研究発表

本稿では人体スケルトンやアピアランス情報を統合した行動理解のための検討を行う。人体スケルトンによる行動理解では間接位置から角度を取得して行動を識別、アピアランス情報を入力とした識別では複数特徴量の統合について記述する。実験では日常生活においていかに認識するかを検証する。

This paper describes a new accurate interpolation method of a low-resolution depth image using a high-resolution color image for the structural analysis of 3d data. Our method adopts the strategy to detect and omit regions where it is difficult to estimate depth value accurately. In order to achieve this, Color geodesic neighborhood is used as the local smooth surface, and depth data is interpolated with checking smoothness of connections between them. Pixels where they can not be smoothly connected are omitted automatically as indeterminable pixels in this process. Experiments with images of the Middlebury stereo datasets demonstrate that in the quantitative sense our method can interpolate depth images more accurately than other interpolation methods.