Systematic selection of local correlation parameters for optical flow-based gesture recognition

We develop a real-time, optical flow-based gesture recognition system for human-robot interactions. In order to robustly estimate the right optical flow related to human gestures by the correlation-based technique, the following parameters must be selected appropriately in advance: the number of grid points, grid point intervals, search window size, pixel thinning rate, image sampling rate, and the size of correlation blocks. In our previous work (1999) these parameters were determined by the operator in a heuristic/empirical way. This paper presents a method to systematically select the local correlation parameters that ensure robust gesture recognition, which was not discussed in the previous study. We verified through various experiments that the combination of an optical flow-based gesture recognition technique with the proposed method can offer high recognition rates (overall 85% or more) for unspecific gesturers over a wide range of the gesturer-camera distance.