In this thesis, we focus on the following two problems in automatic recognition of electric power meter. One is the ability to decrease the sensitivity of an automatic recognition system due to the variation of illumination condition and the other is the automatic extraction of parameters which are necessary to help any non- experts in vision setup the system easily.
To overcome the variation of illumination condition, we introduce a reflectance model and an adaptive Kalman filter. We obtain the reflectance model of a power meter through experiments and use it to determine the threshold in change detection and the initial coordinate for setting up the system. Also, using the adaptive Kalman filter, we overcome the problem induced by a sudden variation of illumination.
For the automatic setup of the system, we use the plane projective geometry which needs four corresponding points between the model and the image. A template matching algorithm and the DDP(divided dynamic programming) based method are developed to compute four corresponding points.
We demonstrate that proposed automatic recognition system can robustly extract visual events from a variety of electric power meters in a realistic laboratory environment.