Loop Closure Detection Algorithm Based On Color And Depth Image

The problem of simultaneous localization and mapping(SLAM)is that the robot calculates the current location and environment map through sensor data in unfamiliar environment.In vision-based SLAM,the pose of the robot is usually computed by the polar constraint based on two frames of images at the adjacent time.Since the current pose is multiplied by the previous frame pose and the transformation matrix,the error accumulates gradually and passes to the next frame.Loop closure detection algorithm is usually used to identify whether the current position of the robot appears in the trajectory of historical motion,thus adding constraints to the graph optimization of pose calculation,so as to restrain the influence of cumulative error growth on the final calculation results.Loop closure detection in VSLAM usually uses color images based on monocular cameras.However,because the two-dimensional color image descriptor is greatly affected by the change of view angle and illumination,it is easy to make misjudgments in the case of scene transformation,and the loss caused by mismatching is much greater than the normal error calculation,which has a great impact on the image.After comprehensive analysis of the research status of loop closure detection algorithm,a 3D global eigenvector matching method based on color depth information is proposed,and the effectiveness of the method is verified by using TUM data sets.The work mainly includes the following aspects.Firstly,the point cloud data in the world coordinate system is obtained from the depth image.According to the coordinate correspondence of the depth image,a large number of matching point pairs are established.Then,the point cloud data near the matching point pairs are extracted and transformed into TSDF(Truncated Signed Distance Function)model.The preparation of convolutional neural network data sets is completed by processing a large number of data.Secondly,the STN(Spatial Transformer Networks)network layer based on 2D affine transformation is extended to 3D space,and a 3D STN network layer is constructed according to perspective transformation.The 3D STN network layer is integrated into the 3dMatch network,and the STN-MatchNet is proposed,which improves the matching ability of the network to local feature blocks.Experiments show that under the same training conditions,for the same test data,STN-MatchNet achieves 3.7% accuracy improvement compared with 3dMatch at 90% recall rate.Finally,in order to improve the point cloud similarity calculation method based on RANSAC and nearest point retrieval,the VLAD network layer is added to the STNMatchNet back-end,and a complete end-to-end network is constructed,which can directly calculate the global eigenvectors of the point cloud image,and then calculate the point cloud similarity according to the distance between the eigenvectors.Experiments show that STN-MatchNet is better than NetVLAD scheme using color images in Closed-loop Detection tasks.