人体行为识别中骨骼特征信息的鸡群优化选择算法

Selecting Optimal Skeletal Features Information in Human Action Recognition by Chicken Swarm Optimization

随着低成本深度传感器的出现,人体行为识别研究吸引了很多研究人员。由于这些设备提供了身体关节的三维位置等骨骼数据,使得基于骨骼的人体行为识别变得简单。但这些关节特征的信息存在部分冗余或者不必要的关节点信息,从而降低识别精度。为此,提出了用改进的CSO(Chicken Swarm Optimization)算法来优化关节点信息的方法,过滤一些不必要的关节点的特征信息,提高了识别精度。改进的CSO算法引入"佳点集"和OS(Ordered Subsets)方法,利用"佳点集"均匀化初始种群;利用OS方法将整个有序种群分成了3个部分,在进行位置更新时,引入猴群算法中的望-眺过程来比拟鸡群在觅食的过程中寻找食物的眺望过程作为其修正公式,加快了算法的收敛速度并且避免算法陷入局部最优。实验结果表明,提出的方法在UTKinect和Florence3D数据集上测试得到的精度分别达到98.01%和93.77%。

With the low-cost depth sensors developed, human action recognition has attracted the attention of many researchers. Since these devices provided skeletal data consisting of 3D positions of body joints, human action recognition became simple. But these might contain irrelevant or redundant features information of body joints that could cut down recognition accuracy. An improved CSO(Chicken Swarm Optimization) is used to optimize features information of body joints by filtering unnecessary data. The Good Point Set and OS(Ordered Subsets) are added to the improved CSO algorithm. The Good Point Set is used to uniform initial population.The OS method divides entire ordered population into three parts correspondingly. During the process of updating position information, the Monkey Algorithm's gaze afar motion is regarded as the process of gazing afar when chicken looking for food, and revised formula is established. It improved the convergence speed of algorithm and avoided falling into local optimal. The proposed approach has been tested on UTKinect-Action dataset and Florence3D-Action dataset. Experimental results show that our method gains test accuracy with 98.01% and 93.77% respectively.