Real-time and Continuous Floor Fingerprint Identification

Abstract

This paper proposes a method to identify the floor fingerprint captured by a camera on a smartphone with limited computation resources. Images captured by the camera on the smartphone are fine enough to recognize microscopic features on the floor because recent smartphones equip high-resolution cameras, but the images are deformed because a user holds a phone by his/her hand. This paper introduces an algorithm to calculate a homography matrix from a gravity vector to rectify the poorly deformed floor image. Limited computation resource is another issue to identify images in real time. We implement cost-sensitive parts in native language C++, and select proper parameters, such as the number of extracted features, to identify floor fingerprint in a reasonable time. In order to determine the accurate user location, it is necessary to identify the angles and the translation against the reference image and identify the position from which the user take a picture image. We implement all these functions in a prototype system. Experimental results show that an image can be identified in a 0.28 seconds, and the accuracy of the position is within some centimeter errors. The prototype system can continuously track human positions, and can generate a map around the reference image.