Deep learning-based dynamic object classification using LiDAR point cloud augmented by layer-based accumulation for intelligent vehicles

摘要

An intelligent vehicle must identify the exact position and class of the surrounding object in various situations to consider the interaction with them. For this reason, the light detection and range sensor, called LiDAR, is widely used in intelligent vehicles. The LiDAR provides information in the form of a point cloud that can be used to localize and classify the surrounding objects. However, unlike vision-based object detection and classification system, the LiDAR-based recognition system cannot provide sufficient classification performance even with deep learning technologies. The reason is that the LiDAR point cloud does not have enough shape information to classify the dynamic object due to the sparsity of the points. To address this problem, we proposed a framework to enhance the deep learning-based classification performance by augmenting the shape information of the LiDAR point cloud. The augmented shape information not only improves classification performance of the networks, but also allows deep learning networks to train effectively by using artificial data-set which is generated with 3D computer-aided design model without tedious efforts of labeling. In order to enhance this shape information effectively, also, this paper proposes a layer-based accumulation algorithm considering the three degree-of-freedom motion of a dynamic object. In the experimental results, the proposed accumulation method outperformed existing registration-based methods. In real-vehicle data test, moreover, the deep learning networks trained with artificial data showed better performance when the LiDAR point cloud was accumulated.