estimation have been greatly developed,but complex environmental factors such as clothing,light conditions,background,and flexible body structure all affect pedestrian detection and human body The accuracy and speed of the pose estimation algorithm are also problems to be solved urgently for the pedestrian detection and pose estimation algorithms based on deep learning.This article focuses on pedestrian detection and pose estimation in complex environments.The main tasks include:(1)Aiming at the problem that YOLOv3 has low accuracy in detecting pedestrians with large scale changes,this paper proposes a fast pedestrian detection method for deep fusion of feature maps based on the YOLOv3 algorithm.The improved YOLOv3 is a pedestrian detection algorithm that can detect pedestrian targets in real time and maintain a high detection effect in complex environments.(2)In view of the slow speed of the current OpenPose pose estimation algorithm and higher hardware requirements,the Mobile Net v1 feature extraction network with deep separable convolution is used,and a lightweight refinement stage with residual connections is designed.Optimize OpenPose to increase the ratio of accuracy to network complexity by more than 6.5 times,while maintaining better estimation results in complex environments.(3)Finally,a bank anti-theft system and a fall detection system were developed based on the pedestrian detection algorithm based on YOLOv3 and the improved OpenPose algorithm proposed in this paper.