CSCI498B/598B Human-Centered Robotics September 05, 2014 A Quick Summary • Definition of human-centered robotics (HCR) • Applications and challenges of HCR • Comparisons of HCR with conventional robotics • Sensors in Robotics • 3D Machine Vision • Kinect 2 What to perceive? • Humans 3 What to perceive? • Human activities 4 What to perceive? • Group behaviors 5 What to perceive? • Human-object interaction 6 What to perceive? • Objects 7 What to perceive? • Scenes 8 General Pipeline 1. Data acquisition (e.g., from 3D sensors) 2. Feature extraction and representation 3. Classification/clustering 4. Decision making 5. Taking an action 9 Data Acquisition from Kinect 1. Color-depth images (i.e., RGBD images) 2. 3D point clouds 10 Data Acquisition from Kinect 3D point clouds and color-depth images are not equivalent! • 3D point clouds contains more information then RGBD images or color-depth images 11 Feature Extraction • Definition: feature extraction is the process of defining a set of features or attributes, which will most efficiently or meaningfully represent the information that is important for classification, clustering, or other analysis • Goal: to improve the effectiveness and efficiency of analysis 12 Feature Extraction • Definition: feature extraction is the process of defining a set of features or attributes, which will most efficiently or meaningfully represent the information that is important for classification, clustering, or other analysis • eliminating redundancy in the data • eliminating variability that is of little, or of no value • restructuring data • Goal: To form a representation that • maximize pattern discrimination (effectiveness) • minimize the number of the features (efficiency) 13 Feature Extraction • Examples: skeleton of 3D models 14 Feature Extraction • Examples: surface norm from 3D point clouds 15 Feature Extraction • Examples: local features from interest points 16 Skeleton features from Kinect 17 Skeleton features from Kinect • Each joint has 3 values (joint world coordinates in meters) • Totally 60 elements in a feature vector • Considering orientations and other elements, we have a large feature with a lot of noise or irrelevant information 18 Skeleton-based features • Solution 1: Only use more descriptive joints 19 Skeleton-based features • Solution 1: Only use more descriptive joints 20 Skeleton-based features • Solution 2: Compute additional features 𝒅 𝜽 21 3D mapping from 2D Lasers http://www.youtube.com/watch?v=IMSozUpFFkU 22