The Unmanned Aerial Vehicle Ground Targets Detection Technology Based On Deep Learning

The increasing number of motor vehicles has led to road congestion and frequent accidents.In order to alleviate the increasingly serious traffic congestion and realize the intelligent management and control of traffic flow during peak hours,it is urgent to carry out effective monitoring and intelligent analysis and research of vehicles on main roads.The Unmanned Aerial Vehicle(UAV)ground traffic monitoring system has the advantages of low cost,strong maneuverability,clear imaging,wide coverage,etc.,which obtains ground vehicle information quickly and efficiently and realizes effective acquisition of traffic flow information.However,due to the long-distance detection of the UAV,there are many types of vehicles,large changes in target scale,and complex backgrounds.Existing deep learningbased target detection methods have problems such as large demand for training samples,incomplete feature extraction,and lower detection accuracy,which makes it impossible to achieve effective monitoring of ground vehicle targets.Based on this,this paper focuses on the research of UAV ground vehicle detection methods based on deep learning.The main work is as follows:(1)Aiming at the problem of poor generalization ability caused by insufficient samples and imbalanced classes,this paper studied and implemented a high-fidelity data augmentation method based on self-attention generative adversarial network.Introduce the spatial selfattention mechanism into the generative network to improve the ability of extracting local features by convolutional layer,the fidelity of the generated target and the fusion of the background.Besides,add a down-sampling module and an up-sampling module at the input and output of the network respectively,in order to solve the problem of difficult highresolution image generation.At the same time,a multi-scale block adversarial network is used for modifying the generated results,which reduces the computational cost while not damaging the quality of sample generation.Several groups experiments shown that the supervised method proposed in this paper can effectively realize the feature conversion from pixel to pixel,and the unbalanced samples generated in the preset noise area of the image have higher fidelity,which effectively improves the performance of the existing target detection methods.(2)Aiming at the problems of few available features and low detection accuracy caused by large changes in target scale and complex backgrounds in UAV ground images.this paper studied and implemented an accurate target detection method based on sparse priors for multi-scale UAV ground vehicles.The detection model based on sparse priors is used for eliminating the influence of dense priors on the detection results and reducing computational complexity,At the same time,the weighted bi-directional feature pyramid network(BiFPN)is used for improving the feature extraction ability of the network.Besides,this method introduces switchable atrous convolution instead of conventional convolution.which reduces feature loss and improves detection accuracy.Several groups experiments shown that the method proposed in this paper has a better detection effect.(3)In order to verify the effectiveness of the proposed method in UAV ground vehicle monitoring,this paper designed and implemented ground vehicle detection software based on UAV platform.The software uses public and self-built UAV image data for offline training,and deployes the trained model on the embedded platform terminal.At the same time,outdoor experiments were carried out to test the functions of online software and verify the feasibility of algorithm.The results proved the effectiveness of the method proposed in this paper.