Research On 3D Feature Extraction And Positioning Technology For Underwater Target Based On Point Cloud Data

Underwater unmanned robots need accurate visual guidance information in the implementation of interactive tasks. Structured light measurement system, whose measuring accuracy is basically not affected by light refraction and attenuation effects, is widely used in 3D reconstruction of underwater target. However, the underwater unstructured scenarios, the overlap and occlusion between target and other underwater objects, the measurement data’s characteristics of single visual and scattered, make underwater target positioning that much more difficult. Aiming at the problems of underwater target positioning above, on the basis of the principle of Random Sample Consensus (RANSAC), a 3D Feature Extraction and Positioning Technology for underwater target is presented based on point cloud data. The research contents in this paper is listed below:1.Based on the surface feature and imaging edge characteristic of regular component model of underwater targets, the target model is described as a combination of imaging line feature and point cloud surface feature, and the combination of the two is used as the prior information for target positioning.2.The segmentation and edge extraction algorithm is operated on the target field image, and the field image is divided into several regions containing the target or not. According to the description of target line feature in prior information, characteristic line is detected in segmentation region in turn, and the region contains the descripted characteristic line is considered as target region.3. A 2D elliptic detection method is proposed based on RANSAC algorithm. According to the coefficient that evaluates elliptic integrity, the ellipse of image edge is extracted.4.According to the mapping relation between image and point cloud measured by the structured light measurement system, the data projected into the target area is extracted and characteristic surface described in accordance with the prior information is detected from these data. Then the approximation of target’s size and pose parameters is estimated and the points belong to target are collected.5.Using 3D surface detection method based on RANSAC algorithm, identification of regular surface in point cloud is performed, such as spherical surface, cylindrical surface, conical surface, cone surface and rectangular decent etc.6.As the model parameters of different surface is not uniform, with superquadrics as target component model, the different types of surfaces translate into superquadrics with consistent parameters.7.With the approximated size parameters and pose parameters as the initial value, a nonlinear objective function is established based on the standard equation of superquadrics. The target parameters are globally optimized by Levenberg-Marquardt (LM) algorithm, thus the positioning result of target is obtained.Neither does the proposed algorithm need to segment the point cloud of target scene, nor does it need target offline modeling. The goal of autonomous positioning can be achieved only by providing the line feature in target imaging edge and surface features itself contained. To verify the correctness of the proposed algorithm, some practical experiments are performed and the experimental results show:the proposed algorithm can yield accurate localization results, be effective for man-made object of unknown size, and has good suitability to underwater complex environment.