YOLO-DC for vehicle detection using deformable convolutional networks and cross-channel coordinate attention

Vehicle detection is crucial for intelligent decision support in transportation systems. However, real-time detection of vehicles is challenging due to geometric variations of vehicles and complex environmental factors such as light conditions and weather. To address these issues, the paper introduces the You-Only-Look-Once with Deformable Convolution and Cross-channel Coordinate Attention (YOLO-DC) framework that improves the performance and reliability of vehicle detection. First, YOLO-DC incorporates Cross-channel Coordinate Attention, which combines channel attention and coordinate attention, to more accurately cover target sampling positions and enhance feature extraction from vehicles of various shapes. Second, to better handle vehicles of different sizes, we employ Multi-scale Grouped Convolution to enable multi-scale awareness and streamline parameter sharing. Additionally, we incorporate channel prior convolutional attention so that the model can concentrate on areas of vehicles that are critical for detection. We also optimize feature fusion by leveraging a highly efficient fusion of C2f (CSP Bottleneck with 2 Convolutions) and FasterNet to reduce the model size. Experimental results demonstrate that YOLO-DC performs better than the state-of-the-art YOLOv8n method in detecting small, medium, and large-sized vehicles, and in detecting vehicles in adverse weather conditions. In addition to its superior performance, YOLO-DC also features fast detection speed, making it appropriate for real-time detection on devices with limited computational power.