A Study on a Lane Keeping System using CNN for Online Learning of Steering Control from Real Time Images

Using CNN for lane keeping in autonomous driving control system has been proposed. Learning of the CNN is performed using road images and steering operations as teaching data. When driving, steering operations are generated from camera images by CNN. Even on roads where white lines are missing or unpaved roads without white lines, the learned CNN can perform autonomous driving control. However, learning of the existing CNN method is “batch learning” which is batch processing, and the prediction phase for driving uses the CNN fixed beforehand. It is difficult to control the driving operations on a road which has different features from the roads used for learning. In this paper, to realize “learning while driving,” we propose using “online learning” in the driving control system, which means learning is performed while receiving input data in real-time. It makes possible to learn additionally for unlearned roads when driving. For evaluation experiments, we implement an online learning system of automatic driving control using CNN on a small robot car to investigate the feasibility. We introduce a divided buffer which classifies and inputs the input data to reduce the influence of the bias of the training data which is the weak point of online learning. From the result of the experiments, there is a possibility that “learning while driving” can be realized by using online learning to a CNN autonomous driving control system.