基於外觀與路徑資訊之物件分群系統

Abstract

近來幾年，越來越多監視攝影機被廣泛地安裝在公共場所，以監控人們的行為，並應用在許多不同的方面。這些攝影機的拍攝時間通常都相當的長，因此，當特殊事件發生後，我們可能需要花許多時間來觀看監視影片並從冗長的影片中找出目標片段。傳統的做法是使用快速播放來節省影片觀看的時間，雖然利用快轉播放可以將影片的播放時間縮短，但同時也會加快目標的動作，可能使其不意被看清或者容易被忽略。 在本篇論文內，我們將影片內物件分類為三種:行人、機車、及汽車。取出一段時間的視訊間空影片，學習出整個場景的模型，基於這個模型，再對其他的物體加以分類。由於透視效應的關係，同一個物件在場景中的不同位置會有不同的資訊，例如大小、行進方向、及行進速度；而不同種類的物件在同一個位置，其長寬比及面積大小也會有不同的表現。基於這個原因，我們將整個場景切為 個方格，蒐集一段時間內出現在每一格子內，各自屬於那三種種類的資訊，每一種類再利用這些訊息去建立三種不同的模型。 當我們對監控影片分類完後，搭配一套時空快速影片檢索系統，來解決在監視影片中尋找目標的問題。讓使用者在快速觀看影片的同時也可以擁有足夠的時間仔細看清楚目標。我們的主要概念是取出並且追蹤影片中所有含有最多資訊的連續物件，並改變他們出現在監視影片中的時間軸位置，藉此來產生一個較短的精簡影片，而不是去改變物件本身的動作速度。若影片內容物太多，而使用者只想看某一種類時，我們的分類機制能讓使用者直接選取想要觀看的物件種類製作濃縮影片。

Recently, more and more surveillance cameras are widely installed in public place to monitor people's behavior for various applications. Since surveillance videos are very long, we need a lot of time to scan a video in order to find a specific target. A traditional approach is to use the fast-forward method to saving the scanning time, while this method can shorten the video playing time, but it will also accelerate the movement speed of the targets, and may leads the targets hard to be seen clearly even be ignored. In this approach, we classified the objects inside the surveillance video into three categories: pedestrians, scooters, and vehicles. A period of video was derived to construct a model of this scene. According the perspective, a same object will have different information at different position: object size, moving direction, and moving velocity. On the other hand, different objects at the same position will have different information either, such as aspect ratio and object size. Based on this reason, we divided the scene into blocks, accumulating data with each kind of category, and construct three different models for them. After the object classification, we develop a temporal-spatial quick browsing system for surveillance video to solve this problem. The user can not only fast browse surveillance videos but also clearly look at the target while retrieving targets. Our basic idea is first to extract all moving objects that carry the most significant information in a surveillance video, and then to rearrange their position on the time-axis of the video to short it. In addition, we try to preserve all the essential activities appearing in the original surveillance video. Also, user can query a certain kind of category they want to see in the condensate surveillance video.