應用電腦視覺處理混合車流之交通衝突分析

Abstract

機車於混合車流中之安全問題，主要由不同車種間複雜的互動行為所引起。汽車與機車由於構造及操作特性的差異，造就了彼此截然不同的駕駛行為。並且，近年來大部分車流分析之研究皆建立於車輛遵守道路標線行駛之前提，但是機車擁有與其假設相異之駕駛行為，而傳統以汽車為導向之交通理論與交通管理之方法無法適用於混合車流。本研究之目的為觀察混合車流之特性及駕駛行為與汽機車間之交通衝突。上述交通資訊之採集係使用無人飛行載具（Unmanned Aerial Vehicle，UAV）拍攝台北市區道路路口之交通影像進行觀察，並透過電腦視覺與影像處理等技術以擷取混合車流之微觀特性，如車輛種類、行車速度、加減速、行車軌跡等。車輛偵測部分採用方向梯度直方圖(Histograms of Oriented Gradient, HOG)描述子(descriptor)描述車輛特徵，並使用支撐向量機(Support vector machine, SVM)分類器完成車輛辨識作業，卡曼濾波(Kalman Filter)則將偵測結果連結完成車輛之軌跡。此外，交通衝突之嚴重程度由計算車輛間之碰撞時間(Time-to-collision, TTC) 以及安全空間(safety space)之被侵犯程度進行評估。本研究所提出之方法顯示了優異的穩定性與偵測表現，於汽機車之偵測上分別達到了98.3%與98.1%的精準度，並且即使於極度擁擠的交通情況下，依然能夠正確地進行車輛的分類與軌跡之追蹤。本研究所提出之結果，可做為道路交通安全分析之參考。

Safety issues of motorcycles are mainly caused by the complicated interactions among various vehicle types in the mixed traffic flow. The structural differences between automobiles and motorcycles results to distinct driving behavior. Moreover, most of the research in traffic flow analytics is carried out in the context of developed countries where vehicle follows lane markings. However, motorcycles have very different behaviors, and the traditional automobile-based traffic theory and transportation management cannot be applied to mixed traffic streams with automobiles and motorcycles. The purpose of this study is to observe the features of mixed traffic flow, driver behavior and traffic conflict between automobiles and motorcycles. The data was collected by using an unmanned aerial vehicle (UAV) at urban intersections in Taipei. The microscopic characteristics of mixed traffic flow such as vehicle types, velocity, acceleration, trajectories are observed through computer vision and image processing methods. The Histogram of Oriented Gradients (HOG) descriptor is adapted for the detection of vehicles utilizing a Support Vector Machine (SVM), and the Kalman Filter is employed for the tracking of the vehicles’ trajectory. The traffic conflict severity was conducted by calculating the time-to-collision (TTC) and invasion of safety space for vehicles. The detection results show superior stability and performance with the precision of 98.3% and 98.1% for automobiles and motorcycles, respectively. In addition, even under highly dense urban traffic conditions, vehicle classification and tracking are successful. The results of this study can serve as a reference for roadway safety guidance.