Image rejection and match verification to improve surface-based localization

Abstract

<p>The capability to localize is paramount for many mobile robots and autonomous vehicles. Key attributes of localization systems include reliability, accuracy, low-latency, minimal cost and robustness to variation in environmental conditions. Typical approaches by self-driving car systems incorporate some combination of LiDAR, camera, radar and GPS sensing technologies; which are all suboptimal with respect to one or more of the aforementioned key attributes. This paper presents new research that translates previous work on surface-based positioning systems, which have appealing latency and accuracy properties, to road networks in the context of autonomous car positioning. To achieve the required performance and robustness to appearance change caused by phenomena such as day-night cycles, we develop two new data-driven statistical and learning-based techniques. One performs self-evaluation with regards to the suitability of the current camera query image, while the other retrospectively determines the quality of the query-reference image comparison outcome. Multiple new road surface datasets spanning day and night cycles were utilized to evaluate the system. Results show that both contributions significantly improve place recognition performance, decreasing the median estimated image position error from 24m to 0:26m. The strengths and limitations of our approach and how it could complement current positioning techniques to improve vehicle positioning capability are discussed.</p>