Improving Lane and Obstacle Detection Using Stereo Vision-Based Image Processing for Driver Assistance

Abstract

The lane and obstacle detection are the critical components of the ADAS because they directly influence the dependability and safety of the vehicle. This work presents a new system for detecting pedestrians based on stereo vision which aims to solve the problems that have remained unsolved up to now, namely real time processing, low texture and occlusion. By incorporating a high disparity calculation, adaptive depth thresholding, and curvature-aware lane detection, the system provides a reliable and flexible driving scenario in various conditions. The methodology focuses on the computational complexity to allow real-time implementation on embedded systems, while not reducing the detection performance. Comprehensive experiments based on the KITTI dataset show that the proposed system has higher detection accuracy than the previous methods, where the F1-score of straight lane detection is 98.2%, and the F1-score of curved lane detection is 94.6%. The obstacle detection module of the system also showed its detection efficiency of 97.5% in highways proving that the system can handle complicated road surfaces. Also, the system can perform data processing at a rate higher than 30 frames per second, which is sufficient for real-time ADAS applications. These results point to the system’s ability to greatly improve driver safety, which demonstrates its value for implementation in semi-autonomous and autonomous vehicles. As the work fills the major shortcomings of the current approaches, it creates the foundation for developing safer and more efficient means of transport in the future.