應用分群法於眾源行車震動事件採探路面異常點

Abstract

此論文中我們提出了一個可透過智慧手機探偵車 (SPC) 所記錄的車輛震動事件群集找尋路面異常點的架構。我們設計了一個具有可適性的震動偵測和評估的演算法，可提供動態手機與車輛配對的能力，因此可大量佈建並使用在眾源資料收集，我們將相關技術實作成“Bumping Sensor”的行動感測APP，提供相關可靠的車輛震動事件偵測，用以評估車輛行經的路面異常點。由於來自群眾所收集的震動事件資料可能包含定位的誤差和錯誤的偵測等所造成的雜訊，因此我們運用Naive Grid-based Clustering (NGC), DENsitybased CLUstEring (DENCLUE), Adaptive Mesh Refinement (AMR) 等三個群聚演算法進行資料彙整並從群眾車輛所回傳的震動事件挖掘出道路表面的異常點。我們進行了兩個震動偵測實驗，分別在台灣交通大學校園和美國柏克萊加州大學裡收集到2059和641個震動事件。由NGC, DENCLUE和AMR三種分群演算法所挖掘出之震動事件正確率分別為64%, 82%, 70%。實驗結果顯示DENCLUE分群演算法相較於NGC和AMR具有更好的路面異常點分辨能力。

In this thesis, we develop a framework to discover road surface anomalies from vehicle bumping events reported by the Smartphone Probe Cars (SPC) crowd. An adaptive bumping detection and assessment algorithm is implemented to develop a mobile sensing App, called “Bumping Sensor”, that provides relatively reliable detection and assessment of the run-over road surface anomalies. Since the crowdsourced data may contain false alerts and be contaminated by imprecise position information. Three clustering algorithms, including Naive Grid-based Clustering (NGC), DENsitybased CLUstEring (DENCLUE), and Adaptive Mesh Refinement (AMR), are implemented to mine road surface anomalies from crowdsourced bumping events. Two experiments were executed on NCTU campus, Taiwan and Richmond Field Station, Berkeley, USA. In the two experiments, 2059 and 641 bumping events were collected respectively. The accuracy results in identification of road anomalies by NGC, DENCLUE and AMR clustering algorithms are 64%, 82%, and 70%. The results show that DENCLUE clustering algorithm outperforms NGC and AMR in identification of road anomalies.