融合球空间下旋转角度编码的人体动作识别

Fusing Rotation Angle Coding in Spherical Space for Human Action Recognition

针对现有的人体动作识别方法较多考虑骨架结构的坐标和位移等平移信息,较少关注代表骨架结构的运动趋势以及代表关节、骨骼运动方向的旋转信息,提出一种融合球空间下旋转角度编码的时空卷积神经网络方法。通过人体动作在三维球空间中的映射,获取具有尺度不变性的角度信息,提取其动态角速度信息作为角度编码,表征动作轨迹中关节点和骨骼边的旋转信息;构建了时空特征提取与共现模块来更好地捕获数据的时空特征;用合适的融合策略对平移特征和旋转特征进行运动特征融合。实验结果证明了旋转角度编码有利于提升运动表征的准确性,以及时空特征提取与共现模块的有效性。

The existing human action recognition methods focus more on the translation information such as the coordinates and displacements of skeleton structure,and pay less attention to the motion trend of skeleton structure and the rotation information representing the motion direction of joints and bones.A spatio-temporal convolutional neural network method combining the rotation angle coding in spherical space is introduced.The angle information with scale invariance is obtained by mapping the human action in three-dimensional spherical space,and the dynamic angular velocity information is extracted as the angle code to represent the rotation information of joints and bones in the action trajectory.A spatiotemporal feature extraction and co-occurrence module(STCN)is constructed to better capture the spatiotemporal features of data.A suitable fusion strategy is utilized to fuse the translation features and rotation features.The experimental results show that the rotation angle coding benefits the accuracy improvement of motion representation and the effectiveness of the spatio-temporal feature extraction and co-occurrence module.