Effectiveness of dynamic image features for estimating daily conversation atmosphere

Abstract

To achieve a conversation atmosphere estimation method, we confirm the usefulness of optical flow values for conversation atmosphere estimation. We extracted only the response scenes from a database of free conversation between elderly individuals, analyzed two-dimensional optical flow parameters, and calculated the standard deviation values (FX-SD, FY-SD) of all frames for each dimension. These values express the dynamic characteristics of conversation gestures. As a result of comparing the FX-SD values and the FY-SD values of “smooth conversation scene (not-need help (NNH))” with that of “not-smooth conversation scene (need help (NH))”, the “NNH scenes” values are smaller than those of the “NH scenes”. We also confirmed the identification rate using support vector machine (SVM). The average identification rate is 72.3%. These results suggest that optical flow parameters are useful for identification, but the performance is not sufficient. In the future, we should evaluate more conversation databases and provide more in-depth discussion of the effectiveness.