基于Android手机和机器学习的室内室外运动状态分类系统

发布时间：2019-01-22 20:34

【摘要】：随着人们生活水平的提高,人们越来越关心自己的健康问题。例如,经常开车会增加二氧化碳排放从而污染环境,长时间久坐会造成腰椎间盘突出等问题。为了打造绿色生活,减少环境污染,让人们健康生活,开发一套可以准确检测人们运动方式的系统,并且通过奖励惩罚规则引导人们更健康生活变得很有价值。本文主要实现了基于安卓手机内置GPS、加速度器和螺旋仪对使用者的运动数据进行收集,使用机器学习算法进行室外和室内运动状态分类并且有高精确度和奖励惩罚规则的一套系统。室外分类的运动状态有静止、走路、跑步、乘坐汽车、乘坐地铁、和骑自行车。室内分类的运动状态有静止、走路、坐下、上楼、下楼、电梯上行、电梯下行、摔倒、和帕金森抖动。系统先通过GPS,加速度器和螺旋仪器进行数据采集。在特征选择上,研究了如何选择更合适的特征来进行分类算法的实行,在加速度器上通过算法选择时间域中的波峰波谷一系列相关的特征和一些统计量特征,计步器统计的步数,Wi-Fi等特征。螺旋仪上通过算法提取合成角速度斜率的最大最小值。然后使用了监督式学习算法:决策树、随机森林、支持向量机、神经网络、朴素贝叶斯、KNN、逻辑回归算法进行运动状态分类。把收集到的数据进行取中位数和原始数据进行实验对比,探究出原始数据的中位数不如使用原始数据进行分类的效果好。通过调整传感器的使用组合(同时使用GPS、加速度器、螺旋仪,只使用其中两个,只使用其中一个)、机器学习算法的参数,进行实验并验证可以提高分类效果的性能。实验对比了机器学习分类算法在室内外运动状态分类中的优劣。选择出了性能最好的算法(随机森林、决策树、和神经网络)。研究表明同时使用GPS,加速度器和螺旋仪收集数据的分类精确度最高。室外运动方式分类上,随机森林算法得到了88.57%的分类精确率;室内运动状态分类上,决策树算法得到了97.54%的分类准确度。并且,用户从5人扩大到了50人去采集数据,利用假设检验验证了样本容量小的时候用户数据的可代表性和算法准确度的可行性。系统可以提高鉴别帕金森抖动和摔倒的准确性。在手机端和服务器端用规则引擎Drools分别制定了奖励惩罚规则,引导人们能够更健康的运动生活。这套系统的开发很具有人类健康方面的应用价值,利用了身边的智能手机和机器学习算法就可以对人们运动方式进行高准确度的检测分类,而且通过奖励惩罚规则,对人们健康绿色生活具有很好的引导作用,利于可持续发展。
[Abstract]:With the improvement of people's living standard, people are more and more concerned about their health problems. For example, frequent driving increases carbon dioxide emissions and pollutes the environment, and prolonged sitting causes problems such as lumbar disc herniation. In order to create a green life, reduce environmental pollution, and make people live healthy, we can develop a system that can accurately detect the way people move, and guide people to live healthier lives through the rules of reward and punishment. This paper mainly realizes the collection of the user's motion data based on the built-in GPS, accelerator and spiral device of Android phone. A system that uses machine learning algorithms to classify outdoor and indoor motion states with high accuracy and reward and punishment rules. Outdoor sports are classified as motionless, walking, running, taking cars, taking the subway, and riding bicycles. Indoor categories of motion are static, walking, sitting, going upstairs, going downstairs, elevators going up, elevators descending, falling down, and Parkinson's jitter. The system firstly collects data by GPS, accelerator and spiral instrument. In feature selection, we study how to select more suitable features to carry out the classification algorithm, and select a series of related features and some statistical features in the time domain by using the algorithm on the accelerator. The number of steps counted by the pedometer, Wi-Fi and other features. The maximum and minimum synthetic angular velocity slope is extracted by the algorithm on the helical instrument. Then the supervised learning algorithms such as decision tree, stochastic forest, support vector machine, neural network, naive Bayes and KNN, logic regression algorithm are used to classify the motion state. The median of the collected data is compared with the original data, and it is found that the median of the original data is not as good as the classification of the original data. By adjusting the combination of sensors (using GPS, accelerator, helical instrument, only two of them, only one of them), the parameters of the machine learning algorithm are tested and verified to improve the performance of classification. The advantages and disadvantages of machine learning classification algorithm in indoor and outdoor motion state classification are compared. The best algorithms (random forests, decision trees, and neural networks) are selected. The results show that the classification accuracy is the highest using both GPS, accelerator and helical instrument. In outdoor motion classification, the accuracy rate of classification is 88.57% by random forest algorithm, and 97.54% by decision tree algorithm in indoor motion state classification. Moreover, the user is expanded from 5 to 50 to collect data, and the feasibility of the representativeness of user data and the accuracy of the algorithm are verified by hypothesis test when the sample size is small. The system can improve the accuracy of identifying Parkinson's jitter and fall. Drools, a rule engine on the mobile and server side, has established reward and punishment rules to guide people to a healthier sports life. The development of this system is of great value in the field of human health. With the help of the smartphone and machine learning algorithms around us, we can detect and classify people's movements with high accuracy, and we can use the rules of reward and punishment. It has a good guiding effect on people's healthy and green life and is conducive to sustainable development.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181;TN929.53

【相似文献】