Research On Pedestrian Detection And Tracking System For Service Robots

Pedestrian target detection and tracking is an important field of computer vision research.With the development of deep learning and artificial intelligence,pedestrian detection and tracking technology is widely used in areas such as autonomous driving,drones,and intelligent surveillance.In recent years,my country's service robot industry has developed rapidly,and intelligent public service robots have become more and more popular,and higher requirements have been put forward for the intelligent perception technology of service robots.This paper takes pedestrian detection and tracking as the research object,studies pedestrian detection,tracking and other perception algorithms,and transplants the algorithms to service robots to provide help for service robots to achieve precise services and active avoidance operations in complex scenes.First of all,for the pedestrian detection algorithm,considering the limited computing power of the service robot equipment and the high requirements for pedestrian detection accuracy and speed in complex scenarios,this article uses the YOLOv4 algorithm as the research object,referring to the network structure of the YOLOv4-Tiny algorithm,and the basic network of YOLOv4 Lightweight improvements were made,and the INRIA data set was manually relabeled.The Canopy algorithm and the K-means++ algorithm were combined to reset the initial value of the anchor frame required by the improved YOLOv4.The improved YOLOv4 and other advanced algorithms are compared and tested on the public data set.The improved YOLOv4 algorithm is 70% faster than YOLOv4 in speed,which meets the higher requirements of service robots for detection speed.Compared with YOLOv4-Tiny,the detection accuracy is increased by 8.21%.Through subjective experiment analysis,it is concluded that the improved YOLOv4 can detect pedestrians in complex scenes well.Then,for the multi-pedestrian target tracking algorithm,in order to meet the real-time tracking needs of the service robot for multiple pedestrian targets,this paper takes the Deep SORT algorithm as the research object,and uses DIOU instead of IOU to join the cascade matching to deal with the occlusion problems and omissions in the pedestrian target tracking.To check the problem,use the improved YOLOv4 target detection algorithm to replace the original Faster R-CNN target detection algorithm in Deep SORT to improve the tracking accuracy and running speed of the Deep SORT algorithm.Compared with Deep SORT,the experimental test shows that the tracking accuracy of the improved Deep SORT is increased by 18%,and the speed is increased by 68%.In the comparison of detection and tracking results in the same scene,it is found that the improved Deep Sort algorithm can effectively improve the detection and tracking capabilities of pedestrians at medium and long distances and improve the problem of pedestrian missed detection under high occlusion.Finally,for the pedestrian detection and tracking system,the overall framework of the system was designed,and the software and hardware platform of the system was completed.The MOT16 data set and the actual campus scene of Nanjing University of Information Science and Technology are used as the test set.Different scenarios such as pedestrian occlusion,dense crowds,pedestrians at night,and light changes are selected to test the system.The experimental results show that the system always detects and tracks accurately,and there is no missed detection.In addition to the occurrence of misdetection,the robustness is good,the processing frame rate is maintained at 8 to 10 frames,and the system runs stably,which basically meets the application requirements of service robots in real-time scenarios.