Research And Application Of Video Pedestrian Detection And Tracking Algorithm

As an important branch of computer vision field,object detection and tracking play an extremely important role in intelligent industry,intelligent construction site,intelligent security and intelligent traffic.Nowadays,researchers have made great progress in the field of target detection and tracking,but there are still some improvements to be made in the field of target detection and tracking in complex scenes,such as small target detection and algorithmic industry implementation.This thesis mainly studies the accuracy of pedestrian detection and tracking and the application of the two algorithms in the Hi3516DV300 development board.The main research work is as follows:(1)An improved pedestrian detection algorithm of YOLOv5 is proposed.Firstly,we propose that adding small and medium targets detection layer to the original YOLOv5 network can improve the detection of small and medium targets.Secondly,we put forward a new Channel Spatial Coordinate Attention Module(CSCAM),which combined the effective information of channel,space and position.In the course of the experiment,it was found that the above two improvement points in Pascal VOC2012 and VOC2007 pedestrian data set achieved a certain improvement effect compared with the basic experiment.(2)An improved DeepSort pedestrian tracking algorithm is proposed.Omni-Scale Network(OSNet),an excellent re-recognition network,is used to replace the original simple feature extraction network of DeepSort,which can reduce the object tracking errors in the case of occlusion.For some cases,the IOU matching method in DeepSort could not match the detection box and tracking prediction box correctly.The popular EIOU was applied to the DeepSort tracking algorithm to optimize the matching process.The experiment is carried out in the open source multi-target tracking data set Mot16 and Mot17 and the improvement effect is obvious.(3)Finally,a passenger counting sensor application software is developed by using the above pedestrian detection and tracking algorithm.This application software takes Hi3516DV300 as the development board,collects images through the TOF camera hardware device,takes the collected images as the input of the software algorithm,and stores the output results of the algorithm in Hisi development board.Various commands,such as program initialization,program start counting,program stop,program upload counting,video upload,etc.can be sent to the Hisi development board through the host,and the Hisi development board will return the required information to the host.In addition,the YOLOv5 detection and tracking algorithm is written in Python language.Due to the Hisi development board only supports C++ language,and the Hisi development board does not support the operators in the original YOLOv5 detection algorithm,such as focus slicing operation.When deploying the Hisi development board,the YOLOv5 network was rewritten to C++ and unsupported operators in the network needed to be modified.The actual application scenario of this thesis is to count the flow of people in the subway.In order to protect the privacy of passengers,depth camera is used.TOF camera is used to measure the depth image within the scope of the train door entrance and the carriage passage,and the algorithm is used to identify people and count the number of passengers entering and leaving the carriage.