Sequential Scene parsing: Semantic Scene Approximation using Temporal and Physics Based Reasoning

Scene parsing;physics-based reasoning;pose estimation;Robotics

Jonathan, Felix

Johns Hopkins University

关键词: Scene parsing;    physics-based reasoning;    pose estimation;    Robotics;   

Others  :  https://jscholarship.library.jhu.edu/bitstream/handle/1774.2/44636/JONATHAN-THESIS-2017.pdf?sequence=1&isAllowed=y

瑞士|英语

来源: JOHNS HOPKINS DSpace Repository

【 摘 要 】

A consistent scene approximation is essential to a performance of a robotic task that involves object manipulation with random poses. Although many computer vision techniques exist for segmenting and computing poses of the objects in scene, these methods still generates inaccurate pose estimates when high amount of occlusions present in the scene. Additionally, since most techniques do not use previous scene results for constructing their current estimate, they can not generate a scene estimate that is stable over time. Building a time-conscious scene hypothesis requires an ability to extract new parameters that determine the time relationship between each scene.In this thesis, we propose an extension of the sequential scene parsing framework by Hager and Wegbreit for constructing a physically consistent scene from RGB-D data received in a discrete-time series. In contrast to the original system, which was limited to geometry-based primitives scene estimate, we utilize semantic segmentation and object pose estimate using an RANSAC-based approach. This framework derives those data into scene components that incorporate collision compliances, pose data fitness, support contribution, stabilities, and pose consistency for each object.Past attempts to tackle this problem, such as using SLAM techniques with geometry-based reasoning and physics, have utilized time variable to achieve acceptable scene modeling consistency; However, they do not estimate object poses in the scene. Other scene estimator algorithms that do not use SLAM usually only focus on creating consistent labeling of the current scene, which makes it lacks the time consistency for resolving object poses in scene. Many research have been done in accepting and rejecting a scene hypothesis based on a physical reasoning, but it does not utilize the data from the previous scene into consideration when constructing a new scene prediction.Experiments done in a simulated environment and real data of an increasingly complex structure has proved that our method was capable of building structures and improved both the accuracy and the precision of pose estimator. In a simulated environment, the average object orientation accuracy shows 40% error reduction, while position accuracy shows by 30% error reduction. The precision of object has improved by 45% in position and 75% in orientation.