多感測器系統與即時性動作辨識

Abstract

現今因人口逐漸老化，如何有效率且自動化的紀錄老人整天的日常生活作息與運動情況 並提供最完善的醫療照護已成為相當重要的議題，此篇論文中我們在 Raspberry Pi 上設計了 一套可同時且準確的接收身體各部位慣性資料 (加速度、角速度) 之系統，將感測器配置於手 腕、手臂、腰、胸、大腿、腳踝，並且利用所蒐集來的慣性資料使用 machine learning 技術 (Support Vector Machine) 進行站、坐、躺、走、跑、舉啞鈴、上樓梯、下樓梯喝水等動作辨 識，我們將資料依 windowSize 1s(約 140 筆資料) 為單位進行切分，再去做 feature extrac- tion 從原始資料中抓出特徵，且分別取出 time domain feature 與 frequency domain feature 詳細情況在文後會有更完整的說明，之後從眾多 feature 中使用 feature selection 方法 (relief feature selection) 從中找出相對重要之 feature 與重要的感測器配戴位置來達到減少計算量之 效果，在實驗部分我們也有去比較不同的感測器數量最終所獲得準確率高低，最後經由實驗證 實系統可及時 (大約每秒可辨識一次) 且準確的預測出使用者當下所進行的動作，此數據亦可提供給專業醫療人員做參考以擬定出更完善之醫療照護。

Because of the serious aging population in these few year.How to efficient and automated record the elderly daily life and their movement situation already become a very important issue. In this paper we build a system on Raspberry Pi it can accurate reception of inertial information(accelerometer, angular velocity )from various parts of the body at the same time, we put the inertial sensors at arm, wrist, chest, waist, thigh and ankle,we use the inertial data collected by system and machine learning technique(Support Vector Machine)to do activity recognition such as Stand, Sit, Lay, Walk, Run, UpStairs, DownStairs, Drink. We segment data as windowSize 1 second(140 piece of data),then take the segmented raw data to do feature selection.Then we get the time do- main feature and feature domain feature ,We will discuss this matter in a little more detail as tje follow paper. After feature extraction we take the feature data to do feature selection (Relief Feature Selection)because we want to know which feature is relevant feature and which is irrelevant, it can smartly reduce calculating time. At experimental result part we compare the relation of sensor number and accuracy. Finally our system can accurately recognize user activity in real-time(recognize every second).This information can also be provided to professional medical staff for reference to develop better medical care.