| Assembly as an essential part of the production chain occupies the most time of the entire product processing. Small batch and diversified production mode put forward higher requirements for assembly in rapid, accuracy and flexibility. Compared to the traditional manual assembly, the use of automated robot can eliminate the impact of human factors on product quality and reduce assembly time consuming, which can greatly improve production efficiency.Research on robot bin-picking is focused on recognizing and estimating the pose of objects. A practical solutions based on a low-cost Kinect is put forward to acquire pose of scattered parts in this paper. Point cloud of targets is access by analyzing and recovering various defects of the depth image output by the sensor. Statistical differences of boundary point in space and planar structure features are introduced to complete segmentation of the unorganized objects. In accordance with random sampling consistent principle, a matching algorithm based on point-pair feature is designed and experiments are carried out to verify the pose estimated by our method. Details are summarized as follows.Firstly, a scattered parts point cloud acquisition method is proposed based on Kinect sensor. The transformation relationship between depth images with the three-dimensional point cloud is establishment. Different methods are proposed for processing the three major defects in depth image. Use Guide Filter to remove noise in the image, the impact of regularization parameter and the window radius on the quality of image and boundary contours is discussed and we also present a method to selecting the optimal parameters for different images. Use Joint Bilateral Filtering to fill hollows in depth image, the influence of parameter on the filter is discussed and experiments are carried out to test the quality and quantity of this method; Use Kalman Filter to eliminate the jitter of depth data, the influence of parameter on the filter is discussed and the variance of filtered data compared to the original is tested. Finally, a complete depth image preprocessing is presented and the accuracy of the obtained point cloud is estimated by a series of tests.Secondly, a completely point cloud preprocessing scheme is designed. Obtain the full target point cloud by calibrating and removing the stage. A method based on voxel space is carried out to lower the density of cloud and decrease the total. By using local statistical features of the data points, an effective discrete point detection method is designed to improve the percentage of effective points in the target point cloud.Thirdly, a scattered parts segmentation method based on statistical distance and planar feature is proposed. Utilizing the differences between border points and interior points on the statistical properties in space, an effective method is proposed to eliminate the connection portion of adjacent target. Thereby obtain point cloud data suitable for spatial clustering. For adjacent target with larger interconnected level, this paper propose a geometry feature based segmentation algorithm, which use planar structure represents the target part. By matching with reference CAD model, single target can be separated completely.Finally, a pose estimation algorithm based on random sample consistent is proposed. The reference CAD model is acquired from STL model by using Halton sequence to generate uniform sampling points. Point pair feature is applied to accomplish matching between target and reference model. We propose a robust normal estimation algorithm based on statistical distance, which is used to compose the feature. Each part of the match scheme is accomplished and discussed according to random sample consistent principle. By introducing Td test and HASH tables, the computation time is greatly reduced. Finally an experimental platform is built and the pose of targets is calculated using method mentioned above, meanwhile, comparing to the actual test data to verify the accuracy of pose estimation. |
PDF Full Text Request