一种简便的视角无关动作识别方法

摘要

It is difficult to recognize the human actions under view changes in daily living.In order to solve this problem, a novel multi-view space hidden Markov model algorithm for view-invariant action recognition based on view space partitioning is proposed in this paper.First, the whole view space is partitioned into multiple sub-view spaces according to the rotation direction of a person relative to camera.Next, a view-robust feature representation by combination of the bag of interest point words in shot length-based video and amplitude histogram of local optical flow is utilized for describing the information of human actions.Thereafter, the human action models in each sub-view space are trained by HMM algorithm.Finally, the unknown view action is recognized via the likelihood proba-bility weighted fusion of the corresponding action models in multi-view space.The experimental results on multi-view action recognition dataset IXMAS demonstrated that the proposed approach is easy to implement and has satis-factory performance for the unknown view action recognition.%针对日常生活中人体执行动作时存在视角变化而导致难以识别的问题，提出一种基于视角空间切分的多视角空间隐马尔可夫模型（ HMM）概率融合的视角无关动作识别算法。该方法首先按照人体相对于摄像机的旋转方向将视角空间分割为多个子空间，然后选取兴趣点视频段词袋特征与分区域的光流特征相融合，形成具有一定视角鲁棒性特征对人体运动信息进行描述，并在每个子视角空间下利用HMM建立各人体动作的模型，最终通过将多视角空间相应的动作模型似然概率加权融合，实现对未知视角动作的识别。利用多视角IXMAS动作识别数据库对该算法进行测试的实验结果表明，该算法实现简单且对未知视角下的动作具有较好识别结果。