An End-to-End Robotic Visual Localization Algorithm Based on Deep Learning

Efficient localization plays a significant role in mobile autonomous robots’ navigation systems. Traditional visual simultaneous localization systems based on point feature matching suffer from two shortcomings. First one is that the method of tracking features is not robust for the environments with frequent changes in brightness. Another one is the large of consecutive visual keyframes consume expensive computation and storage resources in complex environments. To solve these problems, we propose an end-to-end visual localization algorithm to solve the robust and efficiency challenges via a deep learning mode. Firstly, we perform preprocessing operations such as cropping, averaging, and timestamp alignment on datasets to reduce computational cost and time. Secondly, we use CNN networks to autonomously learn the correct features for localization, which is robust to illumination changes. Finally, we utilize LSTM networks to memorize global trajectories to improve the localization precision. We performed a broad range of experiments on both indoor and outdoor datasets. The experimental results demonstrate that the translation and orientation accuracy in outdoor scenes improved by 32.9% and 31.4%, respectively. The average improvement of translation positioning accuracy in indoor scenes is 38.4%, and the orientation improvement is 13.1%. Moreover, the effectiveness of predicting the global motion trajectories of sequential images algorithm has been verified and is superior to other CNN methods.