Vision-based Lane Detection System using CNN and Sliding Window Polynomial Fitting for Autonomous Vehicle

Authors:

Abstract
Enhancing curve detection and accuracy while addressing the challenges posed by environmental conditions is the aim of the proposed vision-based lane detection system. The proposed system employs a three-stage framework consisting of pre-processing, lane marking detection and tracking, and Adaptive Region of Interest (AROI) setup, leveraging machine vision techniques to process greyscale and HSV spaces for analyzing white and yellow lane lines, respectively. Real-time lane detection is achieved through a combination of sliding window polynomial fitting, inverse perspective transformation, and Canny edge detection, ensuring accurate recognition of lane boundaries and curves even under fluctuating illumination. Experimental results demonstrate the system's robustness and reliability across diverse operational conditions, with OpenCV-based testing confirming its ability to maintain high accuracy in lane identification despite varying lighting environments. Additionally, the system is capable of dynamically calculating a vehicle's departure distance based on lane line identification, reinforcing its potential applications in autonomous driving and intelligent traffic systems. By efficiently processing lane lines under complex scenarios, the proposed method establishes a strong foundation for real-world implementation, ensuring stability and precision in automated lane tracking.

This publication has 14 references indexed in Scilit: