High-Accuracy, High-Speed 3D Optical Sensing in Unstructured Environments

Jae-Sang Hyun, Purdue University

Abstract

Over the last few decades, as many companies have released low-cost commercialized 3D sensors, vision-based 3D sensing has been more accessible and ubiquitous. As a result, the range of applications for 3D-sensing technology has been extended to medicine, entertainment, and manufacturing, as well as other industries. However, unlike with well-controlled industries such as manufacturing factories, commercial sensors and resolutions are not yet accurate enough to be applied in unstructured environments, such as construction sites. For example, to inspect the inside of large infrastructures such as steel bridges, robots need high-accuracy 3D maps for inspection and path planning, and robot sensors should be robust enough to withstand harsh weather. To achieve the goal of scanning and inspecting surrounding environments, the 3D imaging system needs to reconstruct 3D images with high accuracy, high speed, and robustness to noise. The first challenge in realizing a high-accuracy 3D imaging system in unstructured environments is noise in captured images. To improve the robustness of 3D images, we developed a computational framework by using geometric constraints for highaccuracy 3D sensing with only two-frequency patterns. A previously existing twofrequency phase unwrapping method has a limitation in accuracy because the scaling factor, which is calculated by the difference in fringe width between low-frequency and high-frequency patterns, significantly amplifies the noise signal. The framework suggested to use the relationship of optical devices for 3D sensing inversely. We can dramatically decrease the scaling factor required to reconstruct 3D images. Without additional patterns, we can measure the geometry of objects within a certain depth range accurately. The second challenge is mainly caused by the dynamic motion of moving platforms. If the sampling rate of 3D sensing is low, it is difficult for robots to localize the platform, generate 3D maps for surrounding environment, and make a right decision in planning a path or inspecting sites based on the information. To increase the speed of 3D sensing, we can reduce the number of patterns used for generating one 3D image. The number of patterns is an important factor in determining the speed of 3D reconstruction because a camera captures the patterns sequentially, which means that the number of patterns is proportional to the time taken to capture a set of images for one 3D image. We developed a method to reduce the number of patterns by using geometric constraints. In addition, by integrating texture image of the object with a phase-coding method, we used a total of five binary patterns to get absolute phase map for 3D reconstruction. By doing experiments with a high-speed camera, the sensing system captures 2D images at 3,333 Hz, and 3D images at 667 Hz. Although the speed of 3D sensing has increased through reducing the number of patterns, the system has fundamental limitations in speed and spectrum of light. The system typically includes at least one Digital Light Processing (DLP) projector because of its accuracy and flexibility. However, the mechanism of the DLP projector, which flips a set of micro mirrors inside the projector for determining whether each pixel is turned on or off, slows down the speed of 3D sensing.

Degree

Ph.D.

Advisors

Zhang, Purdue University.

Subject Area

Electrical engineering

COinS