FlickerFix: Flicker Removal for LED Traffic Lights in Dashcam Videos

I. Introduction

Modern LED traffic lights use PWM, which can cause flicker in dashcam videos when camera frame rates do not match (Figure 1). Prior studies mostly focused on hardware solution [1], while software-based flicker removal was mainly frame-level [2] rather than ROI-specific. As shown in Figure 2, our pipeline directly addresses flickers at the ROI level using a software-only approach. Our pipeline is related to recent work on frame-level video deflickering [2]. While the recent work mitigates the flicker across the entire frame, our proposed FlickerFix targets flicker restoration within LED traffic-light ROIs. Our contribution goes beyond proposing a removal pipeline, as we also constructed dedicated datasets to enable evaluations using established metrics such as PSNR, SSIM and FDI [3].

Fig. 1.

Examples of flickers captured with a high-resolution camera (iPhone 16 Pro, 25 fps)

Fig. 2.

The proposed flicker removal pipeline

Ⅱ. Proposed Method

2. Synthetic Datasets Generation

For testing and training, we build two datasets. One is a segmentation dataset of traffic lights. The other is a flicker dataset where we simulate PWM–camera mismatch with different settings.

2.1 A Large-scale ROI Segmentation Dataset with Pseudo Masks

We first find traffic light areas with object detection. The bounding boxes are then cleaned into pixel-level masks by simple color threshold and morphological filtering. This produces a large paired dataset of 536,028 cropped ROI images in JPEG with corresponding multi-attribute labels(JSON) and pseudo segmentation masks(PNG), as illustrated in Figure 3.

Fig. 3.

Results of the signal synthetic segmentation dataset showing traffic light ROIs and their pseudo masks

2.2 Flicker Synthetic Video Dataset

To test robustness against flicker, we generate 85 real dashcam videos with synthesized flicker using our proposed Algorithm 1, which modulates the brightness of the traffic-light region under varying signal and camera parameters. The flicker synthesis was performed with PWM frequencies of 120 Hz, 150 Hz, and 180 Hz, a camera frame rate of 27.5 fps, exposure times of 2ms and 5ms, and duty cycles of 0.3, 0.5, 0.7. Each sequence includes a paired original and flickered version. Figure 4 shows the results of the flicker synthesis using Algorithm 1.

Fig. 4.

Results of the flicker synthesis

Ⅲ. Experimental Result

1. Qualitative Results

We test our pipeline on real dashcam videos with Figure 5 showing the process and results of our method, and Figure 6 presenting the comparison of temporal brightness consistency.

Fig. 5.

Qualitative results of the flicker removal on real dashcam videos

Fig. 6.

Comparisons of brightness difference between original, with synthesized flicker, and FlickerFix (Ours)

2. Quantitative Results

In this work, we check the restoration performance with three metrics (PSNR, SSIM, and FDI). FDI (Flicker Detection Index) from the IEEE P2020 [3] measures whether a flickering light remains visible against the background, computed from ROI brightness contrast with a threshold τ = 0.10. The FDI is computed as:

$\[ \mathrm{FDI=Prob}\left(\frac{{ X}_{\mathrm{meas}}-X_{\mathrm{ref, off}} }{X_{\mathrm{ref, off }}}\ge \tau \right) \]$

(1)

We focused on the detailed mask-level in ROI evaluation results mainly for the 27.5fps case. In Table 1, our method increased 3.62 of PSNR, 0.3028 of SSIM and 0.1919 of FDI.

Table 1.

PSNR/SSIM/FDI Comparison of 85 pairs of ROI’s of Original/Flicker/FlickerFix videos

Ⅳ. Conclusion

We presented FlickerFix, a software-based pipeline that detects and removes flicker artifacts in dashcam videos of LED traffic lights. Beyond the removal method itself, we contributed two dedicated datasets that enable reproducible training and evaluating. Experiments on 85 synthetic sequences and real dashcam footage demonstrate clear improvements in both perceptual quality and visibility stability. Unlike hardware-oriented solutions, our approach is fully sensor-agnostic and can be applied to existing recordings. Future work will address challenges in detecting very small or distant signals and extend the synthetic dataset generation to support more comprehensive and fair evaluation protocols.

Acknowledgments

This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2022-00155915, Artificial Intelligence Convergence Innovation Human Resources Development (Inha University)). This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) under the Leading Generative AI Human Resources Development (IITP-2025-RS-2024-00360227) grant funded by the Korea government(MSIT).

References

• LED Flicker Mitigation Technology for Automotive Use, https://www.sony-semicon.com/en/technology/automotive/lfm.html
• Jiang, Y., Xue, T., Brown, M., and Huang, J., “Blind video deflickering by neural filtering with a flawed atlas,” Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2023. [https://doi.org/10.1109/CVPR52729.2023.01006]
• Deegan, B., “A review of IEEE P2020 flicker metrics,” in Proc. IS&T Int’l. Symp. On Electronic Imaging: Autonomous Vehicles and Machines, vol. 34, pp 108-1-108-6, January 2022. [https://doi.org/10.2352/EI.2022.34.16.AVM-108]
• Tian, Y., Ye, Q., Doermann, D., “YOLOv12: Attention-centric real-time object detectors,” arXiv preprint arXiv:2502.12524, February, 2025. [https://doi.org/10.48550/arXiv.2502.12524]
• AI Dataset for Traffic Light and Traffic Sign Detection in Seoul Capital Area, https://aihub.or.kr/aihubdata/data/view.do?currMenu=115&topMenu=100&dataSetSn=188, (accessed Sep. 2021).
• Traffic Light Signal Information Recognition Image Data in Various Driving Environments of Autonomous Vehicles, https://aihub.or.kr/aihubdata/data/view.do?currMenu=115&topMenu=100&dataSetSn=71579, (accessed Dec. 2023).