Vehicle detection in CCTV with global-guided self-attention and convolution
摘要
Vehicle detection in closed-circuit television (CCTV) footage has long been a significant area of research due to the unique challenges posed by the nature of camera placement. CCTV cameras are typically mounted at oblique angles, leading to frequent occlusion in dense traffic scenes. Additionally, the varying distances of vehicles from the cameras, combined with the different angles at which the cameras are positioned, introduce the problem of scale variation. To address these challenges and enable accurate and efficient vehicle detection, we propose Global-Guided Mixing (GGMix), a novel module that combines multi-head self-attention (MHSA) and depthwise convolution in a sequential structure designed to enhance feature representation. It integrates multi-attention mechanisms to better capture vehicle-related features while suppressing irrelevant background information. We introduce GGMix blocks to the backbone of YOLOv8 in order to propose VDC-YOLO, a real-time vehicle detection model. Extensive experiments on a proprietary dataset, MLITcctv, and two public datasets, BitVehicle and i2, demonstrate that our model achieves superior performance, in terms of mAP while balancing accuracy and efficiency. Specifically, VDC-YOLO achieved an mAP of 46.9