面向识别与跟踪的局部特征提取方法研究

CORC > 沈阳自动化研究所 > 中国科学院沈阳自动化研究所 > 光电信息技术研究室

题名	面向识别与跟踪的局部特征提取方法研究
作者	赵春阳
学位类别	博士
答辩日期	2016-05-26
授予单位	中国科学院沈阳自动化研究所
导师	赵怀慈
关键词	特征识别与跟踪多模态图像匹配高精度局部特征精确鲁棒单应估计并行SURF局部特征与硬件实现
其他题名	Research on Local Invariant Feature Used in Object Detection and Tracking
学位专业	模式识别与智能系统
中文摘要	本文面向目标识别与跟踪应用，针对多模态图像目标识别、特征跟踪误差和实时性等问题，开展局部特征提取方法相关研究工作。本文主要研究内容和取得的研究成果包括以下几个方面：1. 针对多模态目标识别问题，提出一种新的局部特征描述符及其匹配方法，显著提高了多模态图像匹配性能。首先，基于相位一致性和单演信号理论，本文提出了一种新的角点和线段特征提取方法，由于相位一致性图像具有对比度不变性，因此提取的角点和线段特征是模态变化不敏感的；然后，以角点为中心，基于线段特征构建线段特征描述符(MultiModal Line Segment Descriptor, MM-LSD)；最后，采用基于归一化相关度量的双点、双向匹配方法实现特征匹配。实验结果表明，对于多模态图像，本文方法具有较好的匹配性能，相比于基于表观的次优方法MM-SURF(Multi-Modal SURF)，匹配正确率和重复率提高至2倍左右，相比于基于边缘的次优方法EOH-SIFT(Edge Orientation Histograms SIFT)，匹配正确率和重复率分别提高21%和78%。并且在旋转、尺度变化条件下，匹配性能也显著优于对比方法。另外，本文方法对于不同模态的图像都具有较好的匹配性能。2. 面向平面目标跟踪，针对局部特征跟踪方法的跟踪误差问题，提出一种精确、鲁棒的单应矩阵估计方法，显著降低了帧间跟踪误差。首先，本文分析了SURF和SIFT等斑点特征的定位误差影响因素，选取了一种高精度局部特征提取方法和基于协方差矩阵的特征定位误差表征方法；然后，提出一种基于特征定位误差协方差加权的内点特征优化选取以及精确、鲁棒单应矩阵估计方法。仿真数据实验表明，相比于RANSAC(RANdom SAmple Consensus)、LMedS(Least Median of Squares), MSAC(M-estimator SAC)和MLESAC(Maximum Likelihood SAC)等结合L-M(Levenberg-Marquardt)的标准单应矩阵估计方法，本文方法具有较高的精度和鲁棒性，重投影误差均方根RMSE指标降低25%以上。另外，真实图像实验表明，本文方法的单应矩阵估计精度也优于其他对比方法，重投影误差均方根RMSE指标降低约8~11%左右。3. 针对SURF等局部特征提取方法的实时性问题，提出一种并行SURF特征提取方法及其硬件实时实现方法，满足了特征跟踪算法的实时性要求。本文以SURF算法为研究对象，首先，提出了一种基于圆形特征区域和径向梯度变换的并行SURF算法，相比于标准SURF算法，匹配性能差异小于6%，在保证相当匹配性能的同时，实现了SURF算法可并行流水计算；然后提出了采用FPGA多存储器、多路并行流水机制的并行SURF算法硬件实现方法，对于720×576分辨率的图像，处理速度达到25帧/秒，满足了特定特征跟踪应用的实时性需求。
英文摘要	This paper addresses the following problems: the object recognition problem of multimodal images, the real-time and accuracy problem of local feature based tracking. To solve the problems above, the paper makes the study mainly focused on the local feature extraction method. The main research achievements of this paper are as follows: 1. In order to solve the object recognition problem of multimodal images, a local descriptor and its matching method was proposed in this paper. First, the corner and line segment extraction method was proposed based on phase congruency and local direction information. The extracted corners and line segments are robust to modality invariants because the phase congruency information is insensitive to contrast variants. Then, the MultiModal Line Segment Descriptor (MM-LSD) was constructed by using the corner for a center. Finally, the feature matching method based on normalized correlation function, dual candidate and bi-direction matching was proposed. The experimental results indicate that the proposed method achieves good multimodal image matching performance. Compared with MM-SURF (Multi-Modal SURF), the precision and repeatability of the proposed method is increased to about 2 times. Compared with EOH-SIFT (Edge Orientation Histograms SIFT), the precision and repeatability of the proposed method is increased to about 21% and 78% respectively. The proposed method is also significantly better than the comparison methods under the rotation, scale change condition. On the other hand, the proposed method also demonstrates its robustness to multimodal images. 2. To solve the accuracy problem of local feature-based plane object tracking, a robust and accurate homography estimation method was proposed in this paper. First, the mechanism of feature localization error was analyzed, and a new high accurate localization feature extraction method and location error metric method based on covariance matrix were proposed. Then, an inlier verification method and a high accuracy and robust homography estimation method based on location error metric was proposed in this paper. Compared with state-of-the-art homography estimation methods, such as LMedS (Least Median of Squares), RANSAC (RANdom SAmple Consensus), MSAC (M-estimator SAC) and MLESAC (Maximum Likelihood SAC) combining with L-M (Levenberg-Marquardt) method, the accuracy and robustness of homography estimation is improved greatly, the RMSE of reprojected error is reduced by more than 25%. In addition, experiment results using real image show that the accuracy and robustness of the proposed method also bettern than other comparison methods, the RMSE of reprojected error is reduced about 8%~11%. 3. For improving the real-time of local feature-based tracking, a parallel SURF method and its hardware implementation method was proposed in this paper. First, a parallel SURF algorithm using circular feature region and radial gradient transform method was proposed. The difference of matching performance between SURF method and parallel SURF method is less than 6%. The proposed parallel SURF method not only realizes the parallel process compared with SURF, but also has the same matching performance with SURF. Then, a FPGA-based implementation method of the proposed parallel SURF algorithm using multi-memories and multi-path pipelinings mechanism was put forward. For a video stream with resolution of 720*576, the processing speed reaches 25 fps and satisfies the real-time requirement of local feature-based tracking.
语种	中文
产权排序	1
页码	113页
内容类型	学位论文
源URL	[http://ir.sia.cn/handle/173321/19628]
专题	沈阳自动化研究所_光电信息技术研究室
推荐引用方式 GB/T 7714	赵春阳. 面向识别与跟踪的局部特征提取方法研究[D]. 中国科学院沈阳自动化研究所. 2016.