End-to-End Neural Network for Autonomous Steering using LiDAR Point Cloud Data

Although numerous studies on end-to-end autonomous driving systems based on deep learning have been conducted, many of them used shallow feedforward neural networks, which are unsuitable for extracting useful information from complicated contexts and are mainly focused on video frames. This study investigates a LiDAR point cloud-based end-to-end autonomous steering problem in structured roads. The control command to the vehicle is focused on the steering angle of the wheel, which is discretized into continuous integers as the direction category. The problem is then converted into a classification task, which is a mapping connection between the original point cloud data and the driving direction category. On the basis of the PointNet++ framework, we propose using K-means, KNN, and weighted sampling, to perform the steering decision making. Using the CARLA simulation environment, we have shown that the proposed approach is performing effective autonomous decision making with a rate strictly higher than 91% while requiring less inference speed compared to benchmarks.