SOD-YOLO: Detection Methods for Vehicles with Occlusion and Small Targets

UAV-based vehicle detection has important application value in the fields of intelligent transportation, urban security and emergency rescue. It is the key technology for realizing high-altitude situational awareness and dynamic monitoring. However, in this scenario, the vehicle targets generally have the problems of being small in size, densely distributed, and mutually occluded, which leads to the model being prone to feature loss, missed detections, and high false detection rates. To address these challenges, we propose a vehicle small-target and occluded-target detection method called the small object and occlusion detection method (SOD-YOLO), which is based on YOLOv13. This method includes three main innovations. First, we designed a sliding-window self-attention convolutional structure, a small object and occlusion detection convolution structure (SOD-C2f), which is based on the existing sliding-window self-attention (SWSA) structure. It is suitable for extracting features of small vehicle targets. Through the self-attention mechanism within the local window, we enhance the fine-grained expression of small target features, effectively improving the model's perception and positioning ability for tiny vehicle targets. Afterward, we integrate the coordinate attention mechanism (CA) by explicitly modeling the interchannel dependencies and spatial position information to achieve precise semantic alignment of features at different levels, significantly enhancing the detection robustness in occluded scenarios. Finally, we employ the Space-to-Depth Convolution (SPDConv) module, which splits the spatial dimension and replaces the traditional stride convolution and pooling layer with nonstride convolution. This approach avoids the loss of fine-grained features during downsampling and further retains the key spatial information of small vehicle targets. The experimental results show that the precision, recall, average mAP@0.5 and mAP@0.5:0.95 of the proposed SOD-YOLO algorithm are 77%, 74.8%, 80.9% and 54.9%, respectively. Compared with the benchmark model YOLOv13, the three indicators increase by 3.6%, 1.1%, 3.6% and 2.6%, respectively. In conclusion, the proposed algorithm performs better in the tasks of detecting small targets and occluded vehicles and is able to better meet the precise detection requirements of vehicle targets from a high-altitude perspective.