Research On Pedestrian Detection And Tracking Method Of Security Patrol Robot

In this paper,the pedestrian detection and tracking method of security patrol robot is studied.A model of pedestrian detection and tracking for security patrol robot based on computer vision is proposed,which can detect pedestrians and the knife or lethal weapons they hold in real time.A pedestrian real-time tracking system and a social distance detection model are established.This paper will test the performance of mainstream object detection methods in pedestrian detection.A pedestrian dataset is established to train the pedestrian detection model.Then,the detection accuracy and speed of Faster R-CNN,YOLOv3 and YOLOv4 models on the pedestrian dataset in this paper is evaluated.The YOLOv4 model has the best comprehensive performance for pedestrian detection.A security dataset containing pedestrians,knives,hammers and axes are established to train the detection model.The improvement of the YOLOv4 target detection model in the network structure,the model optimization strategy and the composition of the loss function are analyzed.K-means clustering is used to obtain the best initial anchor frame parameters,and YOLOv4 multi-target detection model is trained to optimize the detection effect.In order to adapt to the application scenarios of mobile devices such as security patrol robots,a lightweight YOLOv4-tiny detection model is proposed and compared with YOLOv4 after training.In the security dataset of this paper,YOLOv4 achieves 86.04% m AP and 47 FPS detection speed,YOLOv4-tiny obtains 71.14% m AP and 143 FPS detection speed.The test result shows it can accurately and quickly detect pedestrians and weapons they held in pictures and videos.The SORT tracking algorithm and the principle of Kalman filter and Hungarian data association are discussed.Then,the pedestrian tracking system of security patrol robot based on YOLOv4+Deepsort is constructed by using Deepsort pedestrian tracker,and the real-time pedestrian target tracking above 20 fps is realized.The detector can be replaced with YOLOv4-tiny to obtain a detection speed up to 80 FPS.Furthermore,the function of pedestrian trajectory visualization is added.The pedestrian social distance detection model is constructed by combining the physical distance approximation of video pixels.