融合注意力机制的人体关键姿态估计

Human key pose estimation based on attention mechanism

为提高与特定动作相关的目标关键姿态识别准确率,提出融合注意力机制的人体关键姿态估计方法。针对特征提取冗余和通道信息损失问题,在网络不同阶段融入注意力机制,帮助模型从空间和通道维度提取对任务更加重要的特征;考虑与特定动作相关的目标关键姿态误识别代价更大,设计一种代价敏感损失函数,提升模型对目标关键姿态的识别性能。实验结果表明,所提方法在COCO数据集和CargoSorting数据集上对目标关键姿态识别精度分别提高了1.4%和1.5%。

To improve the accuracy of target key pose recognition related to specific actions,a human key pose estimation method integrating attention mechanism was proposed.Aiming at the redundancy of feature extraction and the loss of channel information,attention mechanism was integrated into different stages of the network to help the model extract more important features from the dimensions of space and channel.Considering the higher cost of target key pose recognition related to specific actions,a cost sensitive loss function was designed to improve the performance of target key pose recognition.Experimental results show that the proposed method improves the accuracy of target key pose recognition by 1.4%and 1.5%on COCO dataset and CargoSorting dataset respectively.