Paper Title
Lighting-invariant Object Recognition for Robot Operation in Indoor Environments: A Deep Learning Approach

Abstract
Object recognition is a crucial part in a vision based real-world robotics application. This paper leverages transfer learning based on Convolution Neural Networks (CNNs) for object recognition in dynamic ligthing condition in indoor room. As the input, RGB and Depth (RGB-D) image dataset is built and utilized to improve object recognition in challenging lighting conditions. Therefore, a RGB-D integration network is introduced based on combination of features produced as bottleneck from each RGB and Depth network. First, an average accuracy for both rgb and depth are presented to show the impact to each of network’s performance in different lighting scenes. Subsequently, trained integration network indicated more stable recognition in the different lighting condition compared to single RGB based network. Finally, integration network is successfully applied into real-world mobile robot to search and navigate variance of objects in arbitrarily selected order as input from user. Index Terms - Convolution Neural Network, RGB-D datasets, Transfer Learning.