Comparing Deep Learning Models for Road Asset Detection and Classification in LiDAR Point Cloud

摘要

Self-driving cars (or autonomous cars) navigate through an environment without any driver intervention with the help of a complex sensor network. This networks comprises mainly vision sensors working in tandem with accurate algorithms to detect movable and non-movable objects around them. This sensor network typically includes cameras to identify static and non-static objects, Radio Detection and Ranging (RADAR) to detect the speed of the moving objects and Light Detection and Ranging (LiDAR) to detect the distance to objects. In this paper, we explore the use of LIDAR data to classify static objects on the road, For that matter we compare different deep learning models for static road asset detection using only LIDAR point cloud. The different deep learning models are split between 2D models to classify 3D data and 3D models to classify 3D data. For the 2D model, we use an already established pipeline to transform the data from a 3D point cloud to a group of 2D points that is compatible with the input layer of a 2D CNN. Results show an accuracy exceeding 90% in the detection and classification of road edges, solid and broken lane markings, bike lanes, and lane center lines for the 2D model while the accuracy of the 3D models was considerably less.