Research On Pedestrian Detection And Tracking Technology Based On Convolution Neural Network

Pedestrian detection and tracking have a very broad prospect in computer vision.It not only provides pedestrian trajectory in the fields of security monitoring,intelligent transportation,etc.,but also provides the basis for data analysis in IoT applications such as smart cities and new retail.This means that the environment in which we need to perform pedestrian detection and tracking is complex and diverse,and real-time is also necessary.This also refines the research direction of this thesis:based on video research pedestrian detection and tracking technology,design and implement a pedestrian counting system based on Raspberry Pi.This thesis studies object detection and tracking technology based on deep Convolution Neural Network(CNN),and focuses on pedestrian detection and tracking technology under real-time monitoring and introduces Multi-Object Tracking(MOT)algorithm into pedestrian counting.Inspired by the Mobilenet structure,we use Depthwise Separable Convolution Layer to replace the convolution layer of YOLO V3 Tiny.The improved YOLO V3 Tiny network,called Mobilenet-Tiny,is slightly faster than YOLO V3 Tiny network in object detection tasks.This thesis also studies the MOT.Based on SORT algorithm,we use Mobilenet-Tiny as the detector of MOT algorithm.The improved SORT algorithm called Mobilenet-SORT,it can achieve twice FPS speed of the tracking task when the accuracy of the tracking task is slightly less than that of the SORT algorithm.Finally,combined with previous work,Moblienet-SORT,an improved SORT algorithm,is introduced into pedestrian counting tasks.Combined with MOT,time threshold and region information,the algorithm can reduce the counting error of the number of pedestrians directly using SORT output trajectories.