Adaptive Low-Light Enhancement for Road Anomaly Detection: A Multi-Architecture Benchmark and Edge Deployment Study

arXiv:2611.08393 · cs.CV · Submitted 2026-11-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Adaptive Low-Light Enhancement for Road Anomaly Detection".

Jane: Here's the gist of the paper. Road anomaly detection systems on autonomous vehicles must function reliably under degraded visual conditions including darkness, rain, and Harmattan dust haze.

Tom: This paper benchmarks four object detection architectures, YOLOv8n, YOLOv12n, RT-DETR, and Faster R-CNN, under physically grounded lighting and weather degradation.

Jane: The proposed method introduces a lightweight, training-free luminance-adaptive router that dynamically selects between no enhancement, moderate SCI, and denoising plus strong SCI based on per-frame Rec.

Tom: 601 luminance to recover most of the dusk penalty while preserving low-light gains.

Paper summary: Tom: So looking at the whole "Adaptive Low-Light Enhancement for Road Anomaly Detection: A Multi-Architecture Benchmark and Edge Deployment Study," we see that the core finding is that front-end enhancement isn't always a safe idea; it only helps when the input signal is really inadequate, and it actively degrades performance in certain conditions like dusk.

Jane: The authors introduce this luminance-adaptive router as a way to overcome that issue by dynamically selecting between different levels of processing based on the frame's light level, which they found substantially recovered most of that loss.

Lu: They also confirmed that RT-DETR retained a night level gain of plus zero point seven one mAP at fifty on unseen data, suggesting its global attention mechanism was robust even when things got dark or had occlusion.

Meng: From a deployment standpoint, the study points toward YOLOv8n as being the most viable architecture for real-time use on hardware like the Jetson Nano because it's lightweight, though they still flag that YOLOv12n is marginally better in some areas.

Tom: The implication here for us is that we shouldn't just apply a standard enhancement to everything; we need a system that intelligently decides when to enhance and when to bypass the enhancement entirely based on the actual visual input.

Jane: And what this means for drivers or autonomous systems is that you get better results without sacrificing the speed or simplicity needed for on-the-go use, even in challenging conditions like Harmattan dust haze.

Conclusion: Tom: So we've looked at how this paper benchmarks four different object detection architectures—YOLOv8n, YOLOv12n, RT-DETR, and Faster R-CNN—under bad lighting and weather conditions.

Jane: The title is "Adaptive Low-Light Enhancement for Road Anomaly Detection: A Multi-Architecture Benchmark and Edge Deployment Study." It sounds pretty technical, but basically it’s about finding the best way to make self-driving car vision work when it’s dark or dusty.

Lu: Yeah, the authors are tackling a huge problem in places like Nigeria where things get really hazy with dust at night. They show how just using one simple enhancement method isn't always good; sometimes it actually makes the detection worse.

Meng: I mean, from an engineering standpoint, they test this on hardware like the Jetson Nano. That’s where the real world is—running models fast and small enough to actually use in a car or a drone.

Tom: The conclusion is that this adaptive router thing they propose recovers most of the performance loss caused by bad light without needing any extra training data. It dynamically chooses between no enhancement, moderate enhancement, or heavy denoising plus strong enhancement based on how bright the picture looks right there in the frame.

Jane: And they found that RT-DETR actually kept a night-level gain on unseen data, which is pretty interesting considering how much degradation they were fighting against.

Lu: It shows that simple, rule-based decisions based on light levels are enough to capture most of the effect; you don't need to retrain the whole model just for this.

Meng: But here's my concern, and this is what I worry about—they only tested this on a small validation set. When we move it to real traffic, how robust is that simple router when things get truly chaotic?

Tom: That’s the next step, right? Because if we can make these models work reliably under those harsh conditions without needing a massive data overhaul, that opens up so many doors for deploying better safety features on edge devices.

