These days RGB-D cameras, especially Kinect (introduced by Microsoft in 2010), is providing depth map besides the color image of the capturing point of view by triangulating specific infrared patterns [FrEtAl13]. This new feature is beneficial for wide number of problems in the area of Computer Vision, especially for mobile robots to understand the scene and improve their knowledge about its geometry. To create an accurate road map from the input RGB-D data collected by the mobile robots, a significant constrain is to have an accurate depth map which helps to have a better understanding of the desired scene. Having an accurate depth map as input is also an important point in wide number of other problems [ZhEtAl17, ChEtAl16]. Captured depth map using Kinect sensor suffers from both holes and invalid measurements called noise. Holes are the pixels that depth sensor was unable to compute any depth value for them; because of the lighting conditions or being a glass or mirror in front of the IR camera. Invalid measurements which are mostly called as noise in the literature are also involved in the captured depth map due to the lightning condition, the way that the IR pattern is reflecting to the camera, the properties of the object surface that IR pattern is facing with, and finally lacking in calibration and measurement of disparities. It is also important to notify that the value of noise increases according to the distance exponentially (Figure 6.1).

Keywords