LAGS-YOLO: An End-to-End Low-Light Vehicle Detection Model with Luminance-Adaptive Glare Suppression

Vehicle detection under low-light conditions is a critical component of visual perception for Level 3+ autonomous vehicles. Under low-light conditions, existing detection methods are unable to simultaneously enhance illumination in dark areas and suppress overexposure from strong light sources, resulting in reduced detection accuracy. To address this, we have designed LAGS-YOLO, an end-to-end model for vehicle detection under low-light conditions, which incorporates three improvements. The LA-GS-DCE++ adaptive illumination and glare suppression enhancement module has been proposed; by introducing a brightness-gated bidirectional gain mechanism without imposing constraints on learning parameters, it is able to brighten dark areas of the image while simultaneously suppressing overexposure from light sources. Concurrently, we have improved the model’s detection network by introducing a P2 high-resolution detection head and BiFPN to preserve spatial details, whilst optimising bounding box regression using PIoU loss. We designed an image-level serial end-to-end joint optimisation framework, which feeds the detection loss gradients back into the enhancement network to mitigate object misalignment, and employs a two-stage weight scheduling mechanism to maintain training stability. Experiments on the BDD100K night-time subset show that the proposed enhancement module has the lowest ratio of overexposed pixels (1.38%) among all comparative enhancement methods. After end-to-end joint training, LAGS-YOLO achieves [email protected] and [email protected]:0.95 of 75.80% and 44.49%, respectively, with a parameter count of only 2.94 M, which is an improvement of 5.45 and 1.97 percentage points over the YOLOv8n baseline under low-light input. The enhancement module employs bidirectional gain control, and the enhancement network and detection network are interconnected via gradient feedback, enabling a symmetrical design between enhancement and detection. This provides technical support for visual perception in Level 3 and higher autonomous driving under complex lighting conditions.