Multi-Sensors Data Fusion for Navigation of Self-Driving Vehicles

点击翻译按钮获取中文摘要

During the last decade, self-driving vehicles (SVD) have received comprehensive attention in both military and civilian applications due to the significant developments and progress demonstrated by automobile manufacturers (e.g., Ford, Audi, and Mercedes) and technology companies (e.g., Google and Uber). The successful operation of autonomous vehicles relies on a combination of multi-sensors to determine their precise location and sense the environment around (e.g. LiDAR, Cameras, RADAR, Global Navigation Satellite Systems (GNSS), inertial sensors, and odometers), sophisticated algorithms to collate and interpret the acquired multi-sensor data, and powerful processors to execute the implemented algorithms and plan a safe path forward in real time. In addition to cost, trust, reliability, security, and ethical issues, there are significant technical hurdles on the path to widespread adoption of SDV. For example, the employed positioning technology in these vehicles is not accurate enough to be solely trusted and used in safety-critical situations. Moreover, existing digital mapping systems cannot provide highly-detailed maps of surrounding environments that support self-driving applications. Therefore, the development of a multi-sensor system calibration, data fusion, and heterogeneous data processing workflow which is capable of delivering high-accuracy positioning information, high-resolution mapping of the surrounding environment, and reliable path planning is still missing. In other words, the challenges introduced by having a poorly-calibrated multi-sensor system board, the integration and data fusion of the involved GNSS, inertial, and vision-based sensor, precise localization of the vehicle in the absence of GNSS information, efficient processing of large volume multi-sensor data for detailed mapping of the surrounding environment, intelligent information extraction for static/dynamic obstacle detection, and real-time/near real-time decision making and path planning have not been fully understood and addressed by the automotive industry and technology companies pursuing their own self-driving car ambitions. In order to overcome these challenges and prepare self-driving vehicles to hit the roads safely, this proposal aims at developing a comprehensive framework for multi-sensor system calibration, accurate multi-sensor data fusion, and efficient data processing, road/obstacle-related information extraction, highly-detailed mapping of surrounding environments, decision-making, and path planning while addressing the demands of autonomous self-driving vehicles. The proposed framework will provide significant economic, technological, and social benefits to Canadian government and automotive industry. Hence, it will increase the safety and reliability of SVD and promotes their wide adoption in uncontrolled environments.