针对当前动作识别技术中通常忽略上下文信息, 且其局部形状描述对噪声敏感, 难以有效处理运动模式的类内变化等不足, 提出一种3D运动轨迹解析耦合隐马尔可夫模型的动作识别算法。定义四种轨迹基元(直线、平面圆弧、左螺旋线与右螺旋线), 采用基于Kalman 的线性平滑器来降低干扰信息与异常值, 将3D轨迹分解和解析成不同的轨迹基元, 充分减少干扰信息对轨迹数据的影响, 同时保持形状和运动特性。为了避免丢失详细的局部形状信息, 设计一种形状描述符, 将轨迹基元进一步分割成子基元, 将每个运动轨迹表示为子基元的时间序列。引入隐马尔可夫模型, 根据运动数据的先验知识来模拟运动分类, 基于最大后验概率准则来进行运动识别。实验结果表明: 与当前流行的动作识别算法比较, 所提算法能有效识别各种动作类别, 对尺度变化, 以及局部噪声和局部遮挡具有更高的识别精度。

In the field of motion recognition technology, the local shape descriptor is usually insensitive to noise, and it is difficult to effectively deal with the intra class variations of motion patterns, the motion recognition algorithm based on 3D trajectory analysis coupled with hidden Markov model was proposed.Four trajectory primitives (straight line, plane arc, left helix and right helix) are defined, and a Kalman-based linear smoother is used to reduce the interference information and outliers, and the 3D trajectory is decomposed and parsed into different trajectory primitives. The effect of interference information on the trajectory data is sufficiently reduced while maintaining shape and motion characteristics. In order to avoid losing detailed local shape information, a shape descriptor is designed to further segment the trajectory primitive into sub-primitives, and each motion trajectory is represented as a time series of sub-primitives. The hidden Markov model is introduced, and the motion classification is simulated according to the prior knowledge of the motion data, and the motion recognition is performed based on the maximum posterior probability criterion. The experimental results show that compared with the current popular action recognition algorithm, this algorithm can effectively identify the various categories of movement, visual invariance to scale changes, and local noise and occlusion robustness.