Probabilistic lane segmentation using a low-dimensional linear parametrization

Lane detection is an important module for active safety systems since it increases safety and reduces traffic accidents caused by driver inattention. Illumination changes or occlusions make lane detection a challenging task, especially if the detection is performed from a single image. Consequently, this paper presents a probabilistic approach based on the Kalman filter, which uses information from previous image frames to estimate the lane that could not be detected in the current image frame, considering uncertainty in the prediction as well as in the detection. To this end, a principal component analysis of the segmented curvature is introduced with the purpose of dimensionality reduction, moving from a large dimensional pixel representation to a considerably reduced space representation. Furthermore, the proposed approach is compared with a fully connected pretrained CNN model for lane detection, demonstrating that the proposed method has a lower computational cost in addition to a smoother transition between lane estimates.