跟踪-学习-检测框架下改进加速梯度的目标跟踪

引用本文

杨欣, 夏斯军, 刘冬雪, 费树岷, 胡银记. 跟踪-学习-检测框架下改进加速梯度的目标跟踪. 2018, 48(2): 533-538
YANG Xin, XIA Si-jun, LIU Dong-xue, FEI Shu-min, HU Yin-ji. Target tracking based on improved accelerated gradient under tracking-learning-detection framework. Journal of Jilin University Engineering and Technology Edition, 2018, 48(2): 533-538 复制到剪切板

Doi:10.13229/j.cnki.jdxbgxb20160411
Permissions

跟踪-学习-检测框架下改进加速梯度的目标跟踪

杨欣^1,², 夏斯军¹, 刘冬雪¹, 费树岷³, 胡银记²

1.南京航空航天大学自动化学院,南京 210016

2.光电控制技术重点实验室,河南洛阳 471000

3.东南大学自动化学院, 南京 210096

作者简介:杨欣(1978-),男,副教授,博士. 研究方向:模式识别,图像处理. E-mail:yangxin@nuaa.edu.cn

基金项目:国家自然科学基金项目(61573182,61172135,61101198); 光电控制技术重点实验室和航空科学基金联合项目(20145152027)

摘要

单目标持久跟踪的主要难点是由于目标姿态、相似背景及遮挡等因素而导致的漂移问题。基于此提出了一种改进L1APG (L1 tracker using accelerated proximal gradient approach) 的目标-学习-检测(TLD)目标跟踪算法。首先,在L1APG跟踪器中加入遮挡检测判断;其次,将遮挡程度转换为目标模板和背景模板系数的权重;最后,用改进的L1APG跟踪器取代传统TLD框架中的跟踪器,自适应地根据遮挡程度改变模板系数,从而有效地提高了跟踪效果。实验表明:本文算法与传统TLD跟踪框架相比,能更好地处理遮挡和漂移问题,具有较好的稳定性和鲁棒性。

关键词: 人工智能; 目标跟踪; 目标-学习-检测; 遮挡; 漂移

中图分类号:TP391.41 文献标志码:A 文章编号:1671-5497(2018)02-0533-06

Target tracking based on improved accelerated gradient under tracking-learning-detection framework

YANG Xin^1,², XIA Si-jun¹, LIU Dong-xue¹, FEI Shu-min³, HU Yin-ji²

1.College of Automation Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China

2.Key Laboratory of Photoelectric Control Technology, Luoyang 471000, China

3.School of Automation, Southeast University, Nanjing 210096, China

Abstract

The main challenges in a single target persistent tracking are the factors such as the change of the target pose, similar background and occlusion, which account for the difficulty in solving the drift problem. Based on Tracking-Learning-Detection (TLD), we propose an improved Li Tracker Using Accelerated Proximal Gradient Approach (L1APG) tracking algorithm. First, we add occlusion detection to the L1APG tracker. Then, we transform the occlusion problem into the weight of the target template and the background template coefficients. Finally, the original tracker is replaced by the improved L1APG tracker in the traditional TLD algorithm, the coefficient weights vary adaptively according to the degree of occlusion in real time, which contributes to improving the tracking results effectively. Experiments show that, compared with TLD and L1 tracking algorithms, the proposed algorithm can better deal with the occlusion and drift problems and possesses better stability and robustness.

Key words: artificial intelligence; target tracking; tracking-learning-detection (TLD); occlusion; drift

