Abstract: Human action recognition is an important technology in the fields of intelligent monitoring,human-machine interaction,robotics and so on.The action recognition method based on human skeleton sequence has inherent advantages in the face of complex background,changes in human scale,viewing angle and motion speed,and the Spatial-Temporal Graph Convolution Networks model(ST-GCN)has excellent recognition performance in human behavior recognition.Aiming at the problem that the partition strategy in ST-GCN model only focused on local actions,a new partition strategy was designed.The connection between the information of each part of the body and the local movement were strengthened by associating the root node with the farther node.The adjacent regions of the root node were divided into five regions: root node itself,centripetal group,remote centripetal group,centrifugal group and remote centripetal group.At the same time,different weights were given to each region to improve the perception ability of the model to the overall action.The experimental tests were carried out in public datasets and real scenes.The results showed that the Top-1 classification accuracy of 31.1% was obtained on the Kinetics-skeleton dataset,which was 0.4% higher than the original model.The Top-1 classification accuracy of 83.7% and 91.6% were obtained on the two sub-datasets of NTU-RGB+D,which were 2.3% and 3.3% higher respectively than the original model.In the real scene,the proposed model had a high recognition rate for actions with obvious changes and great differences,such as push-ups and jogging.The recognition rate of local movements and movements with similar changes was low,such as clapping and shaking head,which still had room for further improvement.

Key words: action recognition, deep learning, spatial-temporal graph convolution networks model, partition strategy, skeleton sequence

摘要： 人体动作识别是智能监控、人机交互、机器人等领域的一项重要技术,基于人体骨架序列的动作识别方法在面对复杂背景以及人体尺度、视角和运动速度等变化时具有先天优势。时空图卷积神经网络模型(ST-GCN)在人体行为识别中具有卓越的识别性能,针对ST-GCN网络模型中的分区策略只关注局部动作的问题,设计了一种新的分区策略,通过关联根节点与更远节点,加强身体各部分信息联系和局部运动之间的联系,将根节点的相邻区域划分为根节点本身、向心群、远向心群、离心群和远离心群等5个区域,同时为各区域赋予不同的权重,提升了模型对整体动作的感知能力。最后,分别在公开数据集和真实场景下进行实验测试,结果表明,在大规模数据集Kinetics-skeleton上获得了31.1%的Top-1分类准确率,相比原模型提升了0.4%;在NTU-RGB+D的两个子数据集上分别获得了83.7%和91.6%的Top-1性能指标,相比原模型提升了2.3%和3.3%;在真实场景下,所提模型对动作变化明显且区别大的动作如俯卧撑和慢跑识别率高,对局部运动和动作变化相近的动作如鼓掌和摇头识别率偏低,尚有进一步提高的空间。

关键词: 动作识别, 深度学习, 时空图卷积神经网络模型, 分区策略, 骨架序列