摘要
为解决在Kinect平台下人体动作识别中时空复杂性的问题,提出一种基于特征选择的模板识别方法。根据人体不同位置关节点对动作表达的贡献度的不同,将骨骼模型60维的关节点数据转化成24维的距离特征向量,该特征模型能够在空间上对动作进行表示,具有一定不变性,计算复杂度低;结合动态时间规整的思想,解决动作识别在时间轴上不统一的问题;基于所提出的方法实现动作识别系统,定义6种基于交互的上肢动作,在此动作库中进行两个实验共1320次测试,两个实验的平均识别率为93.6%和89.8%,实验结果验证了该方法的鲁棒性和有效性,可以满足交互任务的需求。
To solve the variety and spatic-temporal complexity of human action recognition based on Kinect,a template recognition method based on selecting suitably features was proposed.The full skeleton formulated in a 60 Dfeature vector was tuned to an 24Djoint-distance feature vector according to the contribution of defferent joints to action presentation.The feature model was an action discriptor with scaling invariance and easy-computing characteristics.A dynamic time wraping algorithm was proposed to solve the speed problem of action recognition.An action recognition system was developed based on the proposed method.6kinds of upper limb actions were defined for interactive requirements.Two experiments were conducted to evaluate the method on a dataset with 1320 action instances.The results show that the proposed method can achieve average precisions of 93.6%and89.8%through two experiments respectively,which demonstrates that the method is effective,robust and entirely meets inte ractive requirements.
出处
《计算机工程与设计》
北大核心
2016年第4期1056-1061,共6页
Computer Engineering and Design
基金
国家自然科学基金项目(61100091)
教育部留学回国人员科研启动基金项目([2013]693)
辽宁省高等学校杰出青年学者成长计划基金项目(LJQ2012007)
沈阳工业大学青年学术骨干教师培养基金项目([2011]41)
沈阳工业大学博士科研启动基金项目([2011]30)
关键词
人机交互
骨骼跟踪
动作识别
动态时间规整
KINECT
human-computer interaction
skeletal tracking
action recognition
dynamic time warping
Kinect
