Research On SLAM Technology Of Indoor Mobile Robot Based On Depth Camera

The location and navigation of vehicle in outdoor environment can be realized by GPS and high precision map,but in the case of weak indoor GPS signal,the vehicle can only rely on its own sensors for location and navigation.Therefore,Simultaneous Localization and Mapping(SLAM)technology emerges as the times require.The early research of SLAM is mainly based on lidar and IMU.With the rapid development of computer vision technology,cameras rely on its cheap,information-rich and other characteristics began to be widely used in the field of SLAM.With the release of Microsoft's RGB-D camera,called Kinect,for its direct access to scene depth information,it solves the problems of large computational complexity and scale uncertainty in SLAM research using monocular camera and binocular camera,so it has become a new hotspot of visual SLAM.This algorithm is based on the RGB color image obtained by Kinect and the16-bit single channel depth image.At the front of the algorithm,two adjacent RGB images are first extracted and matched by ORB features.The matching features are solved by PnP(Perspective-n-Point)to get the motion estimation of the camera between two adjacent frames.According to the size of the estimated motion,the key frame is screened out,and the selected key frame is passed to the back end of the SLAM.The selected key frame is input into convolution neural network to extract its feature vector.The similarity degree of image is obtained by calculating the cosine similarity between vectors,and the image similarity is used to determine whether the loop appears or not.If there is an anti-return loop,the constraint is added to the position map.Finally,the global posture is optimized by using the graph optimization theory.In the mapping part,the estimated pose and color map obtained by Kinect are used to concatenate the dense point cloud map with the depth map,and the point cloud map is converted into octree map and compared.The algorithm in this paper is tested on a variety of datasets,and the experimental results show that:1)Compared with the non-loop detection algorithm,the proposed algorithm can successfully detect the loop and improve the mapping accuracy.2)Compared with the traditional word bag method,the method of extracting image features by convolution neural network in this paper has some advantages over the traditional word bag method in the accuracy,recall rate and speed of loop detection.The experimental results show that the average cost of calculating similarity between two pictures by word bag method is 0.5 s,while the average cost for our method is 0.13 s,and the recall rate of this method is 21% higher than that of word bag method under the condition that certain accuracy is guaranteed.3)Compared with point cloud map,octree map can express the space share more intuitively and more compactly,which is more beneficial to improve the running efficiency of the algorithm.The experimental results show that the size of an octree map corresponding to a point cloud map of 47.7m is only 0.439 m.