Vehicle Detection In Aerial Images

Vehicle detection in aerial image is one of the research hotspots of object detection,and it has broad development prospects in smart cities,intelligent transportation,unmanned driving,and traffic control.Although the general object detection algorithm can efficiently detect targets in conventional ground shots,when directly applied to the vehicle detection in aerial image scene,the performance is significantly reduced.The main reasons include the aerial vehicles occupying less pixels,the background is complex,and the shooting angle mostly is the top view angle,the presence of occlusion,etc.In order to efficiently detect vehicles in aerial images,this paper proposes a multidepth hierarchical feature fusion aerial vehicle detection algorithm based on the characteristics of aerial images and previous studies.The specific contributions are as follows:(1)A backbone network based on the stepwise residual module is proposed,this backbone network outputs feature maps with multiple depths.The stepwise residual module divides the input feature maps according to channels.The feature maps of different channels pass through a different number of convolutional layers,and finally the convolutional layer with a selection mechanism selects feature information that is beneficial to detection.Through the above structure,the stepwise residual module can simultaneously obtain a shallow feature map containing detailed texture information and a deep feature map with deep semantics and position information.The surrounding background information in the shallow feature map has less pollution to the target vehicle information,and can transmit more vehicle detailed texture feature information to the subsequent processing process,thereby alleviating the detection caused by the fewer pixels of the vehicle in the aerial image and the complex background.The backbone network based on the stepwise residual module gives full play to the advantages of the stepwise residual module,and can uniformly generate feature maps processed by the 0 to 102 layers of convolutional layers,helping the detection algorithm to better detect vehicle targets.(2)A feature fusion structure based on deep projection deconvolution is proposed to assist in the up-sampling of low-resolution feature maps and to fuse feature maps of different depth levels.By introducing a deep projection unit,the feature fusion module can reduce the loss of key information in the process of down-sampling high-resolution feature maps to low-resolution feature maps,and generate feature information during the process of sampling from low-resolution feature maps to high-resolution feature maps.The combined use of the deep projection deconvolution up/down projection unit constitutes the internal super-resolution sub-module of the feature fusion module,so that the feature map retains more original information during the feature fusion process.The thesis designs a multi-scale detector based on the characteristics of feature fusion,and improves the loss function to solve the common problem of imbalance between positive and negative samples in one-stage networks.The algorithm proposed in this thesis has been verified in detail by UCAS-AOD,VEDAI and DOTA datasets.On the UCAS-AOD dataset,this thesis verifies the performance improvement of the backbone network based on stepwise residuals module and the feature fusion module based on deep projection deconvolution.After the joint action,the vehicle category detection AP reaches 97.34%.On the VEDAI dataset,the AP of the small land vehicle category reached 93.96%.On the DOTA dataset,this thesis verifies the multi-class detection effect,and the m AP of the three categories of boats,planes,and small vehicles reaches 89.86%.Experiments show that the performance of the algorithm proposed in this thesis can exceed other aerial image vehicle detection algorithms on the three datasets.