Learning-based Vehicle Detection And Tracking Under UAV System

UAVs have been widely used in the civilian field due to their small size,light weight,easy concealment and flexible control.At present,traditional detection and tracking of target technology cannot meet the requirements of target detection and tracking tasks in dynamic video under UAV platforms.This paper takes vehicles as the targets to be tested and uses deep learning technology to conduct in-depth research on detection and tracking of vehicles methods under UAV system.The main work of the thesis is as follows:(1)Research on improved method of vehicle detection based on YOLO network.Combining the feature extraction networks of different connection methods,a network model based on YOLO is reconstructed for real-time detection of vehicles under the aerial image of UAV.Firstly,to preserve the fine-grained features of the small target in the horizontal direction,the horizontal resolution of the feature map is increased by adjusting the girds’ number setting.Secondly,the K-means clustering method is used to determine the size and number of anchor boxes in the network,which is used to match the size of the detected vehicles,so as to reduce the computational complexity and improve the network’s capability of prediction.Then,for the problem of random initialization network training slow and difficult to converge,the ResNet-v1-50 pre-training model is selected to replace the shallow structure of the original model.In the deep structure of the network,a dense topology with different connection modes is adopted to strengthen the feature extraction capability of small target.In addition,in order to improve the efficiency of network pooling,a learnable weighting parameters are used to balance the maximum pooling and average pooling.The experimental results of detection of different weather,different road sections and different road directions show that the improved detection method can detect different types of vehicle such as cars,buses and trucks,and the detection accuracy reaches 89.2%,which can meet requirements of real-time detection.(2)A tracking method of vehicles that combines IoU and LSTM networks.In the case where the vehicle does not overlap in the aerial image,the IoU information between the video interframe detection frames is fully utilized to guide the tracking of the vehicle based on the results of the detection network.In the case of partial occlusion of vehicles in the video image,the ResNet-18 feature extractor is used to extract the apparent information of vehicles in the historical frames,and then the LSTM network is used to match and predict the position of the current vehicle.The tracking results of video images in different weather conditions show that the proposed tracking method of vehicles based on IoU and LSTM networks has stable tracking performance,high robustness and can track in real time.