使用Boosting式局部偵測器以適應多角度及部分遮蔽之車輛偵測與追蹤

Abstract

近年來，為了提高分析大量影像資料的效率與準確度，影像式的車輛偵測技術廣泛地應用在智慧型運輸系統中，且相關研究以及應用也越來越多並且受到重視。本研究的目標在於改善影響車流輛偵測的兩大要素，第一，因影像中車側角度不同而造成偵測失誤或遺漏的狀況。第二，各種部分車體遭遮蔽而造成偵測錯誤的狀況。 在本研究中，我們運用兩層式的AdaBoost架構來加強車側的偵測率，後續利用車輪偵測以及有效地追蹤系統來處理車輛遮蔽的問題。由於AdaBoost是一種將資料一分為二的分類器，所以當資料特徵過於分散時，AdaBoost的偵測效果則會變差；我們的方法則是在Training之前，先將車側的資料另外再分成高角度和低角度兩類，並且利用這兩類資料建立兩層式的AdaBoost架構，這樣的方式可以使偵測率有效的提升；為了處理部分車輛遮蔽的問題，我們另外使用AdaBoost來偵測車輪，並且針對各種不同的狀況，制定相對應的追蹤法則，利用我們所制定的追蹤方式，可以處理多數車輛遮蔽的問題。我們提出的方法適用於道路側向的監控環境，高速行進的車輛依舊能有效分辨，也因此可以大幅提高車輛計數的準確率。由此系統所得到的結果可以在車輛偵測的技術中，針對以機器學習為基礎的車輛之偵測系統所遭遇到的問題提供不同的思考模式。實驗的結果證實我們提出的系統可以處理一定程度的遮蔽狀況，並偵測出影像中不同的車輛，我們所實作的系統也能為後續處理，提供有用的資訊。

Visual-based vehicle detection techniques applied to Intelligent Transportation System (ITS) to improve the efficiency and precision of analyzing heavy video information have been studied for years. How to reform the errors and mistakes caused by the angle difference of vehicle’s side view and the vehicle occlusion is the target of this study. In the study, we use the two-layer AdaBoost system to improve the detection rate. After that, we utilize the wheel detection and tracking method to deal with the occlusion problem. AdaBoost is a classifier that separates the data into two. As a result, the detection rate will become worse if the data is too decentralized. Our method is to separate the vehicle’s side view data into high angle and low angle before training. And set up two-layer AdaBoost system to increase the detection rate. In order to handle the occlusion problem, we not only detect vehicle but also detect wheel. According to different detection result, we formulate the tracking rules. With the tracking rules, most occlusion problem can be solved. Our system is capable of vehicle detection with roadside camera and with high speed moving objects. And this study also provides a novel concept to solve the problem faced by classifier-based detection method. Experimental results proved the proposed system can detect different vehicles and handle partial vehicle occlusion. The implemented system also extracted useful traffic information that can be used for further processing.