Monocular Localization Using Invariant Image Feature Matching to Assist Navigation

Indoor positioning is critical for applications like navigation, tracking, monitoring, and accessibility. For the visually impaired this has a huge implication on independent mobility for accessing all types of services as well as social inclusion. The unavailability of indoor positioning solutions with adequate accuracy is a major constraint. The key reason for the lack of growth in indoor positioning systems is to do with the reliability of indoor positioning techniques and additional infrastructure costs along with maintenance overheads. We propose a novel single camera-based visual positioning solution for indoor spaces. Our method uses smart visual feature selection and matching in real-time using a monocular camera. We record and transform the video route information into spars and invariant point-based SURF features. To limit the real-time feature search and match data, the routes inside the buildings are broken into a connected graph. To find the position, confidence of a path increases if it founds a good feature match and decreases otherwise. Each query frame uses a K-nearest neighbor match with the existing databases to increase the confidence of matched path in subsequent frames. Results have shown a reliable positioning accuracy of \(\sim \) 2 meters in variable lighting conditions. We also investigated the error recovery of positioning systems where it easily re-positions the user within the neighboring edges. To promote crowdsourcing, proposed system can add more visual features to the database while performing the matching task.