【摘要】：本文基于PrimeSense1.09深度相機(jī)研究室內(nèi)場(chǎng)景的3D重建問題,對(duì)廣泛研究和使用的傳統(tǒng)KinectFusion算法存在的缺陷進(jìn)行改進(jìn)。研究了針對(duì)傳統(tǒng)KinectFusion算法使用密集體積來存儲(chǔ)3D場(chǎng)景信息,由于內(nèi)存容量等的限制使得KinectFusion只能用于小場(chǎng)景物體的重建的不足,借鑒使用Hash結(jié)構(gòu)的方法來解決內(nèi)存空間限制這一問題。本文使用Hash結(jié)構(gòu),允許進(jìn)行實(shí)時(shí)訪問和隱面數(shù)據(jù)的更新,并且不需要定型或分層網(wǎng)格數(shù)據(jù)結(jié)構(gòu)。而且,數(shù)據(jù)可以高效地進(jìn)出Hash表,實(shí)現(xiàn)GPU和CPU數(shù)據(jù)交換,彌補(bǔ)空間不足的缺陷,這就為傳感器運(yùn)動(dòng)時(shí)存儲(chǔ)空間的可伸縮性提供了方便,并能夠?qū)⑿碌膱?chǎng)景信息不斷拼接到已有的場(chǎng)景模型中,最終實(shí)現(xiàn)大范圍場(chǎng)景信息的3D重建。研究了針對(duì)KinectFusion算法在位姿估計(jì)過程中使用點(diǎn)-面的最近點(diǎn)迭代(ICP)算法,由于其對(duì)初值敏感,且易陷入局部最小,在使用ICP算法進(jìn)行匹配和跟蹤定位時(shí)存在累積誤差,導(dǎo)致相機(jī)突然運(yùn)動(dòng)時(shí)不能有效跟蹤,在基于Hash結(jié)構(gòu)的KinectFusion算法的基礎(chǔ)上,提出一種結(jié)合陀螺儀進(jìn)行擴(kuò)展卡爾曼濾波器(EKF)數(shù)據(jù)融合的方法。使用一種型號(hào)為L(zhǎng)PMS-CU(OEM)的陀螺儀傳感器測(cè)量得到的相機(jī)位姿,與通過ICP算法計(jì)算得到的深度相機(jī)的位姿,使用EKF進(jìn)行數(shù)據(jù)融合得到更加精確的位姿,并在相機(jī)突然運(yùn)動(dòng)時(shí)仍然能夠?qū)崿F(xiàn)有效跟蹤,提高了算法的魯棒性。實(shí)驗(yàn)結(jié)果表明,本文所采用的改進(jìn)算法能在一定程度上解決KinectFusion算法存在的問題,能有效的應(yīng)用于室內(nèi)場(chǎng)景的3D重建中,并取得良好的效果。
[Abstract]:Based on the research of 3D reconstruction of indoor scene with PrimeSense1.09 depth camera, the defects of traditional KinectFusion algorithm, which is widely studied and used, are improved in this paper. Aiming at the traditional KinectFusion algorithm, we use dense volume to store 3D scene information. Because of the limitation of memory capacity, KinectFusion can only be used for the reconstruction of small scene objects. Learn from the use of Hash structure to solve the problem of memory space constraints. In this paper, the Hash structure is used to allow real-time access and update of hidden data, without the need for stereotyped or hierarchical grid data structures. Moreover, the data can enter and leave the Hash table efficiently, realize the data exchange between GPU and CPU, and make up the shortage of space, which provides convenience for the scalability of the storage space when the sensor moves. And the new scene information can be stitched into the existing scene model continuously, and finally the 3D reconstruction of large-scale scene information can be realized. In this paper, the nearest point iterative (ICP) algorithm is studied for KinectFusion algorithm in the process of position and attitude estimation. Because it is sensitive to initial value and easy to fall into local minimum, there are accumulated errors in matching and tracking location using ICP algorithm. As a result, the camera can not be tracked effectively when it moves suddenly. Based on the KinectFusion algorithm based on Hash structure, an extended Kalman filter (EKF) data fusion method based on gyroscope is proposed. Using a type of LPMS-CU (OEM) gyroscope sensor to measure the camera pose, and ICP algorithm to calculate the depth camera pose, using EKF data fusion to obtain a more accurate pose, It can also achieve effective tracking when the camera moves suddenly, and improve the robustness of the algorithm. The experimental results show that the improved algorithm can solve the problem of KinectFusion algorithm to some extent, and can be effectively applied to 3D reconstruction of indoor scene, and good results have been obtained.
【學(xué)位授予單位】：天津大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2016
【分類號(hào)】：TP391.41

【參考文獻(xiàn)】

相關(guān)期刊論文 前1條

1 吳垠;李良福;肖樟樹;劉侍剛;;基于尺度不變特征的光流法目標(biāo)跟蹤技術(shù)研究[J];計(jì)算機(jī)工程與應(yīng)用;2013年15期

相關(guān)碩士學(xué)位論文 前1條

1 琚旋;深度與彩色相機(jī)的聯(lián)合標(biāo)定及其在增強(qiáng)現(xiàn)實(shí)中的應(yīng)用[D];浙江大學(xué);2014年

，

本文編號(hào)：2308024