人体姿势估计中随机森林训练算法的并行化

Parallelization for randomized forests used in human pose estimation

针对用深度图进行人体姿势估计算法中随机森林训练模块的资源消耗大、训练时间长等问题,提出在小规模的集群服务器上用消息传递接口技术对随机森林算法进行并行化加速,并结合算法进行优化以降低存储消耗和占用带宽等,进一步提高训练速度。实验结果表明,在小型集群服务器上不到一天时间完成一次训练,速度相比原来提升约30倍,分类器的像素识别率超过80%,骨架节点的实际误差也足够小,经加速后可以及时进行多次训练,从而完成对训练参数的调整和测试。

The randomized forests training algorithm used in depth image human pose estimation caused problems of large re- sources and training time cost. This paper proposed the parallelization design using message passing interface on small cluster server, then optimized to decrease the storage and bandwidth cost. Experiments show that the processing speed is enhanced by 30 times, one training cost less than one day on the cluster, the pixel precision of training classifier is more than 80% and skele- ton errors are smaller. The proposed parallelization method can finish the training in time to adjust and test the parameters.