Abstract: The aim is to solve the problem that the original adaptive Monte Carlo localization is prone to re-localization failure in similar dynamic environment. Firstly, the odometer, laser radar and vision information are fused, and the instant localization of the robot and the creation of grid map are completed based on the Gmapping algorithm. Secondly, the KinectV2 color image is input into the object detection and recognition method to obtain the boxes and categories of object detection in the image, and the semantic segmentation is realized by combining GrabCut algorithm. The corresponding position of the object in the depth image is calculated by using the registration data. Then, the acquired KinectV2 depth data is converted into point cloud information and projected onto the global map coordinate system to obtain the object semantic map. The relationship between the semantic objects is fully explored, and the dimension chain of semantic objects is constructed. And a fast and effective retrieval method of dimension chain is proposed. At the same time, the Bayesian method is used to reduce the influence of false detection and repeated detection on object semantic map. The semantic map of object and grid map are aligned by origin coincidence and pose alignment, so as to construct two-dimensional semantic grid map. The research proposes an improved AMCL re-localization algorithm base on the dimension chain of semantic objects. Finally, a large number of re-localization experiments in real environments verify the superior performance of the method in similar and dynamic environments.

Key words: similar dynamic environment, the dimension chain of semantic objects, Bayesian method, two-dimensional semantic grid map, re-localization