Abstract

The ActiVis project aims to deliver a mobile system that is able to guide a person with visual impairments towards a target object or area in an unknown indoor environment. For this, it uses new developments in object detection, mobile computing, action generation and human-computer interfacing to interpret the user's surroundings and present effective guidance directions. Our approach to direction generation uses a Partially Observable Markov Decision Process (POMDP) to track the system's state and output the optimal location to be investigated. This system includes an object detector and an audio-based guidance interface to provide a complete active search pipeline. The ActiVis system was evaluated in a set of experiments showing better performance than a simpler unguided case.