Edge Computing Based Electric Bicycle Dangerous Riding Behavior Detection System

ABSTRACT To design and validate a real‐time, low‐cost edge system that detects multiple dangerous electric‐bicycle riding behaviors. We formulate the task as multi‐class object detection plus rider counting and implement it with an enhanced YOLO variant, SDV‐YOLO. Helmet status (Helmet/No_Helmet) and hand status (Grab/No_Grab) are treated as learnable classes, while overloading (rider count > 2) is inferred by associating person and vehicle boxes after class‐wise NMS. SDV‐YOLO integrates (i) SPDConv down‐sampling to preserve detail with fewer FLOPs, (ii) DySample adaptive up‐sampling for content‐aware multi‐scale fusion, and (iii) a VoV‐GSCSP neck for efficient channel aggregation. The processing pipeline is: frame capture → resize to 640 × 640 → normalization → INT8 inference (PyTorch → ONNX → Hailo HEF) → confidence filtering and NMS → behavior logic and alert output. On a 2041‐image, five‐class dataset, SDV‐YOLO raises recall by 10.6% over YOLOv8n, maintains mAP 50 = 0.748, and reduces FLOPs by 13.6%. Deployed on a Raspberry Pi 5 + Hailo‐8L, the system reaches 122 fps (6.8 ms latency) at a hardware cost of ≈USD 150, confirming scalability for roadside supervision.