Doctor of Philosophy, The Ohio State University, 2019, Civil Engineering
Learning the appropriate representation of objects in the real world is an ongoing area of research. Deep learning, specifically convolutional neural networks, have shown promising solutions for representation learning. In order to teach a machine to see and learn about a scene, a wide variety of information must be extracted from that scene. The extracted information can be used in many applications from medical imaging, customer behavior analysis, to worker productivity and safety in a construction site. This dissertation proposes models for extracting information from the scene through geometry and deep learning tools. We tackle the challenging problems of optimized camera placement, object contour detection and tracking, and propose a method to learn from unlabeled data.
Specifically, we design a system that resembles a human. This artificial human has eyes and a brain. We model the eyes of our system as visual sensors, which can be placed anywhere in the scene, e.g., a grocery store, construction site, or on a larger scale in a smart city for traffic systems. In order to configure a network of cameras, a graph formulation is proposed. As a novel constraint, material information has been added to the more conventional constraints, such as geometric and constructive constraints, which are required for maximum coverage and observability. Material information, along with the location of the light source, provides a new perspective in the camera configuration problem as the appearance of objects changes depending on the angle that light hits their surfaces.
Once the cameras are planned, the artificial brain in our system processes the sensor data streaming into the system. We use deep learning to detect, track, and locate objects in the scene. We use both tracking-by-detection and a novel temporally consistent detection and tracking algorithm. The work has been focused on learning new representations and domains in which objects live. Image-joint domain, for (open full item for complete abstract)
Committee: Alper Yilmaz (Advisor); Charles Toth (Committee Member); Rongjun Qin (Committee Member)
Subjects: Civil Engineering; Computer Engineering; Computer Science; Engineering