Lane marker detection and rain removal for autonomous vehicle navigation

Abstract

With the increasing need for autonomous vehicles, the driving safety of vehicles in severe weather conditions is imperative and needs to be addressed. Lane marker detection provides crucial position related information for the vehicles towards autonomous navigation. Also, lane markers detection based on machine vision is playing a crucial role in autonomous vehicle technologies nowadays. Amongst all the lane detection sensors like radar, laser, camera, etc., the camera will be the most economical and practical component to be used for testing autonomous vehicles.

Although camera-based lane marker detection methods are widely used, they are sensitive to noise, such as rain streaks, which would degrade the performance of many machine vision algorithms or may even lead to failure. Therefore, preprocessing mechanisms like rain removal is key to perform lane marker detection, which in turn improves the lane detection accuracy.

In this thesis, a progressive method for lane detection on city roads has been developed. By combining the sliding windows and Kalman filter approaches into a model-based method, we obtained a better performance, when compared to the other existing techniques. Also, a modified neural network structure, combining CNN and LSTM is designed to remove rain streaks before performing the lane marker detection and tracking. Compared to the existing methods in the literature, an average improvement of 2.3% in the peak signal to noise ratio (PSNR) value and an 8% improvement in the Google vision test results has been recorded.