Understanding Scene Geometry and Semantics for Automation and Robotics in Civil Engineering
Understanding Scene Geometry and Semantics for Automation and Robotics in Civil Engineering
Dr. Chen Feng
Abstract:
The construction industry faces challenges including high workplace injuries and fatalities, stagnant productivity, and skill shortage. Automation and Robotics in Construction (ARC) has been proposed as a promising direction to make machinery easier to collaborate with and facilitate better decision-making. However, there are two primary technical challenges in ARC: 1) unstructured and dynamic environments; and 2) differences between the as-designed and the as-built. Therefore, it is impossible to naively replicate conventional automation methods in industries such as manufacturing on construction sites. In particular, understanding the geometry and semantics of jobsite scenes, are fundamental problems that must be addressed to realize the full potential of ARC.
Interestingly, autonomous driving, an active yet seemingly distant research topic, has many aspects in common with ARC, in terms of not only technical challenges but also practical applications. In this talk, I will share two of my recent works at MERL on those problems, demonstrating the intersections and interactions between civil engineering and key technologies in autonomous driving including robotics, computer vision, and machine learning.
The first work focuses on efficiently modeling 3D geometric structures of a scene (e.g., a tunnel), by proposing a new algorithm converting point clouds into T-spline models. In CAD/BIM design and reverse engineering, T-spline is an increasingly popular new 3D representation that is more concise than conventional NURBS or B-spline. This new patent-pending algorithm enables automatic 3D surface modeling from millions of points in seconds, which means 10 to 1000 times faster than conventional methods while achieving comparable or better fitting qualities. Applications of this method range from as-built infrastructure monitoring to map compression for autonomous driving and robotic object manipulation.
The second work focuses on intelligently parsing and understanding semantic information from images, by designing a novel deep active learning framework for largely boosting civil infrastructure inspection and maintenance efficiency through automatic defect detection and classification in surface images. Conventional methods require a large number of images with defect labels to train classifiers such as support vector machines. However, in real world applications, labels are harder to obtain than images, due to the limited labeling resources (i.e., experts). The new framework takes advantage of a state-of-the-art deep convolutional network as the classifier and proposes a novel active learning strategy for its training. This allows the system to automatically select a most informative subset of unlabeled images and query labels from human experts to most economically improve the classifier (saving 30% labeled images than baseline methods while preserving accuracy).
Bio:
Dr. Chen Feng earned his Ph.D. in Civil Engineering in 2015 from the University of Michigan at Ann Arbor, along with a master in Construction Engineering and Management and a master in Electrical Engineering. Before that, he earned his Bachelor degree in Geospatial Engineering from Wuhan University in China.