Robust Place Recognition in Extensively Changing Environment for Robot Navigation

Abstract

Visual place recognition using hand crafted and deep features performs well in static environments. The dynamic environments with extensive changes which are very common are however di cult to be recognised. The envi- ronments may vary in appearance due to many reasons: weather changes, seasonal changes and changes in lightning conditions Visual place recognition can be incredibly enhanced if it becomes possible to estimate the appearance of a speci c scene at a speci c time in view of the appearance of the scene earlier and learning the way in which appearance vary over time. In this the- sis, we examined whether worldwide appearance changes in an environment can be learned adequately to enhance place recognition. We used day night pairs for training a learned model using cGANs that e ciently approximates a night scene based on a day scene. We have used binary descriptor based on color histograms for image matching. The experiments have been done on three datasets collected from di erent environments. The experimental re- sults show that the visual place recognition with images approximated by the trained model outperforms the visual place recognition based on raw images and currently available state of the art methods.