Research On 3D Object Recognition Algorithm Based On Realsense Captured Image Information

3D object recognition is one of the basic research fields in computer vision.It is also one of the hotspots and difficult points of various research institutions.It is widely used in the fields of immersion virtual interaction,target intelligent monitoring,object automatic assembly,mobile manipulation and robot.RealSense is Intel's intelligent depth camera,compared to the existing intelligent depth camera,with a better depth sensor and 1080 p color sensor,more adapt to the dynamic scene,in addition,The camera itself is very thin,in the robot,Internet of Things,virtual interaction and other fields will have a huge application prospects.Therefore,this paper designs and implements the RealSense grabber and management system,realizes and improves the 3D object recognition algorithm based on this system and PCL,with a view to the future application of mobile platforms such as robot with RealSense camera.First,this paper introduces the 2D image and 3D point coordinates of the camera model,and briefly describes RealSense F200 and RealSense R200,and then introduces the most widely used 3D image processing open source library PCL.Due to that RealSense is not supported by the whole platform PCL,which greatly limits the development and the use of researchers,this paper designed the RealSense grabber and management system.The entire system is divided into RealSense grabber and RealSense device manager,RealSense device manager is responsible for managing connected RealSense devices,and abstract the device for RealSense grabber.The RealSense grabber is responsible for providing the interface to the user,and completes the conversion of the original image to the point cloud format data.The system support different RealSense low-level IO library and multi-device access at the same time,and has an easy-to-use user interface,besides,multi-thread is used to ensure the performance of the program.Second,This paper introduces the 3D object recognition algorithm based on RealSense image information,and gives the algorithm and parameter selection for every link in the recognition algorithm.The algorithm is based on the RealSense image information,which captured by RealSense Grabber and management system.Then,this paper analyzes the characteristics of captured image points,and adopts the method of comprehensive filtering to reduce the non-target information points for the problem that the image information contains too many non-target information points.According to the characteristics of target object color space,this paper presents an improved fast point feature histogram descriptor extraction algorithm based on the existing fast point feature histogram descriptor extraction algorithm.The experimental results show that the integrated filter can effectively improve the proportion of interest information points in the image,and it can lay a solid foundation for 3D object recognition.At the same time,the improved fast point feature histogram descriptor extraction algorithm has achieved good results in real scene,compared with the algorithm with the best balance between accuracy and real-time performance,the real-time performance is improved by 27% when the accuracy is only about 3% decrease.Compared with the 3D object recognition algorithm which based on FPFH,the real-time performance of the improved algorithm is 2.45 times and the recognition rate is improved around 27.78%.