Real-Time Deterministic Lane Detection on CPU-Only Embedded Systems via Binary Line Segment Filtering

Cost-sensitive advanced driver-assistance systems (ADAS) increasingly rely on embedded platforms without discrete GPUs, where power-intensive deep neural networks are often impractical to deploy and difficult to certify for safety-critical functions. At the same time, classical geometry-based lane detection pipelines still struggle under strong backlighting, low-contrast night scenes, and heavy rain. This work revisits geometry-driven lane detection from a sensor-layer perspective and proposes a Binary Line Segment Filter (BLSF) that exploits the structural regularities of lane markings in bird’s-eye-view (BEV) images. The filter is integrated into a three-stage pipeline consisting of inverse perspective mapping, median local thresholding, line-segment detection, and simplified Hough-based sliding-window fitting with RANSAC. On a self-collected dataset of 297 challenging frames (strong backlighting, low-contrast night, heavy rain, and high curvature), the full pipeline improves lane detection robustness over the same implementation without BLSF while maintaining real-time performance on a 2 GHz ARM CPU-only platform. To assess generality, we further evaluate BLSF on the Dazzling Light and Night subsets of the large-scale CULane and LLAMAS benchmarks, where it achieves a consistent 6–7% improvement in F1-score over a line-segment baseline under a fixed pre-processing configuration, along with corresponding gains in IoU. These results demonstrate that explainable, geometry-driven lane feature extraction can deliver competitive robustness under adverse illumination on low-cost, CPU-only embedded hardware, and can serve as a complementary design point to lightweight deep-learning models in cost- and safety-constrained ADAS deployments.