基於使用者步頻的音樂輔助跑步系統

Abstract

本論文是有關於我們所開發出的一個手機應用程式 MART (Music Assisted Run Trainer)，這個應用程式的目的是要利用控制音樂的節奏快慢來讓使用者可以跑得更輕鬆。MART可以透過使用者手機中的三軸加速度計跟陀螺儀，來即時推算出使用者的步頻，再藉由波形相似性疊加法的技術，來合成出與使用者步頻相同節奏的音樂。前人的腳步偵測研究，大都是採用閥值判斷的方法，但MART是使用一個類似音樂訊號處理中節拍追蹤的方法，來做腳步偵測跟步頻分析。本論文共提出了四種抓取步頻的方法，三種是使用時域上的自相關函數，最後一種則是使用頻域上的倒頻來做運算。我們設計了兩種實驗來驗證我們的系統，分別是對腳步預測的正確性，以及算出的步頻跟實際步頻的誤差值。最後根據實驗結果，我們發現時域的方法的正確性，都比頻域的方法來得高。

This thesis describes a smartphone app we developed called MART (Music Assisted Run Trainer). The goal is to assist users to jog easily and pleasantly through controlling the tempo of the music. MART estimates the runner’s SPM (steps-per-minute) in real-time via the tri-accelerometer and gyroscope in the smartphone. The music whose tempo is the same as the runner’s SPM is then synthesized via Waveform Similarity Based Overlap-Add (WSOLA). Most of the existing research in step detection uses threshold-based methods, while MART adopts a method which is based on a beat tracking algorithm. We propose four methods for finding the step period. Three of them are based on autocorrelation function in the time domain, while the other one is based on cepstrum in the frequency domain. We design two experiments to evaluate the performance of MART in detecting steps and estimating step frequency. Experimental result shows that the accuracy of the time-domain methods is better than the frequency-domain method.