Research On Visual Perception And Reconstruction Of Regular Object In Indoor Scene For Robots

In recent years,with the continuous development of the field of robotics,robots have been widely used in all aspects of production and life.Accurate scene map is the guarantee of effective interaction between robot and environment,and scene map construction has become a hot issue.At present,SLAM technology has been generally applied in scene map construction and has made some progress.However,the application of regular object information in the scene is insufficient,which results in the scene map constructed without semantic information,the map accuracy is not high,and there are problems such as the lack of depth information and the poor closed-loop effect.In this thesis,an indoor scene semantic map construction method is designed that integrates regular object information and RGB-D SLAM.The main research work of this thesis is as follows:(1)Deep learning and traditional methods are integrated to complete the regular object detection and edge extraction.Firstly,the COCO dataset is used to train the YOLO object detection model to obtain a deep learning model for detecting the regular object.The trained model is used to detect the picture containing the regular object.Then,the image is segmented into images containing only a single object by using the detection results,and the edge was extracted by HED network.Finally,it combined Canny edge detection for optimization.(2)Regular object modeling is accomplished by edge information and geometric constraint of regular object in single image.Mainly realized the reconstruction of cuboid,revolving object,sphere,generalized cylinder and generalized cuboid.The modeling method is simple and efficient,which can complete the modeling quickly.For cuboid reconstruction,the vertex coordinates are solved with the two-dimensional detection frame and vanishing point obtained by edge detection,and the three-dimensional coordinates are solved with the camera parameters.For the revolving object,combined with the constraints of the revolving object after imaging,the ellipses formed on the upper and lower bottom surfaces of the revolving object are fitted.Based on this,the curve and symmetry axis are gradually solved and corrected.For the sphere,the three points on the projection circle are selected as a prior in the projected image to find the center and radius of the circle,and then the camera center is used to find the center and radius of the sphere.For the generalized cylinder and generalized cuboid,Using the edge contour information to obtain the profile and side contour,calculate the axis of symmetry,and then expand the profile along the axis of symmetry to complete the modeling.(3)The regular object information is fused to construct the scene map.Firstly,general framework of scene map construction based on regular object information is designed.Secondly,for the fusion of regular objects and SLAM maps,an initialization method for regular object scale and pose estimation is proposed.Based on the initial pose,the depth data is used to optimize the regular object scale and pose.Finally,a complete scene semantic map combining the regular object information is constructed.Through experiments and analysis,it is shown that the proposed method for constructing a scene map that incorporates regular object information can be used in indoor scenes.The constructed map contains the semantic information and shape information of the specific regular object.Compared with traditional point clouds,the map contains semantic information of regular object.The map is more complete and suitable for more demanding application scenarios.It has certain practical value in the indoor scene service and perception environment of the robot.