面向家庭场景的轻量级婴幼儿姿态估计方法-《中国医学物理学杂志》

文章信息/Info

作者:: 万金亮1; 熊启亮1; 刘苑2; 莫杰义1; 谌颖1; 1.南昌航空大学仪器科学与光电工程学院，江西南昌 330063； 2.重庆医科大学附属儿童医院康复科，重庆 400044

Author(s):: WAN Jinliang1; XIONG Qiliang1; LIU Yuan2; MO Jieyi1; CHEN Ying1; 1. School of Instrument Science and Optoelectronic Engineering, Nanchang Hangkong University, Nanchang 330063, China 2. Department of Rehabilitation, Childrens Hospital of Chongqing Medical University, Chongqing 400044, China

摘要:: 如何有效降低婴幼儿姿态估计网络模型的大小是制约婴幼儿姿态估计技术“家用化”的关键问题，提出一种面向家庭场景的婴幼儿轻量级姿态估计方法，该方法利用轻量级网络MobileNetV3作为编码主干，并在解码部分引入基于PixelShuffle上采样模块以降低模型参数量。同时，本文通过引入坐标注意力机制（CA）的方式，能够更好地捕获位置信息和通道特征信息，突出图像中小目标和遮挡人体关键点的特征信息。最后，进一步修改并行交叉连接卷积部分以增强特征信息提取能力。在人体姿态估计通用数据集COCO以及面向婴幼儿姿态估计的专用数据集SyRIP上分别进行方法性能的验证，实验结果表明，在计算量（GFLOPs）只有0.96的情况下，在COCO和SyRIP数据集中模型平均精度分别达到73.5%、91.0%，证明本文提出的方法在显著降低模型参数和计算量的同时，不会损失姿态估计模型的准确性。本文提出的轻量化估计模型有望部署于智能终端等家用设备上，从而实现家庭场景下婴幼儿异常姿态的智能评估。

Abstract:: Abstract: How to effectively reduce the size of infant pose estimation network models is a key issue restricting the "home-use" of infant pose estimation technology. Therefore, a lightweight method for infant pose estimation in home scenarios is proposed. The method takes the lightweight network MobileNetV3 as the encoding backbone and utilizesa PixelShuffle up-sampling module in the decoder for reducing the quantity of model parameters. Meanwhile, coordinate attention mechanism is used to better capture location information and channel feature information, highlighting the feature information of small targets and occluded human keypoints. Besides, the parallel cross-correlation convolution is further modified to enhance the capability of feature information extraction. The methods performance is verified on the general pose estimation dataset (COCO) and the dedicated infant pose estimation dataset (SyRIP). The results show that, with a calculation volume (GFLOPs) of only 0.96, the method achieves average accuracies of 73.5% and 91.0% on COCO and SyRIP datasets, respectively, proving that it can significantly reduce the quantity of model parameters and calculation volume without sacrificing pose estimation accuracy. The proposed lightweight estimation model is expected to be deployed on home appliances such as smart terminals, thereby realizing intelligent estimation of abnormal infant poses in home scenarios.