基于强化学习的多人姿态检测算法优化

OPTIMIZATION OF ESTIMATION ALGORITHM FOR THE MULTI-PERSON POSE BASED ON REINFORCEMENT LEARNING

在多目标人体姿态检测算法过程中,人体的定位精度依然不够精确。针对该问题,采用速度与精度兼顾的SSD算法作为目标检测器获得人体的初步包围框,定义该包围框为智能体,引入强化学习。采用马尔科夫决策过程以及Q网络组成的目标精细模型对智能体训练其九种动作,分别为左上角与右下角两个点的四方向进行迭代调整以及终止策略,使得包围框达到更贴近人体的效果。结合先进的Stacked hourglass算法作为姿态检测器,对调整后的包围框进行姿态预测。该算法的引入使得多目标人体检测算法在MPII数据集上的精度提升了1.6 mAP,达到了73.7 mAP。

The accuracy of locating people is imprecise in multi-objective human pose estimation. Aiming at this problem, we applied SSD with both speed and accuracy as an object detector to obtain the preliminary bounding box of human body. The algorithm defined the bounding box as an agent and introduced the reinforcement learning. Markov decision-making process and the objective fine model consisting of Q net were applied to train the agent for nine actions. The training actions included iteration adjustments and termination policy directing to two points from the top left corner and points from the bottom right corner respectively, which made the bounding box closer to the human body. The advanced Stacked hourglass algorithm was implemented as pose detector to predict the adjusted bounding box. The introduction of the referred algorithm improves the accuracy of multi-objective human pose estimation on MPII dataset by 1.6 mAP and reaches 73. 7 mAP.