基于大间隔最近邻的人体动作识别

Human Action Recognition Based on Large Margin Nearest Neighbor

  • 摘要: 为实现日常生活中动作的识别,以提高家庭服务机器人的服务质量,为人类提供安全舒适的环境,提出了一种基于马氏距离的度量学习方法进行人体动作的识别.首先,利用Kinect获取人体动作的关节点数据.然后,基于关节点数据构建动作敏感特征集合,即由人体的关节点坐标构造人体的结构向量以及相应的角度,并对每一样本的长度进行归一化处理.采用大间隔最近邻(LMNN)分类算法进行马氏距离学习得到变换矩阵L,将归一化之后的原始数据映射到更优特征空间.最后,采用k近邻算法进行动作识别.在自建的数据集上,得到97%的识别率.实验结果表明,LMNN算法能够改善数据的分布,即缩小类内距离,扩大类间距离,较好地完成人体动作识别的任务.

     

    Abstract: In order to recognize actions in daily life for improving the service quality of the home service robot and providing a safe and comfortable environment for human, a metric learning method based on Manhattan distance is proposed for human action recognition. Firstly, Kinect is used to acquire the joint point data of human action. Then, the action sensitive feature set is constructed based on the joint point data, that is, the structure vectors of human and their corresponding angles are constructed based on the human joint point data, and the length of each sample is normalized. The large margin nearest neighbor (LMNN) method is adopted to obtain the transformation matrix L by learning Manhattan distance. And the normalized data is mapped to a better features space. Finally, the k-nearest neighbor algorithm is utilized to recognize the human actions. Based on our dataset, the accuracy of 97% is achieved. Experimental results show that the LMNN algorithm can improve the distribution of the data (that is, the intra class distance is reduced, and the inter class distance is expanded), and can complete the human action recognition task.

     

/

返回文章
返回