Research On 3D Dense Point Cloud Reconstruction And Target Recognition In Indoor Environments Based On Visual SLA

Visual sensor mapping and target recognition are the research hotspot of mobile robot.Visual SLAM(Simultaneous Localization and Mapping,SLAM)can sense the surrounding environment and reconstruct the dense map containing rich environmental information,which is a key technology for mobile robot to complete the scene understanding work such as finding objects.The 3D object recognition network can detect objects from the dense point cloud map,estimate the boundary box,spatial position and category information of the target,and provide the target information for the robot,which is the basis of path planning,emergency risk aversion and motion estimation of SLAM system.There are some shortcomings in using visual SLAM to reconstruct the 3D dense point cloud map and to make target recognition for the map: First,the traditional visual SLAM system can only build sparse point cloud map,which means the traditional visual SLAM algorithm system needs to be improved,and optimize the calculation complexity and accuracy;Second,the research on indoor 3D target recognition network is still in the initial stage,and the number of its studies is poor and the accuracy is low,so it is necessary to improve the indoor 3D target recognition network,promote the recognition accuracy of the network.In this thesis,in view of the above problems,the methods of conducting dense environment reconstruction based on the traditional ORB-SLAM2 algorithm and indoor 3D target identification on the maps are explored.The main research work is as follows:(1)In order to optimize the accuracy of the map,the feature extraction and matching algorithm of the ORB is improved,in which the FAST algorithm combined with SURF algorithm is used for feature extraction to ensure the scale invariance of feature points.Meanwhile the RANSAC algorithm is improved,through adding the pre-processing of data points,eliminating the redundant internal points,which reduces the number of iterations of the algorithm.(2)In order to reconstruct the dense point cloud map and optimize the computational complexity,this thesis adds a dense reconstruction thread on the ORB-SLAM2 infrastructure,and designs a key frame screening mechanism,through a series of constraints to select high quality key frame.After key frame screening,the dense point cloud map is reconstructed based on the key frame sequence.By experimental comparison,the algorithm performs better than ORB-SLAM2 in reconstruction accuracy and real-time performance.(3)Aiming at the characteristics of large data amount and sparsity of the reconstructed dense point cloud map,this thesis improves an indoor full convolutional3 D target recognition algorithm and designs a novel FCSRM＿3D network.Through improving the backbone network and residual network of the algorithm,the end-to-end point cloud 3D object identification is realized by using 3D sparse convolution instead of 2D convolution.The ablation experiment,the evaluation results of Dense＿mapping＿ORBSLAM2 point cloud map and the results of indoor actual environment on the ROS system show that our algorithm can accurately identify objects and perform better than other algorithms in accuracy.