Vision-based Distance Measurement and Localization for Automated Driving

Abstract

Automated driving vehicles or advanced driver assistance systems (ADAS) have continued to be an important research topic in transportation area. They can promise to reduce road accidents and eliminate traffic congestions. Automated driving vehicles are composed of two parts. On-board sensors are used to observe the environments and then, the captured sensor data are processed to interpret the environments and to make appropriate driving decisions. Some sensors have already been widely used in existing driver-assistance systems, e.g., camera systems are used in lane-keeping systems to recognize lanes on roads

radars (Radio Detection And Ranging) are used in adaptive cruise systems to measure the distance to a vehicle ahead such that a safe distance can be guaranteed

LIDAR (Light Detection And Ranging) sensors are used in the autonomous emergency braking system to detect other vehicles or pedestrians in the vehicle path to avoid collision

accelerometers are used to measure vehicle speed changes, which are especially useful for air-bags

wheel encoder sensors are used to measure wheel rotations in a vehicle anti-lock brake system and GPS sensors are embedded on vehicles to provide the global positions of the vehicle for path navigation. In this dissertation, we cover three important application for automated driving vehicles by using camera sensors in vehicular environments. Firstly, precise and robust distance measurement is one of the most important requirements for driving assistance systems and automated driving systems. We propose a new method for providing accurate distance measurements through a frequency-domain analysis based on a stereo camera by exploiting key information obtained from the analysis of captured images. Secondly, precise and robust localization is another important requirement for safe automated driving. We propose a method for robust localization in diverse driving situations that measures the vehicle positions using a camera with respect to a given map for vision based navigation. The proposed method includes technology for removing dynamic objects and preserving features in vehicular environments using a background model accumulated from previous frames and we improve image quality using illuminant invariance characteristics of the log-chromaticity. We also propose a vehicle localization method using structure tensor and mutual information theory. Finally, we propose a novel algorithm for estimating the drivable collision-free space for autonomous navigation of on-road vehicles. In contrast to previous approaches that use stereo cameras or LIDAR, we solve this problem using a sensor fusion of cameras and LIDAR.