Abstract
This paper employs the SlowFast deep neural network to classify near-miss incident videos, significantly enhancing accuracy through emulation of the human visual system's separation into slow (P-cells) and fast (M-cells) processing two streams. By delaying the analysis time of near-miss incident, more obvious features of near-misses are exposed, making it easier for the model to learn near-miss features. This methodology demonstrates the potential for early driver notification to prevent accidents, thereby encouraging progress in traffic safety.
