Real-time Shape Recognition of a Deformable Link by Using Self-Organizing Map

摘要

Here we present a novel deformable manipulator composed of active rigid joints and deformable links. The deformable link consists of several passive spherical joints articulated with each other with preload force between socket-ball surfaces. Therefore the manipulator can reach more parts of the task space compared with rigid-link manipulators by bending deformable links according to different tasks. However, frequent changes in the links' shape lead to unknown kinematic parameters, which bring difficulties to the planning and control of the manipulator. In this paper, a real-time shape recognition algorithm is proposed for the deformable link by using Self-Organizing Map (SOM). To avoid topological error and local convergence problem, Least Square Method (LSQ) is utilized for initialization according to the link's current position and shape. The reinitialization process is added to improve the robustness when facing noise and occlusion. To meet the demand of real-time tracking, the GPU parallel computation is applied for acceleration. Moreover, an error metric based on Signed Distance Function (SDF) is presented for evaluation. In this paper, our algorithm is implemented on a deformable link with 11 components. Experimental results validate the feasibility and effectiveness of this method. The processing time descends from 700 ms/frame to 40 ms/frame by using GPU, and the overall average of tracking errors is below 4 mm.