Obed Honour Eje, Oluwatosin Kolade, Glory Bagai, Taliha Abiodun Folorunso

Federal University of Technology Minna · Obafemi Awolowo University

cs.CV

Submitted: 2026-11-05

Updated: 2026-11-05

Comments: 5 pages

Code: https://github.com/ultralytics/ultralytics

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 76/100

The gist: Road anomaly detection systems on autonomous vehicles must function reliably under degraded visual conditions including darkness, rain, and Harmattan dust haze.

Key concepts

Luminance-Adaptive Router
This is a lightweight decision mechanism that looks at the frame's luminance (Rec. 601) and decides automatically whether to apply no enhancement, moderate enhancement, or heavy denoising. It uses fixed rules based on brightness levels to save computational power by only enhancing when necessary.
SCI Enhancement
SCI stands for Signal-to-Contrast Enhancement, a technique used to improve image visibility in low light. The study found that applying this enhancement uniformly is harmful at dusk because the input signal is already adequate, which actively degrades detection performance.
Edge Deployment Feasibility
This refers to testing how well the detection system runs on resource-constrained hardware like the NVIDIA Jetson Nano. The results show that only lightweight models like YOLOv8n and marginally YOLOv12n can operate in real-time on this specific hardware tier.
Dusk Penalty Recovery
The 'dusk penalty' is a performance drop caused by low light conditions at dusk. The adaptive router successfully recovered most of this loss, meaning the system maintained high accuracy even when the lighting was poor, without needing expensive retraining.

Terminology

Summary

Road anomaly detection systems on autonomous vehicles must function reliably under degraded visual conditions including darkness, rain, and Harmattan dust haze. p. 0 This paper benchmarks four object detection architectures, YOLOv8n, YOLOv12n, RT-DETR, and Faster R-CNN, under physically grounded lighting and weather degradation. p. 0 The proposed method introduces a lightweight, training-free luminance-adaptive router that dynamically selects between no enhancement, moderate SCI, and denoising plus strong SCI based on per-frame Rec. p. 0 601 luminance to recover most of the dusk penalty while preserving low-light gains. p. 0

Pruning redundant reasoning tokens

The primary motivation stems from the severe public-health and economic burden of road traffic crashes in Nigeria. Poor infrastructure, poor lighting, and adverse weather conditions like Harmattan dust haze significantly reduce contrast and add signal-dependent noise, weakening features that detectors rely on. p. 0 A natural mitigation is a front-end enhancement module to restore visibility before detection. p. 0 However, enhancement optimized for human perception can actively harm machine vision because the objectives are misaligned. p. 0 Retraining on augmented data is computationally expensive and impractical for resource-constrained edge hardware like the Jetson Nano. p. 0

How the training signal is built

The evaluation corpus was constructed by merging two publicly available road-anomaly datasets, specifically the Rajdalsaniya Pothole Detection Dataset and the speed-unmarked-bump dataset v3. The merged corpus comprises 4,444 images across two classes (pothole, speed bump), resized to 640 × 640 px and partitioned into training, validation, and test splits. p. 1 All degradation pipeline development and router threshold derivation were conducted on the 667-image validation split. p. 1

Where the gains come from

The core innovation is a luminance-adaptive router that computes a perceptual luminance score to dynamically route the signal, either bypassing enhancement entirely, passing through moderate enhancement, or heavily denoising and maximally enhancing. This dynamic routing ensures computational resources are only expended when genuinely necessary to recover features. p. 1 The router uses fixed, rule-based decision functions based on Rec. p. 3 601 luminance thresholds derived from mean luminance centroids of different degradation levels. p. 2

Edge Deployment Throughput and Latency

The evaluation assessed edge feasibility by deploying all four detectors, the SCI module, and the adaptive router on an NVIDIA Jetson Nano. YOLOv8n was identified as the only architecture approaching real-time operation across all branches. p. 3 The SCI module added a fixed overhead independent of the detector, with latency varying based on the routing branch (e. p. 3g., 16 ms for Branch B). p. 1

