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

Song, Shiping; Zou, Bowei; Yan, Yan; Lv, Hengxu; Zhang, Cheng; Zhang, Zhiqiang · 2026 · DOAJ

DOI: 10.1186/s44147-026-01005-7

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of accurate multi-lane line detection in autonomous driving systems, particularly in complex road scenarios where lane markings are occluded, worn, curved, or affected by varying illumination. Existing methods often struggle with inconsistent numbers of visible lanes and sensitivity to appearance changes. To resolve these issues, the authors propose FPISNet, a novel algorithm based on feature point instance segmentation. The framework is designed to handle variable lane counts and maintain robustness against environmental variations by integrating appearance-invariant feature learning and adaptive receptive field adjustment. The FPISNet architecture consists of two primary components: a feature extraction network and a key feature point prediction network. The feature extraction network employs three convolutional blocks, each incorporating an Instance-Batch Normalization (IBN) module and a Selective Kernel Network (SKNet). The IBN module splits feature map channels to apply Instance Normalization and Batch Normalization separately, filtering out appearance-related variations (such as brightness and color) while preserving semantic information. SKNet adaptively adjusts the receptive field to strengthen lane-related features. The prediction network utilizes a stacked hourglass structure with four cascaded hourglass modules to predict lane key points. It outputs three branches: confidence scores for key point existence, offset values for precise positioning, and feature embeddings for instance separation. The model is trained end-to-end using a total loss function combining confidence, offset, and feature vector losses. Experiments were conducted on the TuSimple dataset, which includes 3,626 training and 2,782 test samples collected under diverse traffic conditions. The model was trained on an NVIDIA 2080 GPU using PyTorch, with images cropped to 512 × 256 pixels. FPISNet achieved an accuracy of 96.92%, surpassing several state-of-the-art lane detection methods. Ablation studies confirmed that the IBN and SKNet modules significantly contribute to detection accuracy and robustness. The algorithm effectively handles scenarios with varying numbers of lane lines and performs well under challenging conditions such as overexposure and occlusion. The study demonstrates that FPISNet is a feasible and practical solution for robust multi-lane detection in autonomous driving. By combining instance segmentation with appearance-invariant feature extraction and adaptive attention mechanisms, the method overcomes limitations of traditional segmentation and anchor-based approaches. The results highlight the potential for deploying such algorithms in real-world intelligent driving systems, where reliability under diverse and unpredictable road conditions is critical.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• distraction detection algorithms