以智慧型行動裝置進行自動偵測異常路面

Abstract

近年來，台灣的道路工程品質往往給人「地無三里平」的刻板印象。根據法務部的從民國 94 年至民國 96 年的數據顯示，因為道路品質所引發的國賠事件，賠償金額共為一億一仟三百多萬元。施工品質不良除了賠上額外的經費，更進一步危害用路人的安全。本研究利用固定在機車置物箱內的行動裝置，收集三軸加速度器在不同路面的資料，進而分析加速度變化與路面狀況的關係。 本研究的目的在於偵測異常路面及評估路段品質。我們收集了騎乘機車時的加速度資料，共十二個路段，三個小時，約六十公里，並利用監督式(supervised)及非監督式(unsupervised)兩種機器學習的方法評估路面狀況。監督式的機器學習利用已標記的資料，嘗試辨認某個位置是否為異常路面，由此方法得到 78.5% 的異常路面辨識率(precision)。非監督式的機器學習利用分群(clustering)及學習門檻值(threshold)，找出平穩路面的振動模型。實驗最後，以上述兩種方法評估路段狀況，進而建立起一個道路品質的地圖。

Maintaining the quality of roadways is a major challenge for governments around the world. Poor road surfaces pose significant safety threats to drivers and motorists. According to the statistics of the Ministry of Justice in Taiwan, there are 220 claims for state compensation caused by road quality problems from 2005 to 2007, and the government paid a total of 113 million NTD in compensation. This research explores utilizing a mobile phone with tri-axial accelerometer to collect acceleration data while riding in the motorcycle. The data is analyzed to detect road anomaly and to evaluate the quality of the road segments. Acceleration data on motorcycles are collected on twelve road segments, three hours long, with a total length of about 60 kilometers in our experiments. Both supervised and unsupervised machine learning methods are used to recognize the road condition. SVM learning is used to detect road anomaly and to identify its corresponding position from labeled acceleration data. This method achieves a precision of 78.5% in road anomaly detection. To construct a model of smooth roads, unsupervised learning is used to learn the thresholds by clustering data collected from the accelerometer. The results are used to rank the quality of multiple road segments. We compare the rank list from the evaluator with the rank list from human testers who rode on the roads segments. The experiment showed that the automatic rank result is good based on the Kendall tau rank correlation coefficient.