A Lightweight Method for Detecting the Behavior of Electric Bicycle Entering Elevators Based on Deformation Multi-scale Collaboration

Aiming at the problems that existing electric bicycle entering elevators behavior detection algorithms are difficult to adapt to geometric deformation, local exposure and occlusion lead to poor model feature extraction capability, and deformation, multi-scale feature perception and computational efficiency have poor synergy, this paper proposes a lightweight electric bicycle entering elevators behavior detection model based on deformation and multi-scale collaboration. By improving the structure of backbone network, neck network and prediction network of YOLOv5s baseline model, the synergistic optimization of the accuracy and speed of the electric bicycle entering elevators detection model is realized. Experimental results on a self-built electric bicycle entering elevators detection dataset show that the proposed algorithm achieves a mean Average Precision (mAP) of 96.59% and a frame rate of 32.0 frames per second (FPS), representing a significant performance improvement compared to the baseline model.