Abstract:Ensuring the robustness of lane detection systems is essential for the reliability of autonomous vehicles, particularly in the face of diverse weather conditions. While numerous algorithms have been proposed, addressing challenges posed by varying weather remains an ongoing issue. Geometric-based lane detection methods, rooted in the inherent properties of road geometry, provide enhanced generalizability. However, these methods often require manual parameter tuning to accommodate it fluctuating illumination and weather conditions. Conversely, learning-based approaches, trained on pre-labeled datasets, excel in localizing intricate and curved lane configurations but grapple with the absence of diverse weather datasets. This paper introduces a promising hybrid approach that merges the strengths of both methodologies. A novel adaptive preprocessing method is proposed in this work. Utilizing a fuzzy inference system (FIS), the algorithm dynamically adjusts parameters in geometric-based image processing functions and enhances adaptability to diverse weather conditions. Notably, this preprocessing algorithm is designed to seamlessly integrate with all learning-based lane detection models. When implemented in conjunction with CNN-based models, the hybrid approach demonstrates commendable generalizability across weather conditions and adaptability to complex lane configurations. Rigorous testing on datasets featuring challenging weather conditions showcases the proposed method's significant improvements over existing models, underscoring its efficacy in addressing the persistent challenges associated with lane detection in adverse weather scenarios.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: How to improve the robustness and generalization ability of the lane detection system based on Convolutional Neural Network (CNN) under changeable weather conditions. Specifically, the existing lane detection algorithms perform poorly in the face of different weather conditions, especially in bad weather conditions such as rainy days, which limits the safety and reliability of self - driving vehicles. ### Specific Background of the Problem 1. **Limitations of Geometric Methods**: - Geometric methods rely on road geometric characteristics, but require manual adjustment of parameters to adapt to different lighting and weather conditions. - These methods can usually only handle simple straight lanes and perform poorly on complex and curved lanes. 2. **Limitations of Learning Methods**: - Learning methods are trained through pre - labeled data sets and can well locate complex lane configurations, but it is difficult to deal with bad weather conditions in the absence of diverse weather data sets. - Most of the existing data sets (such as TuSimple, Caltech, KITTI) are collected in sunny weather and lack data in rainy and snowy days. 3. **The Need for Hybrid Methods**: - Combine the advantages of geometric methods and learning methods to develop a lane detection algorithm that can maintain high performance under changeable weather conditions. ### Solutions Proposed in the Paper The paper proposes a hybrid method based on adaptive pre - processing parameter adjustment, using Fuzzy Inference System (FIS) to dynamically adjust the parameters in geometric image processing functions to enhance the adaptability to different weather conditions. Specific steps include: - **Noise Suppression**: Use bilateral filters to reduce image noise. - **Edge Detection**: Adopt the Canny edge detection algorithm and dynamically adjust the threshold through a fuzzy parameter regulator. - **Region of Interest (ROI) Selection**: Select a triangular ROI to simplify calculations and improve accuracy. - **Line Detection**: Use the Hough transform to evaluate the effect of the current threshold and adjust the number of candidate pixels. - **Fuzzy Parameter Regulator**: Dynamically adjust the threshold of edge detection through the Mamdani Fuzzy Inference System to ensure the stability and accuracy of the detection results. ### Experimental Verification This algorithm is combined with three different CNN lane detection models (CLRNet, RESA, Enet - SAD) and tested on Tusimple and DSDLDE data sets. The results show that this pre - processing algorithm significantly improves the detection performance of these models in rainy conditions. Although it slightly decreases (<1%) on sunny data sets, it still shows good compatibility and improvement effects. ### Summary The main contribution of the paper is to propose an innovative adaptive pre - processing method, which can effectively improve the robustness and generalization ability of lane detection algorithms under changeable weather conditions, providing strong support for the safe driving of self - driving vehicles.

Improved Generalizability of CNN Based Lane Detection in Challenging Weather Using Adaptive Preprocessing Parameter Tuning

Enhanced SCNN-Based Hybrid Spatial-Temporal Lane Detection Model for Intelligent Transportation Systems

U-Net-Based Learning Using Enhanced Lane Detection with Directional Lane Attention Maps for Various Driving Environments

Enhanced Cross Layer Refinement Network for robust lane detection across diverse lighting and road conditions

LVLane: Deep Learning for Lane Detection and Classification in Challenging Conditions

A Fast and Robust Lane Detection via Online Re-Parameterization and Hybrid Attention

Foggy Lane Dataset Synthesized from Monocular Images for Lane Detection Algorithms

An Efficient Lane Detection Network with Channel-Enhanced Coordinate Attention

Bridging the Gap of Lane Detection Performance Between Different Datasets: Unified Viewpoint Transformation

Effective lane detection on complex roads with convolutional attention mechanism in autonomous vehicles

Lane Mark Detection with Pre-Aligned Spatial-Temporal Attention

DAGL-Faster: Domain adaptive faster r-cnn for vehicle object detection in rainy and foggy weather conditions

Fusion of color images and LiDAR data for lane classification.

Object Detection in Adverse Weather for Autonomous Vehicles Based on Sensor Fusion and Incremental Learning

Spatial-Temproal Based Lane Detection Using Deep Learning

Semi-supervised lane detection for continuous traffic scenes

LHFFNet: A hybrid feature fusion method for lane detection

Fast lane detection by fusing multi-scale contour feature extraction and weight mean normalization

Combining Low-Light Scene Enhancement for Fast and Accurate Lane Detection

Efficient Lane Boundary Detection with Spatial-Temporal Knowledge Filtering.

Fast Multi-Lane Detection Based on CNN Differentiation for ADAS/AD