Computationally efficient navigation system for Unmanned Ground Vehicles

This paper proposes to enhance the existing methods of Self-Supervised Learning (SSL) with application to autonomous navigation systems through efficient computational approaches that are the principal requirements in a practical system. First, confidence-based auto labeling for self-supervised learning is introduced which identifies and eliminates the input samples with low confidence level that are susceptible to be mislabeled. Then, a biologically inspired saliency detection approach for feature biasing is presented which is able to detect the salient features through top-down task specific guidance. The proposed methods are general and can be applied to a variety of applications. Finally, experimental results on real datasets from the DARPA-LAGR program are given to illustrate the effectiveness of the proposed approaches.