Analysis of Factors Improving Accuracy of Passive User Identification with Streams of Face Images for Ubiquitous Commerce

摘要

Ubiquitous commerce services set new requirements for access control methods, e.g. to enable full payment automation it is necessary to passively perform initial customer identification at point of sale. Face biometrics seems to be promising in these scenarios since it does not require user to continuously carry relevant object nor to actively participate. In theory, the accuracy of customer identification should improve with the number of face images, however additional low-quality face images that are included in the recognition stream actually can degrade identification accuracy. Therefore, in this work various criteria of filtering image stream are analyzed to improve accuracy of final identification decision: user attention (face rotation), user mimics, or user height different from the template. The analysis is performed for various lightning conditions, various recognition algorithms, various sensor types, and various recognition distances in the environment simulating real point of sale. In this paper we report on new systematic experiments performed on our earlier context-aware passive payment authorization system. Results have been obtained as an effect of data mining and statistical analysis of log sets.