Recent lane detection methods leverage deep learning techniques, achieving significantly better performance than conventional systems based on traditional hand-crafted algorithms and basic image processing. However, despite their effectiveness, these data-driven approaches are often complex, computationally intensive, and struggle to operate efficiently in real-time environments. The developed tool by CR2C2 researchers introduces a real-time lane detection method that utilizes a lightweight semantic segmentation model for lane feature extraction, combined with a novel clustering technique to delineate lane boundaries. The segmentation network is iteratively refined to reduce the number of convolutional layers without sacrificing detection accuracy. This optimization notably decreases memory usage and accelerates inference, allowing the detector to meet the real-time operational requirements of automated vehicles. The segmentation outputs are post-processed through clustering and then fitted with cubic splines to produce smooth and accurate lane boundary representations. The proposed method achieves detection performance comparable to state-of-the-art (SOTA) approaches, based on ego-lane evaluation using the TuSimple dataset. Furthermore, it has been deployed on a real automated vehicle platform, running at over 92 FPS, demonstrating its capability to meet real-time performance demands in practical self-driving scenarios.

Link to the tool: https://github.com/ACCESSLab/Lane-Detection-using-Segmentation

Related Publications

• Tesfamichael Getahun, and  Ali Karimoddini, “Realtime Semantic Segmentation-based Lane Detection for Automated Driving," Under review.