Research On Vehicle Detection Algorithm In Aerial Images Based On Deep Convolutional Neural Network

In recent years,with the rapid growth of the number of vehicles,road traffic has become increasingly busy,and traffic management and planning are facing great challenges.Aerial vehicle detection offers the potential to acquire precise vehicle perception information from images,providing real-time,efficient,and accurate data support for traffic management and planning.However,aerial images possess unique characteristics of high-altitude and long-distance shooting,diverse shooting angles,and wide coverage angles,which leads to problems such as the interference of geographic information,a high proportion of small objects with weak features,and diverse vehicle directions and attitudes,thereby affecting the performance of aerial vehicle detection.To address these challenges,my thesis examines the vehicle detection algorithms based on the deep convolutional neural network YOLOv5 and proposes two vehicle detection algorithms suitable for aerial scenes.The primary contents and innovations of this thesis are as follows:Addressing the challenges posed by diminutive size and feeble discriminative feature of most vehicles in aerial images,my thesis presents a small object detection algorithm for aerial vehicles based on YOLOv5.The algorithm improves the multi-scale detection structure by combining the shallow feature map with the up-sampled feature map,adding a 4 times down-sampled feature map for small vehicle detection.Moreover,the SPD-Conv(space-to-depth)is employed to optimize down-sampling operation and augment the network’s capacity to express the features of small vehicles,circumventing the incremental loss of small object features during forward propagation.Finally,the Bi C(Bi-directional Concatenation)feature fusion module is incorporated to fuse the feature maps of three contiguous layers,further conveying the locational data of small objects,and improving the detection performance of small vehicles.Addressing the issue that object detection algorithms based on horizontal bounding boxes struggle to deal with vehicles with arbitrary rotation angles,my thesis presents an aerial rotating vehicle detection algorithm based on YOLOv5.The algorithm appends an angle classification prediction branch to the Head Network of YOLOv5,rendering it appropriate for detecting rotating vehicles in aerial images.Subsequently,to suppress the interference of background information and highlight the features of vehicles,the Diamond Mapping Unit module is designed to refine the feature extraction procedure of the shallow backbone network,minimizing information transmission loss.Concurrently,the adaptive spatial feature fusion mechanism is incorporated into the Head Network,effectively merging feature maps of different scales,filtering contradictory and conflicting information,and bolstering the hierarchy of feature information.Finally,to improve detection efficacy,a lightweight feature extraction network named Cheap YOLO Network is constructed to diminish the network’s computational loads and parameters,realizing the lightweight detection of aerial rotating vehicles.In my thesis,the two proposed object detection algorithms are tested and analyzed using the aerial datasets Vis Drone-DET2021,Drone Vehicle,and UCAS-AOD.Concerning the small object detection algorithm for aerial vehicles based on YOLOv5,the optimal performance is achieved in the multiple algorithms on the Vis Drone-DET2021 dataset and m AP(mean Average Precision)reaches 41.6%,which is8.6% higher than YOLOv5;Regarding the aerial rotating vehicle detection algorithm based on YOLOv5,m AP reaches to 73.61% for infrared images and 69.43% for visible images on the Drone Vehicle dataset and the AP is 89.02% on the UCAS-AOD dataset,which is higher than other advanced object detection algorithms.When utilizing the Cheap YOLO Network for detection,the amount of calculation and parameter of the network decrease by 20.79% and 16.57%,respectively,and the real-time detection time of each image is only 17.8ms.In summation,the algorithms proposed in my thesis exhibit commendable detection outcomes in aerial scenes,possessing tangible research significance.