Abstract
Traditional lane detection systems relying on RGB cameras often fail under extreme lighting conditions due to low dynamic range and motion blur. This paper introduces a novel approach using Dynamic Vision Sensors (DVS) and a specialized Structure-Aware Network (SANet). We present the first high-resolution (1280x800) event-based lane detection dataset (DET) and propose a Multidirectional Slice Convolution (MSC) module to capture the diagonal geometric properties of lane markers. Our method outperforms state-of-the-art models, achieving a Mean F1 score of 74.21% while maintaining robustness in challenging environments.
Introduction: The Challenge of Traditional Imaging
Lane detection is a fundamental task for autonomous driving, yet standard frame-based RGB cameras struggle with high-contrast scenes and rapid motion. In scenarios such as entering or exiting tunnels, or driving at night with intense glare, traditional sensors often produce overexposed or underexposed frames, causing lane markers to disappear from the visual field.
Dynamic Vision Sensors (DVS) offer a compelling alternative. Unlike traditional cameras that capture full frames at fixed intervals, DVS pixels independently respond to changes in log-intensity, providing a massive dynamic range (130dB) and microsecond-level latency. However, the sparse and asynchronous nature of DVS data presents unique challenges for standard computer vision architectures.
Event-Based Data Representation
To leverage existing deep learning frameworks, the raw asynchronous event stream must be processed into a format compatible with convolutional neural networks.
The authors utilize "accumulated images," where events within a specific time window are integrated into a 2D frame. This representation preserves the structural edges of moving objects—such as lane markers—while naturally filtering out static backgrounds like the sky or uniform road surfaces, significantly reducing computational noise.
The Structure-Aware Network (SANet)
The proposed SANet architecture is designed to exploit the inherent geometric structure of lane lines. It consists of a feature extraction backbone, a novel Multidirectional Slice Convolution (MSC) module, and an upsampling layer for semantic segmentation.
Multidirectional Slice Convolution (MSC)
Standard CNNs often have limited receptive fields, making it difficult to maintain connectivity in broken or occluded lane markers. While previous architectures like SCNN used slice convolutions to pass messages horizontally and vertically, SANet introduces diagonal and counter-diagonal message passing.
Since lane markers in a vehicle's perspective are typically oblique, these additional diagonal paths allow the network to capture structural context more effectively. This mechanism enables the model to "infer" missing segments of a lane by reasoning about the global geometric flow, ensuring continuous extraction even under partial occlusion.
The DET Dataset
A significant contribution of this work is the DET (lanE marker exTraction) dataset, the highest resolution DVS dataset for lane detection to date.
- Resolution: 1280×800 pixels.
- Scale: 5,424 DVS images derived from 5 hours of raw event streams recorded in diverse urban environments in Wuhan, China.
- Annotations: Multi-class semantic segmentation (background plus four lane classes) using a distance-based labeling convention to ensure feature consistency.
Results and Insights
SANet was evaluated against several state-of-the-art semantic segmentation and lane detection models. It achieved a Mean F1 score of 74.21% and a Mean IoU of 60.86%, notably outperforming SCNN (73.85% F1) and general-purpose models like DeepLabv3 (61.26% F1).
Ablation studies confirmed the importance of multidirectional message passing. Incorporating diagonal directions proved superior to horizontal/vertical-only approaches, highlighting that structural context is more critical than visual texture in event-based lane extraction.
Conclusion and Future Directions
While SANet demonstrates the power of DVS for robust lane detection, the current reliance on accumulated frames does not fully exploit the asynchronous nature of event data. Future research will focus on Spiking Neural Networks (SNNs) and other native event-processing architectures to further reduce latency and improve power efficiency for real-time edge deployment in autonomous vehicles.
Summary updated