Show Figures

0 引言

视频目标跟踪作为计算机视觉研究领域中的一个关键技术, 已广泛应用到国防和民用的各个领域中, 如智能监控、导航制导、医学诊断、人机交互等。同时视频跟踪在实际的应用中面临着多种问题, 比如:跟踪对象的角度、姿态发生改变以及被遮挡等情况, 而长时间目标跟踪更是一个富有挑战性的工作。传统的跟踪方法可分为两类:基于跟踪器的跟踪算法和基于检测器的跟踪算法, 前者主要在于目标特征的选取, 代表性的特征包括目标的质点^[1]、目标的轮廓^{[2, 3, 4]}、链式模型^[5]和光流^[6]。此类方法都是假设目标始终可见, 一旦目标消失就会导致跟踪失败。同样, 检测跟踪方法通常都是基于局部图像特征或者滑动窗口, 对每一帧单独处理检测出目标位置, 此类方法需要离线的学习过程, 并且检测器的性能与学习样本多样性直接相关, 因此基于检测的跟踪算法难以有效地解决过拟合问题。在此背景下, 研究者提出了大量的解决方案, 例如:Mean-shift^{[7, 8, 9]}、粒子滤波^{[10, 11]}、L1跟踪器^[12]、视觉显著性特征^{[13, 14]}、实时压缩感知跟踪 (Real-time compressive tracking)^[15]。这些算法虽然能较好地实现对无遮挡目标的跟踪, 但无法有效地解决跟踪过程中的遮挡和相似背景及目标姿态变化问题, 或者对全部遮挡效果不理想, 因此无法实现持久的跟踪。多特征融合粒子滤波算法^{[16, 17]}对目标遮挡以及背景混淆等诸多问题具有较高的鲁棒性, 但当视频序列的尺寸较大或者背景较复杂需要数量较多的粒子时, 算法处理的时间较长, 会导致实时性变差。为实现持久的目标跟踪, Kalal等^[18]提出了一种新的TLD (Tracking-learning-detection)跟踪算法, 采用跟踪器与检测器并行处理的框架, 并通过学习模块不断更新检测器相应参数值, 最终由整合模块输出跟踪结果。如过跟踪器跟踪失败, 由检测器修正跟踪器的目标模型^[19]。该方法能有效地实现对目标的持久跟踪, 并能在一定程度上克服目标姿态变化问题。但当遮挡前景与跟踪目标灰度特征相似时, 会导致跟踪出现偏差, 尤其当相似干扰与跟踪目标交叉运动时, 会出现漂移现象, 无法再次跟踪到初始目标, 导致跟踪失败。

根据上述分析, 本文提出了一种改进的TLD目标跟踪算法, 用改进L1APG(L1 tracker using accelerated proximal gradient approach)^[20]跟踪器来取代传统TLD算法中基于光流法的跟踪器, 同时加入遮挡检测判断, 设定遮挡阈值, 并根据遮挡程度实时更新目标模板和背景模板系数权重, 进一步解决遮挡问题, 从而抑制漂移问题。

1 TLD跟踪算法

传统TLD算法中跟踪模块主要是用中值流跟踪算法^[18], 中值流跟踪算法就是在光流法的基础上加了个限制, 一个“ 好” 的跟踪应具有正反连续性, 即无论是按照时间上的正序追踪还是反序追踪, 产生的轨迹应该是一样的。根据这个性质规定了任意一个追踪器的FB误差(Forward-backward error):从时间t的初始位置x(t)开始追踪产生时间t+p的位置x(t+p), 再从位置x(t+p)反向追踪产生时间t的预测位置x'(t), 初始位置和预测位置之间的欧氏距离就作为追踪器在t时间的FB误差。选取前、后两帧误差点较小的一半作为追踪点, 以此来估计目标运动位置, 一旦目标消失就会导致跟踪失败。检测器是将方差检测器、集成分类器和最近邻分类器采用级联的方式, 逐次通过每个检测器, 得到最终的目标。根据初始的目标模型, 对每一帧图像进行全面的扫描, 根据每个扫描框与目标的相识度来估计目标可能出现的位置。这种检测方法也会发生错误, 即错误地分类正、负样本。而学习模块中P-N学习主要对这两种错误(错误的正样本和错误的负样本)进行评估, 同时生产新的样本集来更新目标模型, 实现目标的实时更新, 并且对跟踪模块所选取的“ 关键点” (使用的光流法)进行更新, 避免再次出现类似的错误。综合模块主要就是拟合过程:比较跟踪器和检测器的结果, 选择置信度较大者或综合得到目标在下一帧最有可能出现的位置。

