Research On Target Detection And Tracking Based On UAV Autonomous Driving

In recent years,with the development of science and technology,UAV plays an important role in more and more fields.Due to its small size and strong mobility,it is suitable for image and video collection in areas where people are difficult to carry out actions.At this stage,UAVs are widely used in various military and civil fields.People have put forward more and more requirements for the actual application effect of UAVs,so more and more UAVs auto drive system systems appear in people’s vision.Target detection and tracking is an important part of UAV automatic driving process.Its accuracy and detection speed directly affect the practicability of UAV auto drive system,which has important scientific research value and practical value.Target detection and target tracking related to UAV usually face the following problems: The target is too far away from the UAV,resulting in the small target collected;In the process of data collection,there will be problems such as occlusion,defects in the traditional depth learning network,and insufficient correlation of object position information between frames in the tracking process.Aiming at these conspicuous problems,this paper studies the algorithm of target detection and target tracking in the scenario of UAV autonomous driving.The main innovations of this paper are as follows:(1)A small target detection method based on Yolov5 is proposed to solve the problems that the features of each layer in the network are not fully utilized and there is no due context relationship between the feature layers.Firstly,Mosaic is used to enhance data focusing on small targets,increase the number of small targets,and perform random erase to simulate the occlusion problem of targets in practical applications.After that,the original three-layer FPN network is changed into four layers,and then the context feature fusion module is used to process the image features of the last three layers.Finally,the shape loss function is used to regression the training parameters.Experiments show that the proposed method is better than the traditional Yolov5 network in small target detection,and has better performance compared with other small target detection networks.(2)A Yolov5 target detection network based on block relationship enhancement is proposed,which solves the problem that the original baseline network only focuses on extracting high-quality block features and ignores the relationship between blocks.Firstly,an attention-mechanism module based on region division is proposed to learn the relationship between the blocks while learning the internal features of the block image on the network.Then,Softmax function is used to preserve the high value information of the pooled area without losing the low value information.Finally,the SPP module is modified to decompose the large kernel pooling operation into the same small kernel pooling operation.Reduce network parameters and improve network efficiency.Experiments show that this method supplements the defects of baseline method and has better results compared with other methods.(3)A target tracking network based on improved Transformer is proposed,which solves the problem that the existing traditional twin network framework is easy to fall into a local linear matching process,resulting in the loss of semantic information,and only focuses on local optimization.The method first modifies the attention structure of Transformer to reduce the truncation error of the overall network,and then the convolution module is added to the Transformer network to improve the local feature enhancement effect.Finally,the feature splicing operation of adjacent frames is carried out in the attention network of Transformer to enhance the feature timing.A target tracking network with enhanced feature timing is formed.The experimental results show that the proposed method can improve the effectiveness of the existing open data sets compared with other methods.