テンソル分解に基づく多視点映像を用いた人物の動作認識手法の検討

摘要

We propose a Tensor Decomposition based algorithm that recognizes the observed action performed by an unknown person and unknown viewpoint not included in the database. Our previous research aimed motion recognition from one single viewpoint. In this paper, we extend our approach for human motion recognition from an arbitrary viewpoint. To achieve this issue, we set tensor database which are multi-dimensional vectors with dimensions corresponding to human models, viewpoint angles, and action classes. The value of a tensor for a given combination of human silhouette model, viewpoint angle, and action class is the series of mesh feature vectors calculated each frame sequence. To recognize human motion, the actions of one of the persons in the tensor are replaced by the synthesized actions. Then, the core tensor for the replaced tensor is computed. This process is repeated for each combination of action, person, and viewpoint. For each iteration, the difference between the replaced and original core tensors is computed. The assumption that gives the minimal difference is the action recognition result. The recognition results show the validity of our proposed method, the method is experimentally compared with Nearest Neighbor rule. Our proposed method is very stable as each action was recognized with over 75% accuracy.%本稿では多視点映像について，認識用データベース内に含まれない未知人物の動作をテンソル分解に基づく手法により認識する手法を提案する．本手法では，動作，人物，視点の各次元から構成されるテンソルを認識用データベースとして構築することで，任意の視点からの動作誅識を可能にする手法を検討する．動作認識ではデータべ-スに含まれない未知人物の動作を入力した際，入力データを認識用データベース内の一つのデータと置き換え，motion signatureを算出する．この過程を認識用データベース内のすべての人物，動作について繰り返し，それらの場合のmotion signatureを求める．その中で最もmotion signatureが類似するデータから，入力データの動作クラスを判別する．未知人物の動作認識実験を行い，本手法の有効性を評価するために従来法と比較した．6種類の歩行動作を対象に動作認識実験を行った結果，提案手法は75%以上の認識率で各動作について安定的に認識できることを明らかにした．