Efficient location selection for computations of expensive Log-Gabor features using directional enhancement: For robust localization of lane markings in cluttered scenes

Abstract

Vision-based estimation tasks, such as lane marking localization, can be more robust to noise and false signals when utilizing pattern recognition and machine learning techniques as opposed to only low level computer vision operations. Computationally expensive features like Gabor filter responses can be very robust to changes to illumination and other noise. However, machine learning techniques can also be prohibitively slow for time critical applications if such computationally expensive features are calculated for all pixel locations in an input scene. We describe a method to pick the most likely locations for which to compute robust features in order to identify locations of lane markings in highly cluttered scenes. Locations for which features are computed are selected using a novel iterative directional enhancement and thresholding on the perspective image. This drastically reduces the number of locations for which expensive features have to be computed, thus improving latency while retaining precision of the machine learning method. Our method is thus a cascaded classifier scheme that uses low level computer vision operations followed by pattern recognition techniques. We evaluate the performance of our system by checking the overlap of estimates of left and right lane boundaries and lane midline with corresponding annotations.