Efficient road lane marking detection

In recent years, lane detection has garnered significant attention. Mainstream lane detection algorithms can be categorized into two major classes: segmentation-based lane detection algorithm and anchor-based lane detection algorithm. Building upon these two categories, this dissertation proposes se...

Full description

Saved in:
Bibliographic Details
Main Author: Zhu, Ziyuan
Other Authors: Xie Lihua
Format: Thesis-Master by Coursework
Language:English
Published: Nanyang Technological University 2024
Subjects:
Online Access:https://hdl.handle.net/10356/173226
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:In recent years, lane detection has garnered significant attention. Mainstream lane detection algorithms can be categorized into two major classes: segmentation-based lane detection algorithm and anchor-based lane detection algorithm. Building upon these two categories, this dissertation proposes several improvement measures to enhance both the detection speed and accuracy of the models to a certain extent. For the improvement on segmentation-based algorithm, the backbone of the model is replaced by Swin Transformer which possesses an attention mechanism. Attention mechanism enhances the feature extraction capability of the backbone and improves the accuracy of lane detection. However, it cannot meet the real-time detection requirement due to the large number of parameters of the model. For the improvement on anchor-based algorithm, two aspects are optimized for this algorithm: (1) enhancing the feature extraction capability of the backbone by incorporating convolutional block attention module (CBAM) and auxiliary branch; (2) designing a structural loss based on the morphological characteristics of lanes to optimize the loss function. Compared to the optimized segmentation-based algorithm in this dissertation, anchor-based algorithm achieves fast detection speed and can meet real-time detection requirements. The experimental results in this dissertation are obtained from the CULane dataset, which provides a rich variety of scenes that closely represents real-world road conditions.