基于SIFT特征降維的視頻車輛跟蹤研究

詹智財 惠浩添 張松松

摘要：針對尺度不變特征變換(SIFT)算法在匹配時特征數(shù)量過多導(dǎo)致的耗時過長的問題，該文提出一種基于SIFT和主成分分析(PCA)相結(jié)合的SIFT特征降維的視頻車輛跟蹤算法。首先利用SIFT提取出車輛視頻圖像幀中的所有特征點及其特征向量，其次使用PCA算法對其維數(shù)約減并找出各自的具有代表性的特征參數(shù)，達到對特征點向量降維的目的，最后利用歐式距離找出不同車輛圖像幀中相似的車輛。實驗證明，該算法在保證原SIFT算法魯棒性、穩(wěn)定性的同時減少了計算量，增加了匹配效率，增強了實時性。

關(guān)鍵詞：SIFT；PCA；降維；車輛跟蹤

中圖分類號：TP391文獻標識碼：A文章編號：1009-3044(2012)16-3954-04

Video Vehicle Tracking Based on Dimension Reduction of SIFT Features

ZHAN Zhi-cai, HUI Hao-tian, ZHANG Song-song

(School of Computer Science and Telecommunication Engineering, JiangsuUniversity, Zhenjiang 212013, China)

Abstract: In this paper, a video vehicle tracking algorithm based on the combination of Scale Invariant Feature Transform (SIFT) and Prin? ciple Component Analysis (PCA), which is called PCA-SIFT, is proposed to deal with the problem that a long time is taken caused by ex? cessive number of characteristics in the matching with SIFT algorithm. Firstly, SIFT is applied to extract all the feature points and vectors of the vehicle video image frames, and then PCA is used to reduce dimensions, followed by the identification of representative characteristic parameters to achieve the purpose of feature dimensionality reduction. Finally, the Euclidean distance is applied to find similar vehicles in the different vehicle image frames. The experimental results show that the algorithm proposed in this paper has advantages of reducing the cost of computation, improving the matching efficiency and enhancing the real-time performance while maintaining the robustness and sta? bility of the original SIFT algorithm.

Key words: Scale Invariant Feature Transform; Principle Component Analysis; dimensionality reduction; vehicle tracking

在根據(jù)特征值的視頻車輛跟蹤實現(xiàn)過程中，由于SIFT算法提取視頻圖像特征點個數(shù)不同，特征點向量維數(shù)較多，而會使得圖像之間的匹配時間較長，該文將PCA引入到SIFT算法中來，利用PCA可將提取數(shù)據(jù)的主成分而忽略次成分達到不損精度的原理使得SIFT求得的特征描述符的維數(shù)從128維進行大幅度降維。通過實驗分析，PCA-SIFT算法達到了保留SIFT算法的精確性但又解決了匹配時時間過長的問題，因而可以滿足視頻車輛跟蹤的精確性與實時性的要求。

該文中PCA-SIFT算法雖減少了大量的匹配時間，增強了實時性，但在要求實時很高的場景下本算法不能體現(xiàn)很強的優(yōu)勢。因此，在今后的研究中，將針對特征提取算法上進一步的研究，使之有SIFT的準確性，但算法趨于簡單的方法。

[1] Stefano M,Carlamaria M.Vision-based bicycle and motorcycle classification [J].Pattern Recognition Letters,2007(28):1719-1726.

[2] Zhao Z X,Yu S Q,Wu X Y,et al.A multi-target tracking algorithm using texture for real-time surveil-lance[J].Proceedings of the IEEE In? ternational Conference on Robotics and Biomimetics.Bangkok,Thailand:IEEE,2009:2150-2155.

[3] Ryu H R,Huber M A.particle flter approach for multi-target tracking[J].Proceedings of the IEEE/RSJ Interna-tional Conference on Intelli? gent Robots and Systems. San Diego, USA: IEEE,2007:2753-2760.

[4] Staufer C,Grimson W E L.Learning patterns of activity using real-time tracking [J].IEEE Transactions on Pattern Analysis/Machine Intelli? gence,2000,22( 8):747-757.

[5]馮嘉.SIFT算法的研究和改進[D].長春:吉林大學(xué),2010.

[6]吳柯,牛瑞卿,王毅,等.基于PCA與EM算法的多光譜遙感影像變化檢測研究[J].計算機科學(xué),2010,37(3):282-284.

[7]吳春國,梁艷春,孫延風(fēng),等.關(guān)于SVD與PCA等價性的研究[J].計算機學(xué)報,2004(2):286-288.