Design And Implementation Of Indoor 3D Reconstruction Based On Dynamic Sequential Visual-word-vectors

Real-time 3D reconstruction of indoor scenes is the key to home virtual reality and robotic applications.However,scalability brings challenges of drift in pose estimation,introducing significant errors in the accumulated 3D model.Approaches often require hours of offline processing to globally correct 3D model errors.The recent online optimization algorithm,such as the bundle adjustment,demonstrate compelling results,but it suffers from the fact that it needs too much time to perform online correction.the reason for that comes from two aspects:(1)The Loop Closure Detection fails to detect the loop clousres,which results in the inability to perform the online correction;(2)the bundle adjustment needs too much time to perform online correction,preventing true real-time use.In order to address this issue,a Loop Closure Detection based on dynamic sequential Visual Word Vectors(S-VWV)and a hierarchical optimization strategy based on deformation Graph are adopted.Its core is to enhance the reall of detecting loop closures and to reduce the time for online correction,so that it has the ability to update the 3D model after data has been integrated,in accordance with the newest pose estimates.For the Loop Closure Detection fails to detect loop closures in the revisited areas,a Bag of Visual Words(BVW)is used to propose a Loop Closure Detection based on dynamic sequential Visual-Word-Vectors.By analyzing the methods adopted in the existing 3D reconstruction systems,the low recall of the Loop Closure Detection is the main reason for the camera to fail to find the coincident trajectory,and the online correction cannot be executed to correct the model.A Loop Closure Detection method based on dynamic sequential Visual-Word-Vectors uses well established techniques for creating a bag of visual words with a tree structure and describe the visual information of entire regions using Visual-Word-Vectors by extending the notions of word vectors,transforming visual word vectors into sequence visual word vectors.The fact that the proposed approach does not rely on a single image to recognize a site allows for a more robust place recognition,and consequently loop closure detection.The impact of detection ensures the accuracy and the recall of Loop Closure Detection.For the problem that the bundle adjustment in the 3D reconstruction is time-consuming,a hierarchical optimization strategy based on a bundle adjustment and a deformation graph algorithm is proposed.The fact of the growing number of images and map points causes the matrix dimension to be too large when the bundle adjustment correct the 3D model errors,which leads to a disadvantage of needing too much time.Using the deformation map algorithm to extract map points from the local or global map uniformly,adopting the cluster adjustment optimization algorithm to solve the pose of the image and map points,and establishing a hierarchical optimization method for global maps and local maps to effectively reduce the number of images to be optimized.This overcomes the shortcomings of time-consuming increase caused by too large matrix dimension in the bundle adjustment.In the Linux,a real-time indoor 3D reconstruction software was designed and implemented,and three experiments were completed.As the Experiments showed,the Loop Closure Detection based on dynamic sequential visual-word-vectors has a higher recall rate.The hierarchical optimization based on the Deformation Graph can ensure that the optimization takes less than 100 ms.The Loop Closure Detection and The hierarchical optimization guarantees that the 3D Reconstruction has higher reconstruction accuracy.