New Vision Model Boosts Low-Light Traffic Perception, Enabling 12.37% Accuracy Gains

Perceiving objects accurately and swiftly in low-light traffic presents a significant challenge for current computer vision systems, often hampered by poor illumination and limited visual information. To address this, Hulin Li, Qiliang Ren from Chongqing Jiaotong University, Jun Li, Hanbing Wei, and Zheng Liu from the University of British Columbia, along with Linfang Fan from Chongqing Jiaotong University, introduce a new biologically inspired vision model and a comprehensive dataset designed to improve performance in these difficult conditions. The team constructed Dark-traffic, the largest dataset of its kind for low-light traffic scenes, and developed the Separable Learning Vision Model (SLVM), which mimics aspects of biological vision to enhance perception under adverse lighting. Results demonstrate that SLVM achieves substantial improvements over existing state-of-the-art methods in object detection, instance segmentation, and flow estimation, offering a promising step towards safer and more reliable autonomous driving in challenging conditions.

A key component, the Light-Adaptive Pupil Module (LAPM), simulates the human eye’s ability to adjust to varying light levels by amplifying image channels and focusing on luminance, efficiently extracting texture features with minimal computational cost. Experiments involved training SLVM on the Dark-traffic and LIS datasets, demonstrating effective object perception in low light. The success of SLVM and LAPM highlights a promising approach to low-light vision, balancing performance with computational efficiency.

Large-Scale Dataset for Low-Light Traffic Perception

Recognizing the need for better data, scientists created Dark-traffic, a large and detailed dataset specifically designed for training and evaluating object perception systems in low-light traffic scenes, supporting tasks such as object detection, instance segmentation, and optical flow estimation. Building upon this foundation, the team engineered SLVM, a biologically inspired framework designed to enhance perception under adverse lighting. SLVM incorporates a light-adaptive pupillary perception mechanism (LAPM), mimicking the dynamic response of the human pupil to varying light levels, and a feature-level separable learning strategy (FSLConv) that efficiently decouples texture features. A spatial misalignment-aware fusion module ensures precise alignment of multi-feature information. Extensive experiments demonstrate that SLVM achieves state-of-the-art performance with reduced computational overhead, outperforming existing methods on the Dark-traffic and LIS benchmarks. To address the lack of suitable data, the team constructed Dark-traffic, the largest densely annotated dataset to date specifically designed for low-light traffic scenes, supporting detailed analysis of object detection, instance segmentation, and motion estimation. The core of this advancement lies in SLVM’s biologically inspired design, which mimics aspects of the human visual system to enhance performance in adverse lighting. Experiments reveal that SLVM outperforms state-of-the-art models, achieving an 11.

2 percentage point improvement in detection accuracy compared to RT-DETR, and exceeding YOLOv12 by 6. 1 percentage points in instance segmentation. SLVM also reduces endpoint error by 12. 37% on the Dark-traffic dataset and surpasses existing methods on the LIS benchmark, achieving an average of 11 percentage points higher across key metrics when compared to Swin Transformer+EnlightenGAN and ConvNeXt-T+EnlightenGAN. These results demonstrate a substantial leap forward in low-light perception, paving the way for safer and more reliable autonomous systems and advanced driver-assistance technologies. The Dark-traffic dataset and associated code are publicly available, encouraging further research and innovation.

Dark-Traffic Dataset and SLVM Model Advance Perception

This research addresses the challenge of accurate object perception in low-light traffic environments, where existing methods struggle due to poor illumination and limited visual cues. To facilitate progress, the team introduced Dark-traffic, a large and detailed dataset specifically designed for low-light traffic scene understanding, supporting tasks such as object detection, instance segmentation, and flow estimation. SLVM incorporates a light-adaptive mechanism, efficient feature learning, task-specific processing, and precise feature alignment, demonstrating state-of-the-art performance across multiple benchmarks.

Experiments show SLVM outperforms existing real-time models in both accuracy and efficiency, exceeding previous methods on the LIS benchmark and achieving significant gains in detection and instance segmentation on the Dark-traffic dataset. The authors acknowledge potential trade-offs in practical applications, noting that the feature decomposition strategy may introduce latency and that careful scheduling is required to fully utilise the modular architecture on parallel hardware. Future work will likely focus on addressing these limitations and further optimising the system for resource-constrained platforms, with all code and data made publicly available to encourage further research.