摘 要: 運(yùn)動(dòng)目標(biāo)檢測(cè)是計(jì)算機(jī)視覺中重要的課題,運(yùn)動(dòng)目標(biāo)的正確檢測(cè)與正確分割影響著后續(xù)目標(biāo)的跟蹤與識(shí)別;光流法是運(yùn)動(dòng)目標(biāo)檢測(cè)和分析的重要方法,它能夠在不知道任何預(yù)先場(chǎng)景情況下檢測(cè)出獨(dú)立的運(yùn)動(dòng)目標(biāo),并且可適用于動(dòng)態(tài)場(chǎng)景的情況。首先介紹了光流的基本概念,然后介紹了常用的光流的四種算法;接著以智能交通中路口車輛視頻為例,將這四種光流算法用于車輛檢測(cè),然后對(duì)四種光流算法的優(yōu)缺點(diǎn)進(jìn)行分析;最后對(duì)光流法在未來可能研究及改進(jìn)的方向提出展望。
關(guān)鍵詞: 光流算法; 車輛檢測(cè); 智能交通; 計(jì)算機(jī)視覺
中圖分類號(hào): TN911.7?34 文獻(xiàn)標(biāo)識(shí)碼: A 文章編號(hào): 1004?373X(2013)13?0039?04
Comparative analysis of optical flow algorithms
ZHANG Jia?wei1, ZHI Rui?feng2
(1. North China University of Technology, Beijing 100144, China; 2. Tianshui Power Supply Corporation, Tianshui 741000, China)
Abstract: Moving object detection is currently one of the most active subjects in the domain of computer vision. Moving object detection and segmentation influence tracking and classification of the follow?up objects. Optical flow algorithm is an important method for moving object detection and analysis. It can detect the independent moving object in unknown scene and dynamic scene. The basic concept of optical flow is introduced in this paper at first and then the four common optical flow algorithms, which are applied to detection of vehicles at the crossroads in intelligent transportation video. The advantages and disadvantages of the four optical flow algorithms are analyzed while possible researches and modified direction prospect of the optical flow algorithms are proposed at the end.
Keywords: optical flow algorithm; vehicle detection; intelligent transportation; computer vision
0 引 言
基于視頻的運(yùn)動(dòng)目標(biāo)檢測(cè)的目的就是要在序列圖像中將運(yùn)動(dòng)目標(biāo)從場(chǎng)景中提取出來。但是由于光照的影響、風(fēng)吹、樹葉擺動(dòng)、運(yùn)動(dòng)目標(biāo)陰影、攝像機(jī)抖動(dòng)以及運(yùn)動(dòng)目標(biāo)的遮擋現(xiàn)象給運(yùn)動(dòng)目標(biāo)的正確檢測(cè)造成了極大的困難。運(yùn)動(dòng)目標(biāo)能否正確檢測(cè)和分割影響著后續(xù)運(yùn)動(dòng)目標(biāo)能否正確跟蹤與識(shí)別,因此運(yùn)動(dòng)目標(biāo)檢測(cè)成了計(jì)算機(jī)視覺中的一項(xiàng)重要課題。傳統(tǒng)的運(yùn)動(dòng)目標(biāo)檢測(cè)方法有光流算法[1?2],幀間差分法[3],背景建模法[4]和運(yùn)動(dòng)能量法[5]。背景建模法通過建立背景模型,然后將當(dāng)前幀中每個(gè)像素點(diǎn)與背景模型進(jìn)行比較來確定背景圖像。但是背景往往會(huì)隨著時(shí)間的推移發(fā)生變化,需要時(shí)刻更新背景圖像,需要背景圖像自適應(yīng)的更新。幀間差分法可以適應(yīng)環(huán)境的動(dòng)態(tài)變化,可以實(shí)現(xiàn)運(yùn)動(dòng)目標(biāo)的實(shí)時(shí)檢測(cè),但檢測(cè)出的目標(biāo)存在空洞嚴(yán)重且不連續(xù)?!?br>