Object Detection Using Fuzzy Classifier and Color/Depth Features

This thesis applies a fuzzy classifier (FC) to detect objects using color and depth features in color images captured from a Kinect RGB-depth (RGBD) camera. Appearances of the detected objects are assumed to contain multiple colors in non-homogeneous distributions. The FC uses a self-splitting clustering (SSC) algorithm for structure learning. Antecedent and consequent part parameters in the FC are learned through margin-selective gradient descent (MSGD) and support vector machine (SVM), respectively, to endow the FC with high generalization ability. For the problem of detecting objects in a color image, this thesis proposes a new detection method in which color features are extracted according to the color components of an object and their geometrical distributions. The feature of color components consists of two parts. One is pixel distribution and the other is color entropy. By computing the distance between pixels and each cluster generated from the SSC clustering in HS space, the pixels are assigned to the closest cluster. After all the pixels are labeled, the percentage of the number of pixels assigned to each cluster is obtained. The color entropy is obtained from the histograms of the pixels in the clustered HS space. The percentage values and the color entropy are fed as inputs to an FC. The filtered objects are then sent to another FC with entropies of color geometrical distributions to generate the candidate objects. After the detection by using the color feature, the depth information available from the camera is used to extract the shape of the object for a final decision. A histogram-based shape feature is used to filter the candidate objects. The performance of the proposed object detection method is verified through the detection of different objects and comparisons with various fuzzy classification and detection approaches.本論文使用一種模糊分類器（FC），並使用Kinect彩色深度攝影機在物體偵測中使用顏色及深度特徵。偵測的物體設定為表面具多種顏色，且不均勻分布。此模糊分類器（FC）在架構學習的部分使用自我分裂分群（SSC）的分群演算法。前件部學習使用選擇性梯度下降（MSGD），在後件部則使用支持向量機（SVM）學習。 而在彩色影像中偵測物體的問題，本論文提出一個新的物體偵測方法，其根據物體的顏色組成和顏色的幾何分佈，做特徵擷取後進而做物體辨識。物體的顏色組成可分為兩部分，其一是熵值色彩特徵，二是像素分布比例。像素分布比例是藉由計算每個像素在色度飽和度空間中和由SSC分群所產生的群中心的距離，並且標籤此像素屬使其於距離最近的那群，再計算每群中像素的數目的百分比。最後將這些百分比及熵值當作特徵值送入兩階段的模糊分類器去偵測物體。經過篩選的物體會再送入另一個模糊分類器 (FC)，並使用顏色幾何分部的特徵產生候選物體。在使用顏色特徵做偵測後，藉由從攝影機取得的深度資訊，被用來萃取出最後決定的物體形狀。以直方圖分佈為基礎的形狀特徵被用來擷取候選物體的形狀，可用來改善物體偵測的效能。在實驗結果中，本論文比較不同的分類器和物體辨識方法，以驗證所提出的分類器和物體檢測方法的優勢。摘 要 i Abstract ii Content iii List of Tables v List of Figures vi Chapter 1 Introduction 1 1.1. Survey and Literature Review 1 1.1.1 Color-based Object Detection 2 1.1.2 Object Shape Extraction Using a Kinect Camera 2 1.2 Thesis Organization 3 Chapter 2 Fuzzy Classifier 4 2.1. Fuzzy Classifier Rules 4 2.2. SSC-based Structure Learning 5 Chapter 3 Parameter Learning 8 3.1. SVM-based Consequent Parameter Learning 8 3.2. MSGD-based Antecedent Parameter Learning 10 3.3. Overall Parameter Learning Algorithm 12 Chapter 4 Object Detection Using Pixel Distribution and Entropy 14 4.1. Pixel Distribution and Entropy of Color Components in Filtering Phase One 14 4.1.1. Pixel distribution 15 4.1.2. Extraction of the ECC Feature 16 4.2. Entropies of Geometric Color Distributions in Filtering Phase Two 19 4.3 Two phases filter FC-SSCSVM-G process 21 Chapter 5 Object Detection Using Depth Features 23 5.1. Object Detection Process 23 5.2. Shape-based Detection by Kinect D Component 25 5.2.1. Shape Contour Histogram Feature 25 5.2.2. Object Shape Matching Method and Shape Detection 29 5.3. Color-based or Depth-based Detection Selection 32 Chapter 6 Experiment 39 6.1. Hardware and Software of the Object Detection System 39 6.2 Comparisons with Other Classifiers and Object Detection Methods 40 Experiment 1 . Can 44 Experiment 2 . Bottle 47 Experiment 3 . Star 50 Experiment 4 . Octagon 53 Chapter 7 Conclusion 56 References 5