2 基于改进L1APG的TLD跟踪

2.1 L1APG跟踪器

传统TLD算法跟踪器的置信度计算依赖于检测器, 因此最终跟踪结果受检测器的影响较大, 并且TLD算法在复杂背景下的跟踪效果较差, 处理遮挡的能力有待提高。

当目标在相对静止的背景中运动时, TLD算法跟踪效果较好。但当背景产生变化时会出现误判, 并且对于相似目标更是无法准确区别。此外, TLD方法对复杂条件下目标的跟踪处理能力不足, 而L1跟踪器由于具有静态原子, 对漂移有一定的抑制能力, 基于此本文对传统TLD进行改进, 采用改进的L1 APG跟踪器模块, 该模块能够快速地求解L1最小问题。并加入遮挡检测, 使其能够有效处理复杂情况, 增加算法的鲁棒性和实时性, 进而能够对目标进行持久跟踪。

候选目标的选择:假设在第t帧时, 目标所在的位置为x_t, 对目标的预测的观测结果定义为y_1:_t-₁={y₁, y₂, …, y_t-₁}, 可以用下面的两个公式来选取目标在下一帧中的位置:

p(x_t|y_1:_t-₁)=∫ p(x_t|x_t-₁)p(x_t-₁|y_1:_t-₁)dx_t-₁(1)

p(x_t|y_1:_t)= $\begin{matrix} \frac{p (y_{t} | x_{t}) p (x_{t} | y_{1 : t - 1})}{p (y_{t} | y_{1 : t - 1})} \end{matrix}$ (2)

对于t帧的最佳目标位置, 近似于后验概率取最大值时:=argmax_xp(x|y_1:_t)。

目标模型的建立:建立目标模型是为了计算样本在下一时刻所在位置观察可能性p(y_t|x_t)。T_t=[, , …, ]∈ R^{d× n}(d≫n)表示t帧的目标模板; Q_t={, , …, }表示采样的状态; M_t={, , …, }为目标模板空间相应的候补目标图像块集。可以得到稀疏表达模型:

$y_{t}^{i}$ =T_t $a_{T}^{i}$ +I $a_{T}^{i}$ , ∀ y∈ M_t(3)

式中:I为琐碎的模板集(单位矩阵), 表示遮挡和噪声; =[; ]为系数。同时, 对于非负约束有利于提高L1跟踪器的鲁棒性。因此, 对于每个候补的目标块, 其稀疏表达都可以通过求解下式的L1规范最小化得到:

$\begin{matrix} {(y_{t}^{i} - Aa)}_{2}^{2} \end{matrix}$ +β , a≥ 0(4)

式中:A=[T_t, I, -I]; β 为平衡因子。

通过观察, 可以得:

$\begin{matrix} {(T_{t} a - y)}_{2}^{2} \end{matrix}$ ≥ $\begin{matrix} {(T_{t} \overset{\land}{a} - y)}_{2}^{2} \end{matrix}$ , ∀ a∈ R^N(5)

式中:

$\hat{a}$ = $\begin{matrix} \underset{a}{argmin} \end{matrix} \begin{matrix} {(T_{t} a - y)}_{2}^{2} \end{matrix}$ (6)

由于求解L1计算量比较大, 因此可以先计算出样本观测可能性, 然后根据可能性的大小对样本进行过滤, 舍弃可能性小的样本以此来减少计算量:

p(z_t|)=exp{-α $\begin{matrix} {(y_{t}^{i} - T_{t} c_{T}^{i})}_{2}^{2} \end{matrix}$ }(7)

p(z_t|)≤ exp{-α $\begin{matrix} {(T_{t} \overset{\land}{a} - y_{t}^{i})}_{2}^{2} \end{matrix}$ }q(z_t|)(8)

式中:α 为常量, 用来控制高斯核的形状; Γ 为常量; 为式(4)的最小值; z_t为对t帧的预估; q(z_t|)为样本的观测可能性上限, 并且当q(z_t|)< $\begin{matrix} \overset{i - 1}{\sum_{j = 1}} \end{matrix}$ p(z_t|), 样本就排除在样本集里, 以此来减少计算量, 提高实时性。

2.2 遮挡检测

本文算法将跟踪过程分为如下两种情况:

(1)当没有遮挡时, 目标在下一帧中的状态应主要由目标模板的稀疏线性组合表示, 并且背景模板只占很小的一部分。

(2)当有遮挡时, 根据遮挡的程度增大或减小背景模板所占的比例。能够根据不同的情形来改变=[; ]各自所占的比重。而式(4)显然无法区分这两种情形, 因此将其改进为:

