Towards accurate kidnap resolution through deep learning

Abstract

This paper presents a six degree of freedom position regression CNN (convolutional neural network) based on Google's Inception-VLI CNN. This network is then evaluated quantitatively and compared to previous state-of-the-art position regression CNNs. Our model achieves a 22% and 51% relative improvement compared to previous state-of-the-art methods for position and orientation accuracy respectively. A modular system for integrating our model into probabilistic localization algorithms for accurate kidnap resolution and global metric initialization in real-time is also introduced and evaluated. This modular system is able to globally initialize 85% of the time in under 70ms. If the robot is allowed to rotate in place and capture multiple views, this rises to 95%.