A pedestrian detection method in unattended garage based on improved YOLOv7

In the intelligent stereo garage system, automatic detection of people in the garage area is crucial to realize unattended stereo garage and improve the safety of garage operations and access efficiency. In actual operating scenes, pedestrians are often blocked by vehicles or other background objects, which has become a key factor affecting its detection accuracy. Based on the current YOLOv7 framework with strong comprehensive detection performance, this paper proposes a detection algorithm named Cross-YOLOv7 that combines a new attention mechanism and cross-layer feature fusion. First, a new attention module named Coor-NAM is proposed to enhance the feature extraction of pedestrian target areas in garage scenes, especially unoccluded areas. Second, a new non-adjacent scale cross-layer connection method is designed to significantly improve the information interaction between features of different scales. It can enrich the semantic information of the features, and thus improve detection accuracy. The proposed method is evaluated on the garage pedestrian detection dataset and the public WiderPerson dataset. Experimental results show that compared with several typical detection methods, the detection accuracy (AP50 and F1) of blocked pedestrians in the garage scene with our proposed Cross-YOLOV7 can be improved by 4.6% and 5.66% respectively. Moreover, the detection accuracy on the public WiderPerson dataset can also be improved by 2.5% and 1.68%.