近年運動風氣興起，許多運動愛好者在郊區跑步或健走時亦習慣配戴耳機聆聽智慧型行動電話上播放的音樂。因聽覺封閉或分心而降低對週遭環境的感知，可能增加了與路上車輛發生事故的機會。 為了改進行人在此情境下的安全，本論文提出了一個簡單的系統設計並評估，以便將來在智慧型行動裝置平台上實現偵測後方來車的預警系統。 環境中的聲音將可由與行動裝置連接的指向性麥克風來收音，並經過數位訊號處理的方式取出短時間長度的音框內基本的聲音特徵：RMS(Root Mean Square)、ZCS(Zero Crossings)、SPC(Spectral Centroid)、SPR(Spectral Rolloff)。並將所取的得聲音特徵輸入至不同的機器學習分類器：kNN(K Nearest Neighbor)、MLP(Multi Layer Perceptron)、Decision Tree、Random For-est做分類，判斷該音框內是否存在來車的事件。 本研究呈現系統的辨識率及可行性，也檢視了不適用的情境。;More and more people tend to carry smart phones and wear headphones or headsets, listening to the music while they are jogging or walking in the suburb area. This behavior could bring distraction or temporarily losing the hearing of environmental background and cause accident to happen. This work proposes a simple design for audio-based early warning system of vehicle approaching event for improving pedestrian’s safety and gives evaluation. Sound signals were collected by an external directional microphone connected to the smart phone. Multiple feature techniques like root mean square, zero crossings, spectral centroid, and spectral rolloff were applied on the short-time frames of audio samples. Multiple machine learning classifiers like K Nearest Neighbor, Multi-layer Perceptron, Decision Tree and Random Forest were applied to classify the audio frames to detect vehicle approaching sound. The results showed the accuracy and the feasible of the system, also point out the circumstance can’t be applicable.
