Abstract

Three-dimensional information is recently used in various fields such as autonomous driving, robots, and augmented reality because it has advantages in analyzing the location and appearance of objects. The representative technology for obtaining three-dimensional information is the time-of-flight (ToF) method that measures the distance using the fired laser signal and the laser signal reflected from the object. This method has the advantage of simple system configuration and good measurement accuracy. However, compared to other methods, the resolution of the obtained three-dimensional image is low, and the intensity of the reflected signal is a problem, so there may be abnormal pixels with wrong depth values. Due to these disadvantages, since it is difficult to use it efficiently compared to general RGB images, a super-resolution algorithm that increases the resolution and improves the image is required. In this paper, we propose a super-resolution algorithm of a depth image for a time-of-flight measuring camera using a single RGB image depth estimation method. And by using a machine learning algorithm, abnormal pixels with large differences in depth values are improved. This algorithm is expected to be used in intelligent systems using three-dimensional information in the future.