Research On Visual Odometry Of Object Based Mobile Robot SLAM

Visual SLAM technology has been frequently applicated in low-precision localization scenes with benifit of the low cost of camera sensors and the perception ability it could provide.As a collection of environment information,vSLAM is also one of the hotspots of recent research,combined with high-precision positioning systems with rich semantic information.Task-operated robots processing with the need of object-level semantic information in complex unstructured scenarios while positioning and mapping,which presents a new challenge to visual SLAM system.We propose an object-level SLAM front-end tracking algorithm based on RGBD cameras to achieve the specifications of visual SLAM algorithm for high accuracy and high real-time requirement in complex scenes.The algorithm runs in real time,taking RGBD images as input,outcomes three-dimensional pose of the object represented in camera coordinate system.The algorithm refines object pose by co-vision observations within local map,and further improve the localization precision by fusing object-point landmark features in certain states,constructed a semantic map of point-object features in the meanwhile.The specific work content of this article as follows:First,2D bounding box corners are obtained through the object detection algorithm based on deep learning method which takes RGBD images as input.Converting the problem of 3D object pose estimation to a 2D image processing problem by projecting object observation to the supporting plane,and the threedimensional object pose is captured by the method of confidence voting.Secondly,depth filter is adapted to update feature points which are extracted from the converged objects.Take full advantages of object feature points to obtain high precision object pose tracking performance.Object-level pose graph model is constructed by the connection of co-vision objects.Independent of the point-based graph model,object-level pose graph could refine object pose with the connection edge,on the other hand,it maintaining a form of information fusing method.Finally,we tested our algorithm on the open source data set TUM-RGBD visual odometer data to verify the improved performance of the algorithm in accuracy,object recognition speed and robustness.Through the data analysis,we compare the difference of object-based algorithm with traditional visual SLAM system.