Low-Cost Domain Adaptive Experience Based Localization for Autonomous Robots

Localization is one of the key aspect of any automated navigation system, it means finding own position in a given environment, without it an agent might get lost. In a dynamic environment localization is a very difficult task because various factors such as weather and illumination may change over time. We developed a low-cost solution to the localization problem, relying only on a camera sensor, without using expensive sensors like lidars. The solution allows to handle dynamic nature of the environment and uses previous knowledge for localization, hence it is named as Experience Based Localization. STREET-CNN is a core part of the method which is trained on STREET dataset. This dataset contains several images of streets in IIIT-A along with their GPS coordinates as labels. The trained network calculates the similarity score between the two cross domain images. Presented method is equipped with Kalman Filter to include speed variations of agent while predicting its location. Results produced with domain adaptation technique is compared with, without the domain adaptation technique.