Object Classification using 3D Convolutional Neural Networks

Abstract

In the development of autonomous driving and active safety systems, knowledge about the vehicle’s surroundings is critical. When it comes to making decisions in real driving scenarios, the location and relative movement of surrounding vehicles, pedestrians and even static objects gives invaluable information to the system responsible for decision making. To know whether an object is a car or pedestrian, the system has to distinguish between the different features to predict object type. LiDAR sensors are among the most commonly used sensors in the development of modern AD systems as they produce dense images of their surroundings that are relatively resistant to changing light and weather conditions. Many classification methods use feature extraction or transformations to evaluate the 3D information using methods commonly used in 2D image analysis. In this thesis we evaluate the performance of training convolutional neural networks directly on 3D data, bypassing any information loss through data extraction or transformation and allowing the intensity hit of points to be used. The effectiveness of the method is evaluated on a dataset created from the KITTI Vision Benchmarking Suite. Our results show a total accuracy score of 96.35% and with a mean accuracy of 95.67% on a dataset trained on 7 classes.