Video-based Lane Detection using Boosting Principles

Autonomous navigation of road vehicles is a challenging problem that has widespread applications in intelligent systems, and robotics. An integral component of such a system is to understand how the road is structured. Detection of road lane markings assumes importance in this regard, and this problem has been approached with different visual inputbased inference algorithms ([1], [2]), besides other sensing modalities such as the GPS sensor and the internal vehiclestate sensors. But the challenge still remains when there is considerable amount of shadows on the road, variations in outdoor lighting conditions of the scene (transition from day to night), among others. To address such issues, we propose a machine learning approach based on Real Adaboost [3], and train linear classifiers for both the appearance and edge cues of the training examplars. Additionally, we incorporate prior knowledge about the relative importance of the training samples by computing their weights using kernel discriminant analysis [4], before learning the classification function through boosting. The regions identified as lane markings are then analyzed for gradient direction consistency before making the final detection decision. We illustrate the effectiveness of our algorithm on challenging daylight and nighttime road scenarios.