Research On Vehicle Target Detection Algorithm Of UAV Aerial Image Based On Deep Learning

UAV(Unmanned Aerial Vehicle)aerial image vehicle target detection is the core technology of vehicle tracking and traffic monitoring.It is broadly used in highway inspection,traffic flow monitoring,and other fields.Unlike universal target detection,there are problems such as significant differences in the scale of vehicle targets,more small target vehicles,complex vehicle background information,and uneven distribution of the number of vehicle target categories in UAV aerial images,which cause difficulties for vehicle target detection in UAV aerial images;in addition,due to the small size of UAVs and limited computing resources,most target detection models based on deep neural networks are difficult to meet the UAV edge deployment requirements for model lightness and realtime.In this paper,we propose a series of deep learning-based target detection models from the difficulties mentioned above and effectively improve the problems of significant differences in the scale of vehicle targets,more small targets,and uneven distribution of vehicle target categories through the targeted design of data processing,feature fusion,and target detection head in the models,so that they can be better applied in engineering.The main contributions of this paper include the following:1)For the problems of more small vehicle targets,more significant differences in vehicle target scales,and complex vehicle target background information,this paper proposes an anchor frame adaptive High-definition feature fusion model for vehicle target detection of UAV aerial images.First,the model adaptively adjusts the size of the model anchor frame according to the target scale distribution in the UAV aerial image dataset,adapts to different vehicle target scales,and improves the target detection accuracy.Secondly,we propose a novel upsampling module to improve the accuracy of small target detection.Finally,we use a channel attention mechanism to reduce the interference of background information.The experimental results show that the above vehicle target detection algorithm can improve to a greater extent the detection problems caused by the significant differences in vehicle target scales,more small targets,and complex background information in the UAV aerial images in natural scenes.2)To address the problems of insufficient consideration of shallow information transfer and spatial information in the anchor frame adaptive High-definition feature fusion model and difficulty in fully satisfying the scale discriminative characteristics of vehicles,this paper proposes a dual attention mechanism and a multi-scale High-definition feature fusion UAV aerial vehicle target detection algorithm.Firstly,the algorithm designs a feature fusion network based on a jump connection and double attention mechanism,which improves the transfer of shallow feature information of the target using a jump connection.Secondly,the algorithm designs a double attention mechanism to enhance the consideration of feature spatial information and attenuate the interference of background information.Finally,the algorithm designs a multi-head detection module to further enhance the discriminative properties of the model in vehicles at different scales.The experimental results prove that the described algorithm significantly enhances the anti-interference ability and selfadaptive capability of target detection.3)In response to the requirements of real-time and lightweight models for edge deployment of small and low-power devices,this paper proposes a multi-scale UAV aerial image vehicle target detection model based on Transformer joint image enhancement.First,the model selects a lightweight single-stage detection model as the baseline.Secondly,the model fuses the spatial and channel information of vehicle features to highlight vehicle features,attenuate the influence of background noise,and improve the detection accuracy of small targets by the Transformer module.Finally,the image enhancement module is proposed to increase the number of samples in the data set with few sample categories without increasing the complexity of the model,to balance the number of various types of targets,and to reduce the image input size to ensure real-time model inference.The experimental results show that this model can significantly improve the overall performance of vehicle target detection and meet edge deployment requirements.It can meet the requirements of realtime and is lightweight for edge deployment.Order to make the algorithm described in this topic can be better deployed on low computing power,low power consumption,and low-cost UAV devices.In this paper,we build an end-to-end vehicle target detection framework for UAV aerial images on Jeston Xavier NX and i TOP-RK3588low-power development board,respectively,and optimize and compress the final test the actual operating performance of the above vehicle target detection model.The experiments show that the model can still show high detection speed and accuracy on UAV devices with limited computing power and performs well in different application scenarios.Compared with the workstation detection method,deploying the UAV edge can significantly reduce the cost of vehicle target detection and improve the flexibility of accomplishing practical engineering tasks,expanding the scope of the model’s engineering applications and laying a solid foundation for possible future practical engineering applications.In summary,the model described in this paper is characterized by solid anti-interference,high stability,lightweight,and balance of accuracy and efficiency for UAV aerial vehicle target detection,which can improve the problems of significant differences in the scale of vehicle targets,more small target vehicles,complex vehicle background information,uneven distribution of the number of vehicle target categories and difficulty in deploying the model at the edge end of UAV in actual engineering,and according to the actual The hardware deployment framework of the vehicle target detection model on edge,the device is designed according to the application needs of the project,which provides a new idea for the engineering application of UAV aerial vehicle target detection in complex environments.