基于随机森林深度特征选择的人体姿态估计

Human pose estimation based on random forest depth feature selection

针对以随机森林为分类器的人体姿态估计系统内存占用过大的问题,提出一种优化的随机森林模型,该模型在进行Bootstrap抽样前,引入Poisson过程并将其与深度信息相融合组建一个滤过网对原始训练数据集进行过滤,将一部分对后续分类起到非积极作用的特征样本点滤除,使训练数据集得到优化重构,进而较好地弥补随机森林在抽样过程中重复抽样以及重抽样样本代表性不强的缺点。实验结果表明了该优化模型的有效性,大大降低了系统的时间、空间复杂度,使得系统的适用性更强。

The human pose estimation system which uses the random forest as classifier has a problem about taking up too big memory footprint, so this paper puts forward an optimization random forest model to solve the problem above.The new model introduces the Poisson process and combines it with the depth information to form a filter before Bootstrap sampling, and then filter the original training dataset, moving the pixel sample which not plays a positive role away. After that the goal of refactor the training dataset is achieved. So the insufficient about repeated sampling and the weak representative of random forest can be improved. And the experimental results show this optimization is effective, reducing the time and space complexity of the system greatly, and makes the system more general.