Effective Perception System Design Through Accurate Performance Prediction and Resource Optimization

点击翻译按钮获取中文摘要

In an autonomous vehicle, data from radar, video and acoustic sensors are used for collision avoidance and safe navigation. In a border security system, radar and video sensors are used to improve the safety and security of citizens. In both autonomous and surveillance systems, a critical feature is perception: the ability to automatically absorb sensor data, understand the scene therein, identify the salient aspects in the scene, alert human operators of any impending dangers, and take measures to ensure the safety of people and property. Since such safety-critical autonomous vehicles and surveillance systems demand high precision and real-time performance, it is essential to use the best possible sensors, subject to cost or physical constraints, and to get the best performance out of the available sensor resources. While current autonomous and surveillance systems have come a long way, the perception modules therein are still susceptible to failures, mistakes and accidents, as evidenced by aircraft disappearances, vehicle accidents, border breaches and mining disasters. The long-term objective of the proposed research program is to develop end-to-end algorithms to improve and perfect the perception module in autonomous vehicles and surveillance systems so as to improve the quality of life and safety of citizens who increasingly depend on such systems. The program will ensure that the perception module provides the best possible information to the human user (or the autonomous control system) who will then use that information to obtain the best possible outcome in a critical scenario, such as an impending accident. Optimal perception requires the optimal selection, coordination and use of sensors. However, resource management in real-time, while the autonomous system is operational, is challenging due to imperfect sensors, computational complexity and conflicting priorities. Resource optimization at design-time, where design choices are often made through costly trial-and-error processes, is inefficient. Sub-optimal decisions, without accurate perception performance prediction during pre-production design or real-time operation, can increase the cost of autonomous systems, adversely affect their efficiency and jeopardize the safety of users. Thus, the short-term goals of the research proposed here are to 1) derive accurate performance prediction techniques under realistic operational conditions that can be calculated at design-time or in real-time to improve the design and operation of safety-critical perception systems; and 2) develop algorithms for optimal design-time and real-time perception resource usage based on accurate performance prediction. The proposed work, through collaboration with Canadian automotive industry and government agencies, will reduce the cost of perception modules, make them more accurate and efficient, and make the overall system safer, thus improving the safety, security, and quality of life of Canadians.