Development of a Highly Precise Place Recognition Module for Effective Human-robot Interactions in Changing Lighting and Viewpoint Conditions

Abstract

We present a highly precise and robust module for indoor place recognition, extending the work by Lemaignan et al. and Robert Jr. by giving the robot the ability to recognize its environment context. We developed a full end-to-end convolutional neural network architecture, using a pre-trained deep convolutional neural network and the explicit inductive bias transfer learning strategy. Experimental results based on the York University and Rzeszów University dataset show excellent performance values (over 94.75 and 97.95 percent accuracy) and a high level of robustness over changes in camera viewpoint and lighting conditions, outperforming current benchmarks. Furthermore, our architecture is 82.46 percent smaller than the current benchmark, making our module suitable for embedding into mobile robots and easily adoptable to other datasets without the need for heavy adjustments.