Optimizing lightweight CNN-based collision prediction models with optical flow to enhance temporal continuity perception and robustness under low-light conditions.