Utvikling av ansiktsgjenkjenningssystem for bruk på NAO robot

Abstract

The main objective of this thesis was to implement a demonstration behaviour for the NAO robot, with focus on face recognition. To achieve this, a complete framework for face recognition that is capable of real-time processing and learning had to be implemented. A pre-trained database is not needed, as the framework learns new faces on-the-fly. For real time processing and recognition the computation lightness is important, as well as the precision. Therefore the LBP descriptor was chosen to be the main descriptor in the mentioned framework. The K- Nearest Neighbour classifier is used for matching, where the distance metric between the face representations is calculated using the χ2 distance score. To be able to classify an unknown face, a threshold is used when predicting. If the χ2 distance score returned is above a set threshold the learning module is initialized, where only key frames are extracted from the face and stored in the database. These key frames represent the face in different poses and expressions, thus assuring robustness for the real-time face recognition system. The NAO robot acts upon various ”events” based on the classifications done by the system. The performance of the system is evaluated by using available pre-existing face databases consisting of faces under varying conditions regarding illumination, facial expressions and pose. These tests were done by performing a K-fold cross validations. The validation results show high performance for both precision and speed. The face recognition system achieves 91.7% precision when evaluated on the yale face database A, and 99.8% precision for the AT&T database.