智能制造技术与系统研究中心知识库

实时、鲁棒的高精度人体运动感知是一个非常有挑战性的问题。近年来，人工智能、虚拟现实技术快速发展，对人体运动感知技术和设备的需求越来越大。随着微机电系统（MEMS：Micro ElectroMechanical System）系统的发展，微型化、低功耗、低成本的惯性测量单元（IMU：Inertial Measurement Unit）正广泛地被集成到移动终端和智能穿戴设备中，这为利用惯性传感器进行人体运动感知提供了广阔的应用前景。基于微惯性技术的行人航迹推演（PDR： Pedestrian Dead-Reckoning）是用于室内导航和室内定位的众多技术之一，其具有便于携带、成本低廉、对空间环境没有特殊要求等独特优势，因此也逐渐成为目前研究的热点之一。

本文围绕基于微惯性技术的人体位姿感知，特别是行人航迹推演方面展开研究。首先，系统地研究了基于微惯性技术的行人航迹推演（PDR）系统的关键技术；其次，针对行走和混合运动下的行人航迹推演问题，提出了一个有效的解决方案；最后，设计并初步实现了一个人体位姿感知系统。本文的主要工作有：

第一，基于微惯性技术的行人航迹推演（PDR）系统研究。首先，对基于微惯性技术的行人航迹推演（PDR）系统的设计方法、研究热点以及存在的问题进行了系统的分析和总结；其次，提出了一个PDR系统一般性的框架结构，该框架凝练出6个PDR系统所涉及的相对独立的模块；最后，在此基础上，创新性地提出一个PDR系统设计的技术路线图结构，从宏观和微观角度对PDR系统进行了高度概括。

第二，基于微惯性技术的行走和原地踏步混合运动下的行人航迹推演算法研究。首先，提出了一个自适应的步态划分算法，该算法分为预划分和修正两个阶段，并且拥有自适应调节判定阈值的能力。实验测试结果显示所提算法检测精度达到甚至超过现有最好成果。其次，提出了一个基于多层感知机模型的运动分类算法，该算法将按步划分好的运动数据作为输入，识别每步是属于行走运动还是原地踏步运动。实验结果显示该方法对行走和原地踏步混合运动下的运动分类准确率高达99%。此外，在此算法基础上提出了一个步长估计模型，该模型融合了常用的几个步长估计模型，实现了步长估计精度的提升。最后，利用二阶扩展卡尔曼滤波估计IMU姿态，并融合一些航向约束和修正算法来估计每一步的航向。实验结果显示所提算法估计出的运动轨迹精度与现有成果相当。

第三，基于可穿戴传感器的人体全位姿感知系统的设计与实现。首先，设计和实现了基于Unity 3D的数据手套演示系统。其次，实现了基于Unity 3D的人体姿态感知系统。最后，融合上述混合运动下的行人航迹推演算法，初步实现了一个基于WinForm+c#开发的行人航迹演示系统。

Real-time and robust high-precision human motion perception is a very challenging problem. In recent years, with the rapid development of artificial intelligence and virtual reality technology, the demand for human motion sensing technology and equipment is increasing. With the development of micro-electro-mechanical system (MEMS), miniaturized, low-power and low-cost inertial measurement units (IMU) are being widely integrated into mobile terminals and smart wearable devices. It provides a broad application prospect for motion sensing of human body with inertial sensors. In addition, Pedestrian Dead-Reckoning (PDR) based on micro-inertial technology is one of the many techniques used in indoor navigation and indoor positioning, which has unique advantages, such as easy to carry, low cost, no special requirements for space environment, etc. Therefore, it has gradually become one of the hot spots in the current research.

This paper focuses on the human pose tracking based on the micro-inertia technology, in particular the research of PDR. Firstly, the key technology of PDR based on micro-inertial techniques is systematically introduced and studied. Secondly, an effective solution is proposed to solve the problem of PDR under the condition of walking and marking time mixed movement. Finally, the human positioning and posture perception systems are designed and implemented. The main work of this paper is as follows:

Firstly, the systematic study of pedestrian track inference based on micro-inertial technique. First of all, the design methods, research hotspots and existing problems of Pedestrian Dead-Reckoning (PDR) system based on micro inertial techniques are systematically analyzed and summarized in this paper. Then, a general framework of PDR system is proposed. The framework condensed the six relatively independent modules involved in the PDR system. Finally, on this basis, an innovative technology roadmap structure for the design of PDR system is proposed. In this way, the PDR system is highly summarized from both macro and micro perspectives.

Secondly, the PDR algorithm based on the micro-inertial technique is studied under the hybrid motion of walking and marking time. First of all, an adaptive gait partition algorithm is proposed, which is divided into two stages: pre-partition and correction, and has the ability to adjust the threshold adaptively. The experimental results show that the gait phase detection accuracy of the proposed algorithm reaches or even exceeds the best available results. Next, a motion classification algorithm based on multi-layer perceptron model is proposed, in which the step-by-step motion data is used as input to identify whether each step belongs to walking motion or marking time motion. The experimental results show that the accuracy of the proposed method is as high as 99% for motion classification under the mixed motion of walking and marking time. Then, a step-size estimation model is proposed, which combines several common step-size estimation models and improves the accuracy of step-size estimation. Finally, a second-order extended Kalman filter is used to estimate the IMU attitude, and some course constraints and correction algorithms are combined to estimate the heading of each step. The experimental results show that the accuracy of the motion trajectory estimated by the proposed algorithm is similar to that of the state-of-the-art.

Thirdly, the perception systems of positions and postures for human body based on wearable sensors are designed and implemented. Firstly, a data glove demonstration system based on Unity 3D is designed and implemented. Secondly, a human attitude sensing system based on Unity 3D is implemented. Finally, a PDR demonstration system based on WinForm and c # is preliminarily implemented by integrating the PDR algorithm under the hybrid motion mentioned above.