Multi-lane line detection algorithm based on feature point instance segmentation

Abstract
Accurate lane detection is a critical requirement for autonomous driving. However, it remains a challenging task in complex road scenarios, where lane markings can be occluded, worn, curved, or affected by varying illumination. Additionally, the number of visible lane lines is often inconsistent, which limits the effectiveness of conventional lane detection methods. To address these challenges, this paper introduces FPISNet, a novel multi-lane detection algorithm based on feature point instance segmentation. The proposed framework comprises two main components: a feature extraction network and a key feature point prediction network. In the feature extraction stage, an Instance-Batch Normalization (IBN) module is integrated. This module suppresses appearance-related variations while preserving semantic information, enhancing the model’s robustness to changes in brightness, color, and shadows. Furthermore, a Selective Kernel Network (SKNet) is embedded after each convolutional layer. SKNet adaptively adjusts the receptive field and strengthens lane-related features, improving the model’s sensitivity to critical lane cues. Using the extracted feature maps, a stacked hourglass network predicts lane key points and performs lane instance separation. This is achieved through three branches: confidence, offset, and feature embedding, enabling the network to naturally handle a varying number of lane lines in complex scenes. Experimental results on the TuSimple dataset demonstrate that FPISNet achieves an accuracy of 96.92%, surpassing several state-of-the-art lane detection methods. Ablation studies further confirm the contributions of the IBN and SKNet modules in enhancing detection accuracy and robustness. Overall, these results highlight the feasibility and practical potential of FPISNet for robust, multi-lane detection in autonomous driving systems.