結合獨立成份與色彩特徵之平均移動向量人形 追蹤演算法-應用於主動式攝影機

Abstract

近幾年來，基於視覺的偵測和追蹤在電腦視覺領域上是一項很重要的課題，基於視覺的監控系統已經被廣泛的應用在停車，病人監控以及安全監控...等領域上。在本論文當中實現了在主動式攝影機上完成人物偵測和追蹤，在過去主動式攝影機追蹤是利用影像相減的方式找出移動物體的位置，透過移動物體的位置來驅動攝影機的雲台，雖然這樣的方法可以對移動物體作追蹤但是為了要找到移動物體的位置攝影機在移動的過程中必須停下來才可以做影像相減，所以會造成攝影機無法連續的控制雲台，換句話說，假如雲台持續的轉動，因攝影機的轉動會得到模糊的影像而且相減時不僅會得到移動物體也會得到包含背景的資訊。針對利用相減的方法無法找到移動物體精確的位置，因此我們提出結合獨立成份與色彩特徵之平均移動向量人形追蹤演算法應用於主動式攝影機來解決上述的問題。 本論文主要分成三大部分，分別是人物偵測、追蹤和攝影機的控制。利用人物偵測系統獨立成分分析和支持向量機(Support Vector Machines)分類現在畫面中的移動物體是人還是非人。當我們系統偵測到人之後就會被鎖定，利用平均移動向量演算法在每一張輸入的影像計算出相似度並且送出控制命令給主動式攝影機，使畫面中的人可以保持在我們所監控的影像中。有時候人會整個或部分的被其他物體遮蔽，因此相似度會劇烈的下降，使平均移動向量在追蹤的時候無法追蹤目標物，為了解決遮蔽的情況我們採用卡爾曼濾波器對目標物做位置的預測。 當目標物和背景有相同顏色的情況下只使用顏色當作平均移動向量追蹤特徵會有遺失的現象。為了解決遺失的現象，我們提出一個結合獨立成分分析和顏色當特徵的先進平均移動向量演算。因為獨立成分模組在訓練時所輸入的灰階影像具有人物的特性，因此我們提出來的驗算法可以有效在相同顏色下判斷人或背景。

In recent years, detection and tracking are important tasks in computer vision for visual-based surveillance system. Visual-based surveillance system is a widespread application in parking, patient monitoring and security surveillance fields. In this thesis, we use human detection and tracking algorithm based on active camera. In the past, active camera based object tracking used temporal difference to find object position and then drive pan or tilt command to control the active camera. Although this process can achieve moving object tracking. However, to find moving object position, the active camera should be stopped for the computation of temporal difference. Therefore, the active camera can not pan/tilt continuously and smoothly. In the other words, if the active camera is able to keep moving the whole time, we will capture blur images, and temporal difference will extract not only moving object but also background. Therefore, it is impossible to accurately locate the position of moving object by using temporal difference while the active camera is moving. So we propose mean-shift human tracking based on combination of ICA and color feature on active camera to solve above problem. This thesis consists of three major parts: Human detection, human tracking and pan/tilt control. In human detection system, the independent component analysis (ICA) and support machine vector (SVM) classifier are applied to classify moving objects into human or non-human. When a human is detected then we need to track it. Mean-shift algorithm will track the target by computing the similarity value in every frame and send the position to active camera, and then active camera will drive PTZ to keep the target in the center of FOV (field of view). Sometimes human will be partially or fully occluded by other object, thus the similarity will drastically decrease. Consequently, mean-shift will miss the target. To overcome the above problem, the Kalman filter is applied to predict the target’s position in next frame. The mean-shift using only color feature will miss the tracking target, when the target object and background have the same color. In order to solve the missing problem, we propose a novel mean-shift algorithm based on combination of ICA and color feature. Since the ICAs modeled human characteristic with gray-level training data, the proposed algorithm can distinguish human and background with the same color.