Drone-TOOD: A Lightweight Task-Aligned Object Detection Algorithm for Vehicle Detection in UAV Images

Vehicle object detection using UAV images is a crucial undertaking in urban traffic management and the advancement of autonomous driving technologies. Conventional networks fail to achieve accuary in detecting vehicle objects from a drone’s perspective due to the significant variations in the size of the target items, unequal distribution of their positions in the image, and image degradation induced by the drone’s movement.In order to surmount this challenge, this research suggests an enhanced TOOD object detection model named Drone-TOOD. The first proposal is to create a lightweight network skeleton called CSPRegNet by merging CSPblock with Regblock. Secondly, we incorporate Regblock into CSPPAFPN to enhance CSPRegPAFPN and incorporate EVCblock at the upsampling location of deep features to capture corner area details and minimize the degradation of feature information. In addition, a efficient task decomposition attention module is also proposed to enhance the interaction ability of positioning and classification tasks. This task decomposition module can highlight the characteristics of a specific task while retaining the characteristics of another task, thereby improving detection capabilities. Experiments conducted on the Drone Vision Challenge Benchmark (VisDrone) demonstrate that the enhanced model can obtain superior performance compared to TOOD. The average precision (mAP) achieved by our approach is 64%, surpassing TOOD by 7.9%. The frames per second (FPS) stayed constant at 27.2. Drone-TOOD demonstrates superior performance compared to other lightweight models on the VisDrone-2021 dataset. In order to demonstrate the robustness of our approach, we additionally performed ablation experiments and conducted tests on the UAV Detection and Tracking Dataset (UAVDT), resulting in an achieved mean average precision (mAP) of 64.6%. Furthermore, Drone-TOOD possesses a total of parameter approximately 21.7 M.