Advanced Lane Detection and Tracking Using Robust Thresholding and Frame Coherence

Despite the development of several vision-based methods over past decades, lane detection is still a challenging issue in computer vision community. Traditional research on lane detection is mainly based on inter-frame similarity. Conversely, our methodology challenges this assumption, allowing the algorithm to adapt in diverse dynamic driving scenarios. In this paper, we proposed an advanced lane detection and tracking method using robust thresholding and frame coherence. A novel two stage feature extraction methodology for lane detection is introduced. In the initial stage, adaptive thresholding is employed based on the local statistical and geometrical constraints of lane markings at the sub-image level. Subsequently, artifacts are addressed and removed using a vertical aggregation technique combined with the Hough transform at the regional level. In this lane detection framework, tracking is integrated to establish a dynamic area of interest for lane search in the subsequent frame. For robust tracking, Kalman filter parameters are adjusted based on measurement reliability. To adapt to the non-linear variations inherent in real-world driving scenarios, outcomes of inter-frame similarity analysis are compared with periodic full-frame detection and a moving average window estimator is employed within the verification protocol. The proposed method is evaluated using Caltech dataset and our dataset, which was recorded at Indian highway and urban roads. Experiment results shows that the proposed method is robust to varying illumination conditions, dense shadows, lane change scenarios, abrupt change in lane width, and occluded lanes with a whooping average detection rate of 98.37%.