Research On Pedestrian Deiontect And Trackng Of UAV Based On Deep Learning

Combining UAV technology with target detection and tracking algorithm for real-time target detection and tracking has become a hot topic in UAV application field.It is of great significance to use UAV in civil field to carry out security inspection and early warning of public security emergencies in dense scenes.With the rapid development of artificial intelligence technology,deep learning methods train deep neural networks through a large amount of data to express image information features.Compared with traditional methods such as HOG and SIFT,deep learning object detection makes up for weaknesses such as window redundancy and manual feature extraction,and has outstanding advantages such as higher detection accuracy,feature expression ability and stronger reliability.In a crowded environment with pedestrians,UAV is easy to miss target detection and fail to track pedestrians due to dim light,mutual occlusion of pedestrians and small target under the perspective of UAV.Therefore,a series of researches are carried out on pedestrian target detection and following methods of UAV in complex environment.The main research work of this paper is as follows:The main content of pedestrian detection and tracking system of UAV is to improve the accuracy of algorithm.This paper studies related algorithms of detection and tracking system.By comparing and analyzing the advantages and disadvantages of mainstream algorithms and basic theories,this paper selects YOLOv5+DeepSort algorithm as the basic framework of pedestrian detection and tracking.Firstly,the theory and method of convolutional neural network are studied.An improved YOLOv5 target detection algorithm is proposed to solve the problem of low detection accuracy caused by insufficient feature extraction of small pedestrian targets with little pixel information and weak feature representation in complex environment from a high-altitude perspective.Adding a small target detection layer in the network enhances the feature information fusion between high and low layers.Based on CBAM,an improved channel-space(CAM-SAM)attention mechanism is proposed.By changing the connection structure of the channel and space,the attention weight is allocated to the feature graphs of different sizes.The non-maximum suppression of the original NMS is replaced by Gaussweighted Soft-NMS in the predictive network to solve the problem of the overlapping occlusion target being mistakenly removed.Moreover,the data set Vis Drone-2019 and the training results of the extended pedestrian data set MOT16 are used to verify the effectiveness of the improved YOLOv5 algorithm in improving the accuracy value of pedestrian small targets.Secondly,the multi-object tracking(MOT)algorithm based on DeepSort is studied,and the differences between generative and discriminant tracking models are expounded.Finally,the multi-object tracking method based on deep learning is selected.Aiming at the problem of non-uniform motion such as pause and acceleration in the process of pedestrian target tracking,the acceleration parameter is introduced into the Kalman filter to form the uniform acceleration model.When pedestrians are in crowded scenes,they tend to block each other and cause the IDswitch problem,which leads to the failure of matching track.Therefore,DIOU association matching is proposed to replace the traditional IOU matching method.The improved method improves the accuracy of pedestrian tracking algorithm.Finally,a pedestrian detection and tracking system is designed using the hardware and software platform of UAV.Moreover,simulation verification and actual flight test were carried out on the autonomous UAV software platform based on the improved YOLOv5+DeepSort algorithm.