Motorcycle Blind Spot Detection Through Computer Vision Techniques

The rise in vehicles on our roads has led to a noticeable increase in traffic accidents, especially those involving motorcycles. Motorcyclists face many problems, one significant issue being unsafe lane changes. In this research, a real-time blind spot detection system is being proposed. The proposed system is aimed at accurately identifying vehicles in the motorcycle's blind spot, thereby enhancing road safety without requiring the rider to divert their attention from the path ahead. This research addresses the limitations of the pre-trained YOLO model, such as vehicle detection during the night, by implementing a custom dataset. This adaptation enhances both precision and recall without compromising real-time inference speeds. The experimental process entailed capturing video from a helmet-mounted camera, segmenting the footage into train, validation, and test splits, annotating the video frames, and finally evaluating both pre-trained and custom-trained models against the test dataset to gather necessary metrics. Despite the custom dataset's relatively smaller size compared to the comprehensive COCO dataset, the custom model demonstrated marked improvement in night-time vehicle detection compared to the pre-trained model, with respective precisions of 82.6% and 22.1%. Additionally, the recall stands out as a significant metric, with the custom model providing a high value of 80.3%, a substantial improvement over 0.4% by the pre-trained model.