[Japanese | Thesis | Researches in Minoh Lab | Minoh Lab]
The goal of research presented in this report is to obtain basic knowledge useful to analyze a user's footstep features which are applicable to the user identification system of footstep features. With the expansion of ubiquitous computing, many appliances will be connected via network. Already several researches have been done for appliances to offer multiple services to the users. Since the users would like to access their favorite services easily, the appliances should identify which user is accessing them. At the same time user identification should not be troublesome. Now many user identification systems are proposed and used, for example: RFID (radio frequency identification) tags, iris recognition, finger print matching, and password checking. In those systems users should have or handle an appliance, which could be troublesome in operation. Moreover, iris recognition or finger print matching could be a mental burden for the users, since their sensitive personal data could be stored and possibly mishandled in such systems. User's footstep features identification has a little load and no mental burden for the users since they only walk on the floor with pressure sensors. In the previous research on footstep features, possibility of identifying users based on their walking was discussed. The difference of weight could have influence on footstep data, but the distribution of the subjects' weight was not analyzed in the previous research. The method to collect data could have influence on footstep analysis. In this report, first we discuss how to collect footstep data and we create the proper system to collect footstep data. Then, we analyze user identification performance of footstep features. To collect the footstep data, we use the weighting plate with four load cells. This plate measures 50cm x 50cm x 3cm. We collect footstep data in two following scenarios based on previous research:
Then we evaluate distributions of features obtained from the collected data. These features are same as in previous research. In consequence, it is shown that the features obtained in these scenarios have different distributions. We examine the possible effects of users' being aware of the position when they step on the plate. In order not to make them aware of the position where they step, we create the system using many plates to collect footstep data. We collect footstep data in the following two scenarios:
Then we evaluate distributions of features obtained from the collected data. In consequence, we show that features obtained in these scenarios have different distributions. As a result footstep data should be collected with condition that there is no height difference and that the users are not aware of the position. We collect footstep data with the above condition and we analyze user identification performance using these data. A procedure to analyze the user identification is as follows. We calculate covariance matrix from the collected data, and project the data onto three dimensions using principle component analysis. Then we hypothesize distributions of footstep data as :
Finally, we determine some distributions of user randomly in the above constraints and we analyze user identification performance from the overlapped domain of footstep data. As a result, the possibility of user identification is shown with condition that detection rate is 95%, error rate is 5%, for ten users. To examine the influence of users' weight, we analyze footstep data for similar conditions. As a result, the error rate rises much higher and it is revealed that weight is important for user's footstep features identification. A possibility of user identification could be shown with the condition that detection rate is 95%, error rate is 5%, for three users.