Depth Recovery Method Based On Color Texture Information

Depth sensing is a fundamental task in the field of computational vision, relatingto many practical applications such as robot orientation,3D model reconstruction andhuman-computer interaction. Laser scanning requires slice by slice scanning, which istoo time-consuming and unfit for the dynamic scenes. Stereo matching fails in theocclusion region and the smooth region. TOF (Time-of-flight) camera is a kind ofimaging device which achieves active depth sensing. TOF camera provides reliablemeasurement within the range of a wavelength, compensating the drawbacks of laserscanning and stereo matching. However, TOF camera also exists some defects: lowresolution, high noise and inaccurate calibration, which severely limits its application.This paper focuses on research of depth super-resolution reconstruction for thedepth maps captured by TOF cameras. Main achievements are listed as follows.1. We propose a depth super-resolution method based on non-local meansfiltering. Referring to the algorithm of stereo matching, the cost function is formedusing the depth map, and the non-local means is chosen for cost aggregation. Underthe guidance of registered high resolution depth and color images, every slice of thecost volume is recomputed based on weighted summation.2. A non-local depth recovery method is proposed based on minimum spanningtree. A minimum spanning tree is derived from the depth and color guidance, and thepixel similarity criterion is defined on the tree structure. Every pixel contributes to theprediction of the other pixels. The experiment results show the superiority of ourmethod over the traditional local algorithms.3. Besides, we construct a stereo imaging system using a TOF camera and PointGrey camera. The depth images are super-resolved, and the virtual views are renderedwith depth and texture information, which are ready for3D display.