Research And Improvement On Real-time Performance Of Recognition Algorithm Based On Eye-in-hand System

This paper is based on existing eye-in-hand system architecture and vision recognition algorithm.We focus on real-time performance of 3d object recognition and grasping,including: recognition speed,success rate of pose estimation,and stability of automatic observation and grasping.Based on the research and analysis of existing algorithms,we propose some methods to improve its real-time performance.The main contents of this paper are as follows:(1)We propose an icosahedron recursive uniform template distribution strategy based on LINEMOD algorithm.This strategy utilizes distribution of polyhedron vertexes to fit camera viewpoints,and each layer of new viewpoints is generated on the last layer through bisecting triangle.With simulation and comparison,under the same experiment scenario,the proposed template generating strategy greatly improved the success rate of recognition algorithm while ensuring the matching score.(2)We propose a hierarchical template search strategy based on region matching.We firstly perform partial hierarchical search on r dimension.Then,we establish the template index structure based on region hierarchy,and use regional hierarchical search strategy to simplify the search process.After experiment and comparison,the search time of template matching was greatly reduced.(3)Integrating recognition algorithm with neural network segmentation.We propose to take the result of segmentation as the matching target of the template matching algorithm.In this way,searcharea of each image is reduced and matching speed is improved.Firstly,we use artificial image segmentation to verify the feasibility of the idea,then integrate FCN semantic segmentation with template matching algorithm.Finally,the recognition speed was improved within 100 ms per frame.(4)Designing experiments to demonstrate and compare the performance of eye-hand system.We establish the experimental framework through ROS,and add Moveit collision detection to ensure the safety and reliability.Based on established framework,we compared optimized recognition algorithm with the previous algorithm under real scenario.The experiment result shows that real-time performance of robotic observation and grasping is significantly improved by using established experimental framework and optimized strategies.