基于粒子滤波的个人导航系统算法研究

发布时间：2018-03-21 01:54

本文选题：个人导航　切入点：活动识别　出处：《厦门大学》2014年硕士论文　论文类型：学位论文

【摘要】：导航技术作为众多信息技术的代表,正悄然进入人类生活的细枝末节。如何适应复杂环境、融合多传感器信息实现更加精确的定位成为导航技术的关键所在,贯性导航系统避免了对于信号源的依赖,使用更加灵活,正逐渐成为个人导航技术研究的重要课题。然而惯性导航系统存在误差累积、容错能力差等特点,本文意在设计一种基于惯性传感器的个人导航系统,并结合机器学习支持向量机方法和粒子滤波,实现对于定位结果的优化。本文通过提取基于卡尔曼滤波的惯性导航系统所解算步长、航向变化角度等信息,建立基于步长、航向变化角度的航位推算运动模型,通过粒子滤波算法对运动轨迹进行优化。优化模块包括平面地图信息融合和活动识别纠正点融合两方面：首先,平面地图信息为航位推算正确性提供了重要的判断依据,本文假设在室内平面地图已知的情况下,利用平面地图信息,判断粒子滤波推算的正确性,即对每一步中每一个粒子分别进行推算,剔除错误粒子,对粒子权重进行二次优化,保证运动轨迹符合客观事实,从而实现纠正；另外,对惯性传感器以及气压传感器数据进行预处理,包括坐标变换、高通滤波、计算气压差值等过程,抽象出训练集进行训练,通过两层的支持向量机对人的活动进行识别,主要识别静止、走路、上下楼梯、上下电梯等活动,针对其中包含了地理信息的活动提取纠正点,并将其提供给粒子滤波模块,在粒子滤波推算的过程中纠正定位解算结果。通过实验可以发现,融合平面地图信息使系统修正了穿越墙壁的错误解算结果；加入活动识别纠正模块后,在基于卡尔曼滤波的导航系统解算结果误差较大情况下,累积误差控制在2%以内。本文的研究证明：通过融合平面地图信息,影响粒子权重的更新与传递,有效的剔除了错误粒子；同时借助支持向量机,对惯性传感器数据进行活动识别,识别准确率较高,通过二次优化识别结果,将带有地理信息的活动作为纠正点,粒子滤波融合纠正点信息完成修正,提高了系统整体精度；算法具有一定的可行性。
[Abstract]:Navigation technology, as the representative of many information technologies, is quietly entering the details of human life. How to adapt to the complex environment and integrate multi-sensor information to achieve more accurate positioning become the key of navigation technology. The penetration navigation system avoids the dependence on the signal source and becomes more flexible. It is gradually becoming an important subject of personal navigation technology. However, the inertial navigation system has the characteristics of error accumulation and poor fault tolerance. The purpose of this paper is to design a personal navigation system based on inertial sensors and combine machine learning support vector machine method with particle filter to optimize the localization results. In this paper, by extracting the information of step size and course change angle of inertial navigation system based on Kalman filter, the motion model of dead-reckoning based on step size and changing angle of heading is established. Particle filter algorithm is used to optimize the motion trajectory. The optimization module includes two aspects: the fusion of plane map information and the fusion of activity recognition correction points. Firstly, the plane map information provides an important basis for the correctness of the dead reckoning. This paper assumes that if the indoor plane map is known, the accuracy of particle filter calculation is judged by using the plane map information, that is, each particle in each step is calculated separately, the wrong particle is eliminated, and the particle weight is optimized twice. In addition, the data of inertial sensor and pressure sensor are preprocessed, including coordinate transformation, high-pass filtering, calculation of air pressure difference and so on, and the training set is abstracted for training. Through the two-layer support vector machine to identify the human activity, mainly to identify the activities such as static, walking, up and down stairs, up and down elevators and so on, to extract correction points for the activities containing geographic information, and to provide them to particle filter module. In the process of particle filter calculation, the result of location calculation is corrected. Through experiments, it can be found that the system can correct the result of error calculation through the wall by the fusion of plane map information, and after adding the activity recognition and correction module, the error of the result of navigation system based on Kalman filter is large. The cumulative error is controlled within 2%. The research in this paper proves that by merging the plane map information, which affects the updating and transmission of particle weight, it can effectively eliminate the wrong particles, and at the same time, with the support vector machine, the inertial sensor data can be recognized with high recognition accuracy. Through the quadratic optimization recognition result, the activity with geographical information is taken as the correction point, and the information of correction point is corrected by particle filter fusion, which improves the overall accuracy of the system, and the algorithm is feasible.
【学位授予单位】：厦门大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN966

【参考文献】