Red-light running by non-motorised vehicles seriously disrupts traffic order and poses significant safety hazards. To address this problem, this paper proposes a system for detecting red-light running by non-motorized vehicles, using an improved YOLOv5-DeepSORT framework based on video analysis. The system is constructed based on a dataset containing eight types of targets and uses an Inverse Frequency Weighting (IFW) strategy to address the sample imbalance problem. At the algorithmic level, the YOLOv5 detector is deeply fused with the DeepSORT tracker, and a complete red-light running behaviour determination logic is constructed by introducing a Selective Tracking Mechanism (STM), a Motion Trajectory Analysis Module (MTAM) and a Signal Priority Decision Mechanism (SPDM). The system achieves a processing speed of 25 Frames Per Second (FPS), with accuracies of 92.3% for direction judgement, 88.7% for red-light running under ideal lighting, and 81% on average across different weather conditions, with a low misdetection rate of 3.5%. Compared with traditional and mainstream deep learning methods, this system shows significant advantages in detection accuracy, tracking stability and environmental adaptability.