Low-light VRU detection using channel-attention zero-DCE++ with Raspberry Pi deployment

• Channel-attention is added to Zero-DCE++ to enhance illumination channels and stability. • Zero-DCE++ iterations are reduced to four with reweighted losses to improve PSNR and SSIM. • A low-light dataset from Kollam captures Indian nighttime roads and VRU classes. • Modified DCE++ is benchmarked on BDD100K and Indian datasets with ablation studies. • Quantized DCE++ is deployed on Raspberry Pi 5, achieving real-time edge performance. Vulnerable Road User (VRU) detection under low-light and varying traffic conditions remains a significant challenge, resulting in object localization inaccuracies. This challenge is further amplified in Indian road characterized by unstructured road layouts and predominantly low illumination where large portions of the scene remain poorly illuminated despite the presence of localized artificial lighting from sources from streetlights and vehicle headlights. To address these limitations, our work introduces an integrated low-light image enhancement and object detection system that couples a modified Zero Reference-Deep Curve Estimation++(Zero-DCE++) network for low light enhancement with YOLOv8n for object detection. Compared to the baseline Zero-DCE++ (n = 8) + YOLOv8n pipeline, the modified Zero-DCE++ introduces integration of channel attention mechanism after each of the first six convolutional layers of Zero-DCE++ enabling the network to selectively focus on dark and poorly illuminated regions. In addition, the illumination smoothness loss and Exposure control loss were modified along with reduction in the curve enhancement iterations from 8 to 4. These design changes produces a cleaner and more spatially coherent illumination map, mitigating over-enhancement artifacts and improving structural preservation in low-light scenes which in turn benefits downstream detection performance. To support a realistic evaluation, we introduce a custom Indian low-light dataset collected from night-time low-light scenes in sub-urban areas of Kollam city, India. On this dataset, the proposed method improves the enhancement quality with PSNR = 19.2 dB outperforming the baseline PSNR = 18.2. In terms of downstream detection, it achieves mAP@50 = 63.47 ± 1.28 with statistically significant improvement over the baseline mAP@50 of 58.9 (paired t -test, p < 1e − 4). For edge deployment, the proposed architecture is optimized using INT8 dynamic quantization and evaluated on a Raspberry Pi 5 achieving 8.35 FPS (119.7 ms per frame) with a compact model size of 3.2 MB demonstrating practical feasibility for real-time intelligent transportation applications. Additional validation was performed on public BDD100K dataset which shows consistent improvement of PSNR = 20.4 dB over baseline 18.4 dB.Comprehensive benchmarking and ablation studies including channel attention , loss modifications, and enhancement iterations were conducted on both the custom Indian dataset and BDD100K to validate the effectiveness of each design choice.