DNN-Based Recognition of Pole-Like Objects in LiDAR Point Clouds

Abstract

In this paper, we present a novel method for recognizing pole-like objects in LiDAR point clouds, which is useful for landmark-based localization and high-definition (HD) map generation. Our method utilizes a state-of-the-art deep neural network relying on learned encodings of the point cloud input. Here, we modified an existing network architecture to improve the detection of small objects such as poles. To enable the estimation of bounding cylinders for pole-like objects, we propose a respective object anchor design with an accompanying strategy for matching ground truth objects to object anchors during network training. Furthermore, we examine the impact of two different data representations of the point cloud on the detection performance, as well as the impact of topological alternatives. The performance of our method is demonstrated on a dataset including various challenging classes of poles. We plan to publish this dataset as part of this work, which fills a gap regarding publicly available LiDAR point cloud datasets covering various elements of HD maps such as pole-like objects. Our method achieves a mean recall, precision, and classification accuracy of 0.85, 0.85, and 0.93, respectively, and may serve as a future baseline for other approaches.