CHLAC特徴の周期性解析による料理映像中の繰り返し調理動作区間の抽出と識別

摘要

This paper proposes a method for extracting segments that contain repetitive cooking motions such as "cutting" and "mixing", from cooking videos, and for recognizing the cooking motions in the segments. The proposed method extracts repetitive cooking motion segments by two types of features; One is a feature that depends on the location of the motion within video frames and the other is a feature invariant to the location. As a result, high identification accuracy is expected to be maintained on both a repetitive motion whose oscillation center moves along time and a repetitive motion with a constant oscillation center. Next, the proposed method analyses the periodicity of these feature values by Fourier transform and extracts the segments. On the other hand, in cooking motion classification, considering that the location of the cooking motion within video frames is various, the proposed method classifies the repetitive cooking motion segments by using a feature value invariant to the location of the motion within video frames. In an experiment for extracting segments that contain repetitive cooking motions, the method obtained an accuracy of 0.78, and for cooking motion classification, an accuracy of 0.77 was obtained. From these results, the effectiveness of our method was shown.%本報告では，料理映像から「切る」や「混ぜる」といった繰り返し動作が行われている映像区間を抽出し，その区間の調理動作を識別する手法を提案する．繰り返し動作区間の抽出には，映像フレーム中の動作の位置に依存しない特徴量と依存する特徴量という2種類の特徴量を用いる．これにより，繰り返し動作の振動中心が移動する動作および一定である動作の両方に対して，高い抽出精度を維持する．そして，これら特徴量の周期性をフーリエ変換により解析し，区間抽出を行う．一方，調理動作の識別では，映像フレーム中で調理動作が行われる位置が多様であることを考慮し，動作の位置に依存しない特徴量を用いる．繰り返し動作区間抽出の実験では0.78，調理動作識別の実験では0.77の精度が得られ，このことから本手法の有効性を確認した．