Reliable perception of highly changing environments: implementations for car-to-pedestrian collision avoidance systems

摘要

Robots have been given sophisticated ”eyes” to make them ”see” and understand their environments. These eyes (cameras, ladars, sonars, radars, etc...) collect a huge amount of data that need to be correctly processed to be useful. Processing this information is what a perception system is intended to perform. For almost half a century now, various perception algorithms have been proposed to tackle one or several of the underlying issues that arise when addressing the perception problem. Well known tracking, detection, mapping, localization and classification algorithms can consequently be combined to design complete perception algorithms that work well for a given application in most situations. The problem is that some real world applications (autonomous driving, etc...) require perception systems that do better than working well in most situations. An autonomous vehicle driving in a crowded urban center would need indeed to be equipped with a perception system that works well in every situation. This dissertation addresses the specific problem of perception systems reliability when confronted to highly changing dense environment. First a detailed analysis of the fundamental limitations undermining the performances of existing approaches is given. Then an original approach - based on a unified grid-based formulation of the five perceptual subproblems - is proposed and proves to be capable of solving issues that most existing systems cannot solve. The relevance of this analysis and the experimental validity of the proposed approach is assessed through an experimental comparison of two fully detailed original perception systems specifically designed for pedestrian detection purposes in urban environments.