Conclusion

The study demonstrates that uniform SCI enhancement is actively harmful at dusk, degrading mAP@50 by 16–31% across architectures. The luminance-adaptive router substantially recovered the plain-SCI dusk penalty, maintaining performance within 0. p. 3007 of baseline on the held-out test split. p. 3 RT-DETR retained a night-level gain of +0. p. 3071 mAP@50 on unseen data. p. 3 Edge validation confirmed that only YOLOv8n, and marginally YOLOv12n, are suitable for real-time deployment at this hardware tier. p. 4 p. 4

Future Work

Immediate extensions include collecting real-world data under Nigerian night-time, Harmattan, and rainy-season conditions to validate synthetic findings. Further work involves compounding degradation benchmarking combining lighting and weather simultaneously, as well as developing weather-specific restoration modules and model compression for sub-Jetson-Nano deployment. p. 4

References

[1] Federal Road Safety Corps (FRSC), “2024 annual road traffic crash report,” press briefing by Corps Marshal S. Mohammed, Abuja, Nigeria, Jan. p. 4 2025.

[2] World Health Organization, Global Status Report on Road Safety 2023. p. 4

[3] World Bank Global Road Safety Facility, “The single organizational model for road safety in Nigeria,” World Bank Group, Washington, DC, USA, 2023. p. 4

[4] G. Jocher, A. Chaurasia, and J. p. 4 Qiu, “Ultralytics YOLOv8,” 2023. p. 4

[5] Y. Safyari, M. p. 4 Mahdianpari, and H. p. 4 Shiri, “A review of vision-based pothole detection methods using computer vision and machine learning,” Sensors, vol. p. 4 24, no. 17, p. 5652, 2024. p. 4

[6] A. C. Anuforom, “Spatial distribution and temporal variability of Harmattan dust haze in sub-Sahel West Africa,” Atmos. p. 4 Environ., vol. 41, no. 39, pp. 9079–9090, 2007. p. 4

[7] C. Guo et al., “Zero-reference deep curve estimation for low-light image enhancement,” in Proc. p. 4 IEEE/CVF Conf. p. 4 p. 4 p. 4 Comput. Vis. Pattern Recognit. p. 4, (CVPR), 2020, pp. p. 4 1780–1789. p. 4

[8] Y. Jiang et al. p. 4, “EnlightenGAN: Deep light enhancement without paired supervision,” IEEE Trans. p. 4 Image Process. p. 4, vol. 30, pp. 2340–2349, 2021. p. 4

[9] L. Ma, T. Ma, R. Liu, X. Fan, and Z. p. 4 Luo, “Toward fast, flexible, and robust low-light image enhancement,” in Proc. p. 4 IEEE/CVF Conf. Comput. Vis., (CVPR), 2022, pp. p. 4 5637–5646. p. 4

[10] H. Liu, F. Jin, H. Zeng, H. Pu, and B. Fan, “Image enhancement guided object detection in visually degraded scenes,” IEEE Trans. p. 4 Neural Netw. p. 4, Learn., Syst., vol. 35, no. 10, pp. 14164–14177, 2023. p. 4

[11] Y. Wang et al. p. 4, “Is underwater image enhancement all object detectors need? p. 4,” IEEE J. p. 0 Ocean. Eng., vol. 49, no. 2, pp. 606–621, 2023. p. 4

[12] Y. Tian, Q. Ye, and D. Doermann, “YOLOv12: Attention-centric realtime object detectors,” in Proc. p. 4 Adv. Neural Inf. p. 4 p. 4 Process., (NeurIPS), vol. p. 4 38, 2026, pp. p. 0 78433–78457. p. 4

[13] Y. Zhao et al. p. 4, “DETRs beat YOLOs on real-time object detection,” in Proc. p. 4 IEEE/CVF Conf. Comput. Vis., (CVPR), 2024, pp. p. 4 16965–16974. p. 4

[14] S. Ren, K. He, R. Girshick, and J. p. 4 Sun, “Faster R-CNN: Towards realtime object detection with region proposal networks,” in Proc. p. 4 Adv. Neural Inf. Process., (NIPS), vol. p. 4 28, 2015.

