轻量位姿估计算法

Lightweight pose estimation algorithm

针对实时二维姿态估计复杂模型中网络参数多、计算复杂度高的问题,文中提出一种轻量级人体姿态估计算法(SASNet)。该方法采用轻量型MobileNet v3作为骨干网络,引入空间域自注意力模型(Self-Attention)、通道域SENet模块两种维度的注意力模块,在此基础上融合构建SASM(Self-Attention SENet Module)模块,从而在减少参数、降低运算复杂度的同时,提升网络模型的性能。COCO校验集上的验证结果表明,所提算法与常见的轻量型位姿估计算法相比,能够以更少的参数量和复杂度实现更高精度的检测;SASNet与复杂模型OpenPose相比,参数量只有其11.9%;与轻量型模型Lightweight OpenPose相比,参数量减少20.5%。文中的轻量化改进方法可有效减少参数量,实现轻量化,精度较高。

In allusion to the problem that multiple network parameters and high computational complexity in the complex model of real-time 2D pose estimation,a lightweight human pose estimation algorithm(SASNet)is proposed. In this method,lightweight MobileNet v3 is used as the backbone network,and the Self-Attention module in spatial domain and SENet module in channel domain are introduced. On this basis,the SASM(self attention SENet module)is construction by fusion to improve the performance of the network model while reducing parameters and computing complexity. The verifition results on COCO check set show that,in comparison with the common lightweight pose estimation algorithms,the proposed algorithm can realize higher detection accuracy with fewer parameters and complexity. The parameter quantity of SASNet has only 11.9% compared with complex model OpenPose,and the parameter quantity is reduced by 20.5% compared with Lightweight OpenPose. The improved lightweight method proposed in this paper can effectively reduce parameter quantity,and achieve lightweight with high accuracy.