DIONet: An Illumination-Adaptive Network for Real-Time Road Obstacles Detection Under Low-Light Scenarios

Road obstacles detection under low-light conditions faces challenges such as feature loss, blurred boundaries of small targets, insufficient feature representation, and bounding box regression biases. To address these issues, we propose dark is ok network (DIONet), an end-to-end real-time algorithm designed for low-light road obstacle detection. First, we propose an obstacle illumination transformer (OIT) module, which employs a multiscale transformer architecture to decouple illumination and reflection components in low-light images, effectively recovering detailed obstacle features. Second, we design a frequency-adaptive dilated convolution (FADC) module to enhance the network’s perception of multiscale obstacles, mitigating boundary blur in low-light scenes. Third, we propose an attention-based intrascale feature interaction (AIFI) module, which optimizes feature fusion through an attention mechanism to enhance the localization of small-sized obstacles. Finally, we propose an Inner-IoU approach to refine localization strategies, enhancing detection precision for small obstacles and alleviating boundary-induced biases. Experimental evaluations demonstrate that DIONet achieves state-of-the-art performance, attaining mean average precision (mAP)@50 of 93.5 and mAP@50:95 of 65.9 on the seek and secure (SAS) dataset, outperforming the YOLOv11 by 2.8% and 2.7%, YOLOv12 by 2.0% and 3.7%, LED-YOLO by 1.5% and 2.0%, EnlightenGAN-YOLO by 1.5% and 2.1%, PE-YOLO by 1.7% in both metrics, while leading in precision and recall. Moreover, deployment experiments on the PIXKIT2.0 autonomous driving hardware platform confirm the real-time applicability and effectiveness of our proposed approach.