[15] A. Dhiman and R. p. 4 Klette, “Pothole detection using computer vision and learning,” IEEE Trans. p. 4 Intell. Transp., Syst., vol. 21, no. 8, pp. 3536–3550, 2019. p. 4

[16] D. Arya et al. p. 4, “Crowdsensing-based road damage detection challenge (CRDDC’2022),” in Proc. p. 4 IEEE Int. Conf. Big Data (Big Data), 2022, pp. p. 4 6378–6386. p. 4

[17] D. Arya, H. Maeda, S. K. Ghosh, D. Toshniwal, and Y. p. 4 Sekimoto, “RDD2022: A multi-national image dataset for automatic road damage detection,” Geosci. p. 4 Data J., vol. 11, no. 4, pp. 846–862, 2024. p. 4

[18] Y. Zanevych et al. p. 4, “Evaluation of pothole detection performance using deep learning models under low-light conditions,” Sustainability, vol. p. 4 16, no. 24, p. 10964, 2024. p. 4

[19] J. Terven, D.-M. Cordova-Esparza, and J. p. 4-A. Romero-González, “A comprehensive review of YOLO architectures in computer vision: From YOLOv1 to YOLOv8 and YOLO-NAS,” Mach. p. 4 Learn. Knowl. Extr., vol. 5, no. 4, pp. 1680–1716, 2023. p. 4

[20] K. He, J. Sun, and X. p. 4 Tang, “Single image haze removal using dark channel prior,” IEEE Trans. p. 4 Pattern Anal. p. 4 Mach. Intell., vol. 33, no. 12, pp. 2341–2353, 2010. p. 4

[21] Z. Chen et al. p. 4, “Object detection for autonomous vehicles under adverse weather conditions,” Expert Syst. p. 4 Appl., 2025, Art. p. 4 no. 128994.

[22] S. S. Halder, J.-F. Lalonde, and R. p. 4 de Charette, “Physics-based rendering for improving robustness to rain,” in Proc. p. 4 IEEE/CVF Int. p. 4 Conf. Comput., Vis., (ICCV), 2019, pp. p. 4 10203–10212. p. 4

[23] W. Liu et al., “Image-adaptive YOLO for object detection in adverse weather conditions,” in Proc. p. 4 AAAI Conf. Artif. Intell., vol. 36, no. 2, 2022, pp. p. 0 1792–1800. p. 4

[24] S. Kalwar et al. p. 4, “GDIP: Gated differentiable image processing for object detection in adverse conditions,” arXiv preprint arXiv:2209. p. 414922, 2022. p. 4

[25] Q. Qin, K. Chang, M. Huang, and G. p. 4 Li, “DENet: Detection-driven enhancement network for object detection under adverse weather conditions,” in Proc. p. 4 Asian Conf. p. 4 Comput., Vis., (ACCV), 2022, pp. p. 4 2813–2829. p. 4

[26] R. Gao, M. H. Omar, and M. p. 4 Mahmuddin, “YOLO-DER: A dynamic enhancement routing framework for adverse weather vehicle detection,” Electronics, vol. p. 4 14, no. 24, p. 4851, 2025. p. 4

[27] Z. Wei, Y. Wang, L. Sun, A. V. Vasilakos, and L. p. 4 Wang, “ClassLIE: Structure- and illumination-adaptive classification for low-light image enhancement,” IEEE Trans. p. 4 Artifif., vol. 5, no. 9, pp. 4765–4775, 2024. p. 4

[28] H. Feng, G. Mu, S. Zhong, P. Zhang, and T. p. 4 Yuan, “Benchmark analysis of YOLO performance on edge intelligence devices,” Cryptography, vol. p. 4 6, no. 2, p. 16, 2022.

[29] D. K. Alqahtani, M. p. 4 A. Cheema, and A. p. 4 N. Toosi, “Benchmarking deep learning models for object detection on edge computing devices,” in Proc. p. 4 Int. Conf. Service-Oriented Comput. p. 4, Singapore: Springer, 2024, pp. p. 4 142–150.

