Image Segmentation Technique Based On 3D Camera

Image segmentation is the key research of the image processing field, meanwhile, it is also an important step before image recognition and image analysis, which can be widely used in the intelligent mobile robot obstacle avoidance and scene understanding. Image segmentation algorithm has an important prerequisite that camera has the ability to capture high-quality image data. With the research and development of three-dimensional visual sensor, the camera to obtain high-quality space scene color image information and depth information has become increasingly convenient. The three-dimensional image data acquired camera sensor combines with the basic algorithm for traditional two-dimensional color image segmentation, taking full advantage of the pixel included space coordinate point information of image data as well as the color space information of the corresponding point to segment the object of three-dimension scene. Not only has this approach better stability than the ordinary two-dimensional color image segmentation, but also has a better practical significance and engineering application.Since the image segmentation is studied in many ways, we have achieved great progress, but there are many technical problems of three-dimensional space division needs to break through, such as:the complexity and uncertainty of three-dimensional space. In this paper, using the Kinect camera to obtain three-dimensional scene information to research the 3D point cloud segmentation algorithm and corresponding optimization algorithm based on the normal vector and curvature characteristics. The main work is divided into the following parts.Firstly, obtaining color images and depth information by Kinect camera, the existence of black holes and the area of the noise at the depth information acquired by the camera, using the threshold can be settled is based on the save format of this article experimental data. After recovering the depth data, in order to smooth the hardware and the process of data transmission bringing noise, a variety of non-linear filtering methods are discussed and comparison, deciding to select the bilateral filtering as the the final way to smooth our data.Secondly, Build the kd-tree for the raw data acquired by the Kinect camera which was converted into a three-dimensional point cloud data after preprocessing, can take advantage of the information of three-dimensional points to form triangular facets to calculate the formation of corresponding normal vector and curvature feature, relying on the angle between the normal vector and curvature change in the threshold range, utilizing region growing algorithm to cluster these points with similar characteristics all together to complete the initial three-dimensional point cloud segmentation.Finally, Converting the color image data into HSI color space, and then synthesize point cloud format data including XYZHSI information with the depth information. The introduction of noise region determination principles for the determination region calculated corresponding average value out of the normal vector and color component, relying on the principle of the normal vector angle difference and the color difference between the mean area to determine regional integration or removed to obtain a final three-dimensional point cloud segmentation result.