Object Location Prediction Based on Motion Estimation with Application on Color-Based Foreground Object Detection

在很多電腦視覺以及機器視覺的應用當中，一些前景偵測的方法經常被使用來做為前置處理。影像之中有許多的特性被用來作為分辨前景以及背景的依據，其中物體的動態資訊可以有效的分辨出移動中的物體。 雖然動態資訊非常的有效，但是取得整張影像的動態資訊需要非常大的計算量。在過去，只有一些快速的演算法用應用在動態估測的處理上面來增加估測的速度。事實上，並非整張影像裡面的所有動態都非常重要，只有前景物體的動態才真正是被前景偵測系統所需要，而背景的動態則並非必要的，這意味著動態估測的處理不需要實行在背景的區域。 這個研究提出了一個在影片中預測前景物體位置的方法。這個方法同時使用了動態以及交通密度來預測前景物體在影像中的位置。其中物體的動態可由動態估測獲得而交通密度則可以由過去的判斷結果得到。 一個前景偵測的程式被設計來驗證這個預測方法。對於移動以及大小改變的物體的預測能力將藉由一些特殊的影片來解釋。最後，使用使用預測方法的優勢將藉由三個不同的輸入影片加以說明。Many computer vision and machine vision applications employ some foreground detection methods as the first stage for detecting object location. Many characteristics of image data have been used to segment images into background and foreground elements. Motion is effective information for detecting moving objects in two continuous images. Although motion is helpful to detect foreground objects, it requires a heavy computational load when detecting all motions of an image. In previous applications, some fast search algorithms are proposed to reduce the computational load of motion estimation. In fact, not all motions are important in an image. Only the foreground object motion is required in a foreground detection system. The background motion is not necessary for detecting foreground, and it means that the motion estimation process has no need to be applied in the background area. In this research, a method is proposed for predicting foreground object location in a video. The method uses both motion and traffic density to predict object location in an input image. Motion is obtained by motion estimation, and traffic density is obtained by the analysis of historical detection results. A program of foreground detection is designed to verify the prediction method. The prediction capabilities with moving and size-changing object are explained by the experiments with some special videos. Finally, the advantages of using the prediction method are illustrated through the experiment with three different input videos.摘要 I ABSTRACT III CONTENTS V LIST OF FIGURES IX LIST OF TABLE XVII CHAPTER 1 1 INTRODUCTION 1 1.1 Motivation 1 1.2 Fundamental Components of Foreground Detection 3 CHAPTER 2 9 LITERATURE SURVEY 9 2.1 Models and Feature for Building Background 10 2.2 Motion Estimation 11 2.3 Per Pixel Model and Region-based Model 12 2.4 Applications of Foreground Detection 12 CHAPTER 3 15 FUNDAMENTAL KNOWLEDGE OF IMAGE ANALYSIS AND PATTERN RECOGNITION 15 3.1 Image Acquisition 16 3.1.1 Digital Image Format 16 3.1.2 Spatial Filtering 17 3.1.3 HSI Color Space 20 3.2 Maximum Likelihood Estimation 21 3.3 Morphological Image Process 24 3.3.1 Dilation 25 3.3.2 Erosion 27 3.3.3 Opening and Closing 29 CHAPTER 4 33 OBJECT LOCATION PREDICTION IN VIDEO FLOW 33 4.1 Predict Method Based on Object Motion 34 4.1.1 Motion Estimation 35 4.1.2 Adaptive Prediction Block Size 43 4.2 Predict Method Based on Traffic Density 45 4.3 Combination of Motion Estimation and Traffic Density Map 47 CHAPTER 5 51 EXPERIMENTAL RESULTS OF FOREGROUND DETECTION SYSTEM 51 5.1 Experimental Environment and Framework of the System 51 5.2 The Classifier Used in the Experiment 59 5.3 Experimental Results of Object Location Prediction 62 5.4 Morphological Image Processing Results in the Experiments 72 5.5 The Experimental Results of Foreground Segmentation 73 CHAPTER 6 87 CONCLUSION AND FUTURE WORK 87 6.1 Conclusion 87 6.2 Future Work 88 REFERENCES 8