Research on Autonomous Driving Performance Enhancement in Low-Light Scenarios Based on Improved Object Detection Algorithm

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

In low-light scenarios, autonomous driving perception systems face problems such as weak contrast, high noise, blurred boundaries, and missing texture details, which pose great challenges to object detection tasks. Existing methods often use image enhancement as a preprocessing step to improve detection performance; however, such methods usually require additional computational overhead and struggle to simultaneously meet the requirements of detection accuracy and real-time performance. To address these issues, this paper takes YOLOv8n as the baseline model and proposes an improved object detection algorithm, BiF-VoV-YOLO, for low-light autonomous driving scenarios. First, a lightweight feature fusion module, VoVGSCSP, is introduced into the backbone network to replace the original SPPF module. Through grouped feature reuse and channel shuffling, it reduces the number of model parameters and computational cost while enhancing low-light feature representation. Second, a BiFormer fusion layer is embedded in the neck, using a content-adaptive sparse attention mechanism to filter low-light background noise and improve global dependency modeling and detection accuracy. Finally, the upsampling node nn.Upsample is reconfigured and added in the neck. To address the lack of content-awareness in original upsampling, which leads to the loss of small-object details, this design improves information fidelity during the upsampling stage. The synergistic effect of the three modules significantly alleviates missed detection and false detection of multi-scale objects in low-light scenarios on the DarkFace dataset while maintaining the model's real-time inference capability. Experimental results show that the improved model achieves an mAP@0.5 of 53.9%, an improvement of 8.9 percentage points over the YOLOv8n baseline, and a reduction of about 0.05MB in parameter count. It achieves the best detection accuracy among various classic models and state-of-the-art models, verifying the effectiveness and generalization of the proposed algorithm in low-light autonomous driving object detection.