Lane detection based on improved feature map and efficient region of interest extraction

In this paper, the authors propose an efficient algorithm to detect lanes on an improved feature map based on Inverse Perspective Mapping (IPM). IPM removes the perspective effect from the road image and creates a bird's eye view of the road scene. A conventional approach to estimate the lateral offset of the lanes from the IPM image is by adding feature points column by column. However, when the vehicle is drifting from the lanes, lanes are not vertical lines anymore. In this paper, first an edge orientation histogram is calculated in the IPM image and a global lane orientation is estimated. Then, the Signal to Noise Ratio (SNR) of the feature map is improved based on the initially estimated global lane orientation by shifting and matching the feature map. Since lanes are already oriented in that direction, while lane feature points match with each other, other feature points do not. Thus, improvement in the SNR is achieved. Then, the lane marking positions are estimated by 1-D Hough transform searching lines in polar representation with a known orientation, and finally lanes are detected by fitting parabolic lane models to the feature points, in the region of interests, initialized by the Hough transform. The proposed algorithm, achieved to detect lanes with above 96% of detection ratio, tested in complex scenarios.