基於自適應增強演算法的串聯式分類器及物件估測之車輛偵測與追蹤系統

Abstract

近年科技的發展人們對於車輛的需求越來越高，科技所帶來的方便但仍伴隨著危險性而產生，人們在行車的過程中往往一個不注意就引發交通事故，因此本論文發展出一套車輛偵測系統，希望透過本系統的警告能夠必免不必要的交通意外事故，不同於其他複雜的環境設置，對於內部軟體本論文只需要將事前所訓練好的資料庫輸入道系統端就能夠進行偵測，對於外部設定只要將網路攝影機架設到車向就能夠結合電腦就能夠即時車輛偵測，同時本論文也發展出一套人機介面，透過單一按鍵就能輕易控制系統去偵測車外環境資訊。 本論文主要分為三大部分:物件估測、車輛偵測與物件追蹤，第一部分物件估測透過事前決定的正負樣本，使用二進制範梯度(Binary Normal Gradient)形成特徵結合線性支持跡向量求得影像中物件的可能位置與其尺寸，第二部分車輛偵測本論文使用自適應性增益來訓練一組弱分類器，組成一組強分器，同時弱分類器是使用哈爾特徵的基本五個型態來描述事前已知的樣本，第三部分車輛偵測本論文使用粒子濾波器結合偵測到的物件顏色來描述並追蹤物件，最後本論文依據上架構實現出一套針對車輛的偵測與追蹤系統，所得的最高偵測率有94%以上而誤偵測率最低只有7%左右的數值。

In recent years, the traffic accidents happen because drivers are not concentrating on the front vehicles. This thesis develops a vehicle detection and tracking system which can avoid traffic accidents. Unlike other complex hardware or software system, this dissertation uses USB camera and computer to build a hardware architecture. The software architecture contains an offline training model. This paper has three part: object estimation, object detection and object tracking In first part, the object estimation calculates the object probable position and scale size in processing image by offline training model .The offline model is combined with positive samples and negatives samples which using the BING(Binary Normal Gradient) and support vector machine. In second part , the several weak classifiers use AdaBoost algorithm to estimate the cascade classifiers, and then the weak classifier is calculated by five kinds of Haar like features which are described with offline positive and negative samples. In third part, the detection object tracked by particle filter, the particle filter’s likelihood function is yielded with the object’s color features. Finally, our method is combined with three parts and applied to vehicle detection and tracking system. The thesis has the highest 94% detection rate and the lowest 7% false alarm in our experiments