Detection Model for Small Vehicle Targets from Unmanned Aerial Vehicle Perspectives

To address the challenges of vehicle target detection in aerial images from Unmanned Aerial Vehicle (UAV) perspectives, this paper proposes a novel vehicle detection model named UAV-YOLO, based on the YOLOv7 object detection framework. First, Omni-dimensional Dynamic Convolution (ODConv) is introduced into the backbone of the feature extraction network to enrich the acquisition of small target information, enhancing the network's capability to extract features of small objects. Second, the Mixed Local Channel Attention (MLCA) module is added to the ELAN and feature pyramid structures within the backbone feature extraction module to integrate both local and global information, further improving the model's perception and information fusion for small targets. Lastly, a joint loss function, Normalized Wasserstein Complete IoU Loss (NWC-IoU Loss), is designed to enhance the model's detection performance and stability for small targets. Experimental results show that the improved detection algorithm increases the mean Average Precision ([email protected]) by 4.8%, reduces the FLOPs by 27.3%, and achieves a detection speed of 62.2 FPS, meeting real-time requirements. The proposed method demonstrates superior detection performance in various complex scenarios of the VisDrone2019 dataset. The UAV -YOLO model exhibits excellent performance in vehicle target detection in aerial images from UAV perspectives.