CoP-YOLO: A Light-weight Dangerous Driving Behavior Detection Method

With the continuous increase in vehicle ownership, the incidence of traffic accidents has also escalated, with 90% attributed to human aspect. To mitigate the impact of dangerous driving behaviors, this study introduces a lightweight detection method for hazardous driving behaviors based on visual perception. This research uses YOLOv10 as the baseline model, employing partial convolution to minimize unnecessary computational overhead and memory access, while integrating the coordinate attention mechanism to enhance feature extraction and improve the representation of regions of interest. The research achieves a significant reduction in model parameters and computational complexity, alongside an improvement in detection accuracy, culminating in an efficient system for monitoring dangerous driving behaviors. The system's performance is evaluated using a proprietary dataset, demonstrating that this method not only enables precise real-time recognition and detection of driving anomalies but also maintains a compact model size, and the inference speed can reach 87fps on the NVIDIA ORIN NX embedded device.