FlickerFix: Flicker Removal for LED Traffic Lights in Dashcam Videos

초록

Many LED traffic lights control brightness with pulse-width modulation (PWM). When the frame rate of a dashcam does notmatch the PWM frequency or even when there is a small mismatch, flicker artifacts appear, making the signal look as if it blinksor disappears. Instead of relying on hardware-integrated solutions, we propose a three-stage post-processing method for flickerremoval. First, we fine-tune a YOLO detector on domestic traffic light datasets. Second, a segmentation model trained on oursynthesized signal-pseudo mask dataset is used to extract the active light region. Finally, corrupted signals are restored throughreference-mask based synthesis. We also constructed two datasets: a large-scale ROI segmentation dataset with pseudo masks, anda synthetic flicker video dataset generated by simulating PWM-camera mismatch. We evaluate the method on 85 pairs of syntheticflicker videos, achieving improvements of +3.6dB in PSNR, +0.30 in SSIM, +0.19 in FDI. We also validate on real dashcamvideos, showing consistent qualitative improvements. Because the approach is sensor-agnostic, it can be applied to existing dashcamfootage and can support more reliable traffic-light recognition in future ADAS and autonomous driving systems.

키워드