YOLOv5s-FAC: enhanced feature association detector for person-vehicle counting in smart park

Currently, tracking by detection technique is widely used in crowd or vehicle counting. However, it is difficult to meet detection and counting with mixed pedestrian and vehicular traffic of smart park scenes in emergencies, for occluding small objects recognition problem. An improved detection model YOLOv5-FAC is proposed based YOLOv5s. First, a P2 detection layer is added to expand the detection range of the model and improve its detection ability of different sizes. Second, an auxiliary inference network is constructed using programmable gradient information to provide the model with a stronger information fitting capability. Finally, a cascading triplet attention mechanism is added to the head of model to increase the feature fusion capability. Then, a collision line counting method combined with OCSORT track technology is proposed, in which both direction movement of traffic is considered. The experimental results show that YOLOv5s-FAC has a significantly improved detection quality, with a mAP of 69.5%, and the counting accuracy reached 94.4%.