DiffuFlicker: Diffusion-Based LED Traffic Light Flicker Removal in Dashcam Videos

초록

The flicker phenomenon caused by the mismatch between pulse-width modulation (PWM)-driven LED traffic lights and camera frame rates degrades the reliability of traffic-light perception in autonomous driving systems. To the best of our knowledge, this is the first deep learning-based approach specifically designed for LED traffic light flicker removal. We present DiffuFlicker, a diffusion-based restoration method that addresses this underexplored problem. DiffuFlicker employs a class-conditioned temporal 3D U-Net together with an SDEdit-based adaptive restoration strategy, enabling robust recovery under both high-frequency (HFF) and low-frequency (LFF) flicker conditions. Experimental results on 144 synthetic video sequences demonstrate that DiffuFlicker achieves an average peak signal-to-noise ratio (PSNR) improvement of +6.35 dB, outperforming the rule-based baseline (+5.27 dB) by 1.08 dB. Video results are available at https://diffuflicker.github.io/. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2027.

키워드