A visual place recognition procedure with a Markov chain based filter

摘要

Recognizing a place with a visual glance is the first capacity used by humans to understand where they are. Making this capacity available to robots will make it possible to increase the redundancy of the localization systems available in the robots, and improve semantic localization systems. However, to achieve this capacity it is necessary to build a robust visual place recognition procedure that could be used by an indoor robot. This paper presents an approach that from a single image estimates the robot location in the semantic space. This approach extracts from each camera image a global descriptor, which is the input of a Support Vector Machine classifier. In order to improve the classifier accuracy a Markov chain formalism was considered to constraint the probability flow according the place connections. This approach was tested using videos acquired from three robots in three different indoor scenarios - with and without the Markov chain filter. The use of Markov chain filter has shown a significantly improvement of the approach accuracy.