采用 SOM 算法的软体机械臂三维形状实时感知

To realize the accurate 3D shape real-time estimation of soft manipulator and lay the foundation for deformation control and application, a real-time 3D shape recognition method based on the SOM algorithm is proposed for three-section soft manipulator. Firstly, the left and right frame data captured by ZED binocular camera are preprocessed to obtain the left and right binary images, with the 2D contour data of the soft manipulator extracted in real time. Then, the self-organizing map (SOM) algorithm is run to cluster the contour data to obtain the 2D centerline of the soft manipulator in order, and compared with K-means, Gaussian mixture model and thinning of three centerline extraction algorithms, revealing that SOM algorithm is more suitable for solving the centerline identification of complex shape of the soft manipulator. Finally, the 3D shape reconstruction of the soft manipulator is completed by using the triangulation model based on the disparity. Furthermore, the algorithm framework adapts data downsampling, SOM parameter optimization and other methods to improve real-time performance. The real-time shape sensing and comparative verification are carried out during the continuous deformation of the soft manipulator. Experimental results show that the algorithm presents a high shape sensing accuracy and a superior real-time tracking effect. Compared with other shape detection algorithms locally and globally, the proposed algorithm displays a better performance.