In heavy-duty vehicle automatic parking technology, accurate parking space detection is a critical prerequisite for reliable parking. However, existing object detection methods perform poorly in complex parking lot environments. They struggle to effectively handle issues such as blurred parking space lines, occlusions, and diverse layouts. To address this, this paper proposes the YOLO-HG detection framework and designs a specialized Hierarchical Global Perception (HGP) attention mechanism. This mechanism uses non-local self-attention to model global spatial relationships. It combines geometric feature extraction and direction-sensitive filtering techniques to accurately capture the regular arrangement and boundary features of parking lots. Through a channel-spatial joint attention mechanism, it effectively fuses multi-dimensional features such as color and edges to improve feature extraction performance. Additionally, this paper integrates ODConv convolution into the C2f module to optimize the efficiency and effectiveness of feature extraction. Meanwhile, it adopts the WIoU loss function to further improve bounding box localization accuracy. The YOLO-HG built on the YOLOv8 architecture achieves excellent performance on our self-built heavy-duty vehicle parking space dataset. Experimental results show that compared to YOLOv8, this method improves mAP50 and mAP50–95 by 1.3% and 1.2% respectively. It maintains real-time inference speed, validating its effectiveness and practicality in complex parking scenarios.