| The quadruped robot,which has excellent environmental adaptability,movement flexibility and unparalleled advantages on rough roads and complex environments,has become one of the research hotspots in the field of mobile robots.The environment perception technology represented by SLAM technology can give quadruped robots the ability to locate,map,and autonomously navigate in complex and unknown environments.However,the high dynamic characteristics of the quadruped robot itself,as well as the large scene and high complexity of the application environment,pose a huge challenge to the traditional laser SLAM technology,especially its key technology point cloud registration.In response to the above problems,this paper proposed a SLAM technology that combines deep semantic information to achieve precise positioning and mapping of quadruped robots in complex environments.The main contents are as follows:First,the statistical filtering method is used to preprocess the point cloud filtering on the three-dimensional point cloud collected by the 16-line laser scanner.On the premise of retaining the geometric characteristics of the point cloud,the noise points and outliers in the original point cloud are eliminated.The deep learning-based end-to-end cloud segmentation framework PointNet++is used to obtain scene semantic information,which is used to assist SLAM positioning.Then,to solve the problem that the ICP registration algorithm is easy to get local optimum because the high dynamic characteristics of quadruped robot leads to a low point cloud overlap rate and the single sensor cannot provide the initial registration value.Use the Super4PCS algorithm that allows any initial posture to perform pre-registration to build a laser odometer frame that combines coarse and fine registration.The environmental semantic information is used as a priori knowledge into the laser odometer,removing unstable obstacles in the environment,improving the quality of the point cloud to be registered,and performing point cloud feature extraction.A stepped ICP algorithm based on semantic features is proposed to improve the registration accuracy and robustness of the laser odometer.In addition,in order to solve the problem that the error of the laser odometry will accumulate over time,the back-end optimization algorithm based on graph optimization is used to convert the optimization problem of the SLAM system into the form of graph.The graph is constructed by a laser odometer and a reliable loopback detection method based on semantic information,and then the graph optimization algorithm is used to adjust the pose node to modify the trajectory of the quadruped robot,thereby improving the overall positioning and mapping accuracy of the SLAM system.Finally,this paper uses the public data set Semantic KITTI to train PointNet++,and uses the quadruped robot platform to collect and manually label the data set to fine-tune the model,which improves the generalization ability of the model.In addition,with the help of real scenes,the experiments of each algorithm are verified,and the accuracy and performance of the algorithm before and after combining semantic information are compared.The results show that the SLAM algorithm combined with semantic information has better robustness,accelerates the speed of registration,improves the accuracy of positioning and mapping,and can create a global consistent map in large-scale scenes,which meets the application of quadruped robots demand. |
