Online covariance estimation for novelty‐based visual obstacle detection. Issue 8 (29th May 2017)

• Record Type: Journal Article
• Authors: Ross, Patrick; English, Andrew; Ball, David
• Is Part Of: Journal of field robotics. Volume 34:Issue 8(2017)
• Page Start: 1469
• Page End: 1488
• Publication Date: 2017-05-29
• Subjects: learning -- perception; Robots, Industrial -- Periodicals; Automatic control -- Periodicals; 629.892
• DOI: 10.1002/rob.21724
• Languages: English
• ISSNs: 1556-4959

Abstract

Abstract: Robust obstacle detection remains a challenge for mobile robots traversing outdoor field environments. Obstacle detection systems that combine multiple cues can potentially overcome deficiencies in individual cues. A key challenge in designing multi‐sensor obstacle detection systems is to automatically and appropriately combine these cues in an unsupervised manner. This paper presents a method for obstacle detection, which continuously adapts its obstacle definition and the weighting of each cue for the current conditions. The key contribution of this paper is a method for online covariance estimation for Parzen windows probability density estimation, which in this application determines the relative importance of each descriptor dimension. By iteratively estimating the covariance using small subsets of the available data, the proposed method is capable of converging to an approximate solution order of magnitude faster than standard optimizers, making it suitable for use online. The application of this covariance estimation method to our novelty‐based obstacle detection system improves obstacle detection precision and reduces learning duration after environmental transitions. It also removes all environment‐specific parameters from the method, and allows the descriptor to contain arbitrary data without time‐consuming hand‐tuning.