针对当前按摩机器人穴位自动定位技术存在着装限制、适用范围局限以及定位精度较差等问题，提出了一种基于背部形态分类与多层感知机网络的人体背部穴位定位方法（morphology classification-multilayer perceptron-based accurate point location, BMC-MPAPL）。通过采集大量不同人体背部图像数据结合骨骼关键点定位技术、核密度估计和最大类间方差方法，研究了背部形态的统计分布与有效分类。为克服衣物覆盖对定位的限制，基于分类结果制作包含背部关键点和大椎穴的各类坐标数据集，同时利用多层感知机网络深度学习模型实现大椎穴的自动定位。依据大椎穴的定位结果建立人体坐标系，结合中医骨度分寸法实现人体背部60个常见穴位的自动定位。结果表明：基于不同背部形态的大椎穴定位模型，实现了对大椎穴的高精度定位，在允许误差为20 mm时的平均精度为94.87%，较未进行背部分类的大椎穴定位模型精度提高13.37%。对于背部其他常见穴位，在患者服饰、背景没有限制的情况下，其定位精度在允许误差为20 mm时为91.58%。可见，所提方法有效提高了穴位自动定位的精度和适用性。

The current acupoint automatic positioning technology for massage robots is faced with issues such as dress restrictions, limited application scope, and poor positioning accuracy. A new method for human back acupoint positioning, based on back morphology classification and a multilayer perceptron network morphology classification-multilayer perceptron-based accurate point location(BMC-MPAPL) has been proposed. A substantial collection of diverse human back images, along with skeletal key point positioning techniques, kernel density estimation, and the maximum interclass variance approach, was used to investigate the statistical distribution and effective categorization of back morphologies. To counteract the restrictions imposed by clothing on positioning, a dataset encompassing key back points and the Dazhui acupoint was developed based on classification outcomes, and the automatic positioning of the Dazhui acupoint was accomplished through a deep learning model of the multi-layer perceptron network. Utilizing the Dazhui acupoint's positioning results, a human coordinate system was established, and the automatic positioning of 60 common back acupoints was achieved with the integration of ancient Chinese bone measurement methods. The results show that the Dazhui acupoint positioning model, tailored to various back morphologies, has realized high-precision positioning, with an average accuracy of 94.87% at an allowable error of 20 mm, marking a 13.37% enhancement over models without back classification. For other common back acupoints, the positioning accuracy stands at 91.58% within an allowable error of 20 mm, irrespective of patient attire or background constraints. It is concluded that the method presented effectively enhances the accuracy and applicability of acupoint automatic positioning.