﻿ 面向小运动目标的压缩域跟踪方法
 计算机系统应用  2018, Vol. 27 Issue (12): 143-149 PDF

Compressed-Domain Object Tracking for Small Moving Targets
ZHANG Xin-Sheng, LIU Hao, SUN Xiao-Fan, KUANG Qi-Gang, WU Le-Ming
School of Information Science and Technology, Donghua University, Shanghai 201620, China
Foundation item: Natural Science Foundation of Shanghai (18ZR1400300)
Abstract: The compressed-domain object tracking approaches utilize the information that is directly extracted from the compressed bitstream, such as motion vector and block coding modes. Because the existing compressed-domain tracking methods have poor tracking performance for small moving targets, this study proposes a compressed-domain tracking algorithm for small moving targets. By analyzing the shortages of the existing algorithms, the performance of small target tracking is improved from the acquisition of initial frame mask, the setting of outlier boundary and the edge control of the predicating small target, and some system parameters of the block-coding system are optimized through data-driven methodology. Experiential results on three small-target video sequences show that compared with other object tracking methods, the proposed method can effectively improve the tracking performance for small moving targets in terms of accuracy and F-measure.
Key words: compressed domain     object tracking     block coding     motion vector     small target

1 序言

2 现有方法分析

 图 1 基于MRF的目标跟踪算法流程图

 ${Precision = \frac{{TP}}{{TP + FP}} = \frac{{TP}}{{TP + \lambda \cdot TP}} = \frac{{\rm{1}}}{{{\rm{1}} + \lambda }}}$ (1)

 图 2 Precision和倍数 $\lambda$ 的关系图(Precision为小数形式)

 图 3 City序列起始帧变化图

3 所提算法 3.1 算法框架

3.2 算法数学模型

 $P = ({w^t}|{w^{t - 1}},{K^t}) = \frac{{P({w^t}|{w^t},{K^t}) \cdot P({K^t}|{w^t}) \cdot P({w^t})}}{{P({w^{t - 1}},{K^t})}}$ (2)
 图 4 Ground序列起始帧变化图

 ${w^t} = \mathop {\arg \max }\limits_{\psi \in \Omega } {\left\{ {P({w^{t - {\rm{1}}}}|\psi ,{K^t}) \cdot P({K^t}|\psi ) \cdot P(\psi )} \right\}}$ (3)

 ${w^t} = \mathop {\arg \min }\limits_{\psi \in \Omega } \{ - \log P({w^{t - 1}}|\psi ,{K^t}) - \log P({K^t}|\psi ) - \log P(\psi )\}$ (4)

 $\begin{array}{l}{w^t} = \mathop {\arg \min }\limits_{\psi \in \Omega } \{ - \gamma \cdot \log P(\alpha \cdot {w^{t - 1}}|\psi ,{K^t}) - \\\quad\quad\log P({K^t}|\psi ) - \log P(\psi )\} \end{array}$ (5)

 $P({K^t}|\psi ) = {{{\rm{min}}\left\{ {\frac{{{d'}(n)/{\sigma _{{d'}}} - {\rm{2}}}}{{\rm{2}}}{\rm{, 1}}} \right\}}/\rho }$ (6)
 $d'({\bf{n}}) = \left\{ \begin{gathered} d({\bf{n}}){\rm{ }}d({{n}}) \leqslant \beta \cdot {\sigma _d} \\ 0{\rm{ }}d({{n}}) > \beta \cdot {\sigma _d} \\ \end{gathered} \right.$ (7)
 ${d({\bf{n}}) = {{\left\| {{v'}({\bf{n}}) - \hat v} \right\|}_2}}$ (8)

3.3 块编码感知的系统参数优化

 图 5 本文算法流程图

4 实验分析

 图 6 Ground序列在64组实验参数下的Precision结果. Precision为百分比形式.

 图 7 三种算法的指标随Sky序列帧号的变化图

 图 8 本文算法在不同QP下的平均跟踪结果

 图 9 三种算法在不同QP下的指标比较

 图 8 本文算法在不同QP下的平均跟踪结果

5 结论

 [1] Wu Y, Lim J, Yang MH. Online object tracking: A benchmark. Proceedings of 2013 IEEE Conference on Computer Vision and Pattern Recognition. Portland, OR, USA. 2013. 2411–2418. [2] Yilmaz A, Javed O, Shah M. Object tracking: A survey. ACM Computing Surveys, 2006, 38(4): 13. DOI:10.1145/1177352 [3] 张微, 康宝生. 相关滤波目标跟踪进展综述. 中国图象图形学报, 2017, 22(8): 1017-1033. [4] Chen YM, Bajić IV. Compressed-domain moving region segmentation with pixel precision using motion integration. Proceedings of 2009 IEEE Pacific Rim Conference on Communications, Computers and Signal Processing. Victoria, BC, Canada. 2009. 442–447. [5] 李旭超, 朱善安. 图像分割中的马尔可夫随机场方法综述. 中国图象图形学报, 2007, 12(5): 789-798. DOI:10.3969/j.issn.1006-8961.2007.05.004 [6] Zeng W, Du J, Gao W, et al. Robust moving object segmentation on H.264/AVC compressed video using the block-based MRF model. Real-Time Imaging, 2005, 11(4): 290-299. DOI:10.1016/j.rti.2005.04.008 [7] Babu RV, Tom M, Wadekar P. A survey on compressed domain video analysis techniques. Multimedia Tools and Applications, 2016, 75(2): 1043-1078. DOI:10.1007/s11042-014-2345-z [8] Khatoonabadi SH, Bajić IV. Video object tracking in the compressed domain using spatio-temporal Markov random fields. IEEE Transactions on Image Processing, 2013, 22(1): 300-313. DOI:10.1109/TIP.2012.2214049 [9] Xu J, Tu Q, Li CW, et al. Video saliency map detection based on global motion estimation. Proceedings of 2015 IEEE International Conference on Multimedia & Expo Workshops. Turin, Italy. 2015. 1–6. [10] Richardson IEG. H.264 and MPEG-4 video compression: Video coding for next-generation multimedia. New York: John Wiley & Sons, 2004. [11] 王闪, 吴秦. 基于马尔可夫随机场模型的运动对象分割算法. 传感器与微系统, 2016, 35(7): 113-115, 119.