[30] Rajdalsaniya, “Pothole detection dataset,” Kaggle, 2022. p. 1

[31] “Speed-unmarked-bumb dataset, v3,” Roboflow Universe, 2023. p. 4

--- Page 6 ---

Fig. 1. The proposed luminance-adaptive routing architecture. p. 1

[24], and DENet/DE-YOLO [25] integrate learned, alwayson enhancement end-to-end with the detector, while YOLODER [26] and ClassLIE [27] make a binary enhance/bypass decision after additional training. p. 1 Benchmarks confirm lightweight YOLO variants are typically the only models approaching real-time on a Jetson Nano, often requiring TensorRT, quantisation, or pruning. p. 1 p. 1

The architecture integrates learned enhancement end-to-end with the detector through DENet/DE-YOLO [25] and YOLODER [26], or makes a binary enhance/bypass decision after additional training via ClassLIE [27]. Benchmarks confirm lightweight YOLO variants are typically the only models approaching real-time on a Jetson Nano, often requiring TensorRT, quantisation, or pruning.

RT-DETR’s global attention mechanism consistently gave it an advantage under partial occlusion (fog, rain). This robustness is plausibly attributable to its global self-attention mechanism, which is less reliant on local high-frequency texture than the convolutional backbones of the YOLO models and Faster R-CNN. p. 3

The luminance-adaptive router demonstrates that a simple, training-free, threshold-based decision is sufficient to capture most of this effect, recovering the majority of the dusk penalty without any additional model training.

The study confirms that front-end enhancement is not a universally safe intervention; SCI improves detection only when the input signal is genuinely inadequate and actively degrades it when applied to scenes like dusk that already carry sufficient luminance. The adaptive router recovers most of this loss, and RT-DETR retains a night-level gain on unseen data. p. 3 Edge validation confirmed that only YOLOv8n, and marginally YOLOv12n, are suitable for real-time deployment at this hardware tier.

Immediate extensions include collecting real-world data under Nigerian night-time, Harmattan, and rainy-season conditions to validate synthetic findings. Future work also involves compounded degradation benchmarking combining lighting and weather simultaneously, as well as developing weather-specific restoration modules and model compression for sub-Jetson-Nano deployment. p. 4

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Adaptive enhancement routing for road anomaly detection systems can be implemented by computing a lightweight perceptual luminance score to dynamically select between no enhancement, moderate SCI, or denoising plus strong SCI based on per-frame Rec. 601 luminance (Equation 2). This ensures that computational resources on constrained edge hardware are only expended when genuinely necessary to recover features.

  2. The proposed system can be deployed on NVIDIA Jetson Nano edge hardware, achieving real-time operation with YOLOv8n and marginally YOLOv12n models up to 18.54 FPS under the adaptive routing branch B (Dusk). This confirms that only YOLOv8n, and marginally YOLOv12n, are suitable for real-time deployment at this hardware tier.

  3. The system can provide improved accuracy under dusk lighting by utilizing the luminance-adaptive router, which substantially recovered the plain-SCI dusk penalty on validation data by reducing residual loss to 0.001–0.073 mAP@50. This is achieved through a fixed, rule-based decision function that involves no learned parameters and required no additional training.

  4. The system can be more robust against uniform low-light enhancement by demonstrating that uniform SCI enhancement is actively harmful at dusk, degrading mAP@50 by 16–31% across architectures. This finding allows the system to avoid applying enhancement indiscriminately, preserving feature integrity when the input signal is already adequate.

  5. The architecture can be optimized for specific detector types by utilizing per-architecture threshold calibration for twostage detectors. This refinement is necessary because the fixed thresholds are not uniformly optimal for two-stage detectors, allowing the system to tailor its routing decisions based on whether it is using Faster R-CNN or a one-stage detector like YOLOv8n.

Sources

Related papers