Object Detection and Classification in 3D Point Cloud Data for Automated Driving

Abstract

A 3D LIDAR provides 3D surface information of objects with the highest position accuracy, among available sensors that can be utilized to develop perception algorithms for automated driving vehicles. In terms of automated driving, the accurate surface information gives the following benefits: 1) the accurate position information that is quite useful itself for collision avoidance is stably provided regardless of illumination condition, because the LIDAR is an active sensor. 2) the surface information can provide precise 3D shape-oriented features for object classification. Motivated by these characteristics, we propose three algorithms for a perception purpose of automated driving vehicles based on the 3D LIDAR in this dissertation.

A very first procedure to utilize the 3D LIDAR as a perception sensor is segmentation that transform a stream of the LIDAR measurements into multiple point groups, where each point group indicate an individual object near the sensor. In chapter 2, a real-time and accurate segmentation is proposed. In particular, Gaussian Process regression is used to solve a problem called over-segmentation that increases False Positives by partitioning an object into multiple portions.

The segmentation result can be utilized as input of another perception algorithm, such as object classification that is required for designing more human-likely driving strategies. For example, it is important to recognize pedestrians in urban driving environments because avoiding collisions with pedestrians are nearly a top priority. In chapter 3, we propose a pedestrian recognition algorithm based on a Deep Neural Network architecture that learns appearance variation.

Another traffic participant that should be recognized with high-priority is a vehicle. Because various vehicle types of which appearances differ, such as a sedan, a bus, or a truck, are present on road, detection of the vehicles with similar performance regardless of the types is necessary. In chapter 4, we propose an algorithm that makes use of a common appearance of vehicles to solve the problem. To improve performance, a monocular camera is additionally employed, where the information from both sensors are integrated by a Dempster-Shafer Theory framework.