$\begin{matrix} \{\begin{matrix} \min_{a} {(y - A'a)}_{2}^{2} + β {(a)}_{1} + u_{t} {(a_{I})}_{2}^{2} \\ s.t. a_{T} \geq 0 \end{matrix} \end{matrix}$ (9)

式中:A'=[T_t, I], a=[a_T; a_I]; u_t为用来控制背景模板的权重。

为了更好地处理遮挡情况, 引入遮挡检测判断, 并且根据遮挡程度自适应调节稀疏矩阵系数, 具体如下:设稀疏系数=[; ], 对模板和背景系数分别进行转置可以得到 $\begin{matrix} a_{Τ}^{iT} \end{matrix}$ =[a_T₁, a_T₂, …, a_Tn]∈ R¹^{× n}, 同样=[a_I₁, a_I₂, …, a_Id]∈ R¹^{× d}, 根据式(10)(11), 可以得到判断遮挡的目标模板阈值σ _T以及背景模板阈值σ _I:

σ _T= $\begin{matrix} \overset{m}{\sum_{t = 1}} \end{matrix} \begin{matrix} \overset{n}{\sum_{j = 1}} \end{matrix}$ a_Tj(10)

σ _I= $\begin{matrix} \overset{m}{\sum_{t = 1}} \end{matrix} \begin{matrix} \overset{d}{\sum_{j = 1}} \end{matrix}$ a_Ij(11)

式中:m为常数, 表示无遮挡情况下的初始帧数, 通过实验来确定。

最终跟踪目标是由目标模板和背景模板的稀疏表达, 因此, 目标区域中任意像素点的灰度值可由目标模板与背景模板相应点的线性之和来表示, 改变两者系数值与改变其一皆可以达到同等的效果, 即:

$Z_{p}^{i}$ = $\begin{matrix} w_{p 1}^{i} \end{matrix} \begin{matrix} a_{T_{p}}^{i} \end{matrix}$ + $\begin{matrix} w_{p 2}^{i} \end{matrix} \begin{matrix} a_{I_{p}}^{i} \end{matrix}$ (12)

$Z_{p}^{i}$ = $\begin{matrix} a_{T_{p}}^{i} \end{matrix}$ + $\begin{matrix} w_{p 3}^{i} \end{matrix} \begin{matrix} a_{I_{p}}^{i} \end{matrix}$ (13)

式中:、、分别为第i帧图像预测目标、目标模板及背景模板中第p像素点的值(0~255), 对应于单个像素点=1, 并且的范围为(0~255)。改变权重的值, 能使式(12)(13)同时成立。因此, 只需改变背景模板的权重系数, 就能达到预期跟踪效果, 同时大幅度减少计算量, 进而有效提高运行速度。

在后面每一帧都计算相应的参数值, 与σ _T、σ _I进行比较, 当满足判断遮挡条件时, 更新稀疏表达式的参数所占权重:

(1)当i> m时, 分别计算, 。

(2)当< σ _T& > σ _I时:

$\begin{matrix} \{\begin{matrix} S_{T}^{i} = \sqrt[]{σ_{T} / σ_{T}^{i}} \\ S_{I}^{i} = \sqrt[]{σ_{I}^{i} / σ_{I}} \\ λ_{t} = λ S_{T}^{i} + (1 - λ) S_{I}^{i} \\ u_{t} = λ_{t} u_{t} \end{matrix} \end{matrix}$ (14)

式中:学习率λ 是个固定的参数, 本文算法设定为0.95, 该参数主要更新稀疏表达式中背景模板的比重, 通过线性叠加来更改最终跟踪目标模型, 使其能更好地处理遮挡情况。

2.3 总体算法

本文算法的整体思路为:通过对跟踪器的有效改进, 同时检测器和跟踪器并行处理, 并且学习模块不断更新, 最终结果由综合模块对比输出。具体算法步骤如下:

Step1 设定初始图像中目标所在位置矩阵框参数[x, y, w, h]。

Step2 通过仿射变换, 在目标周围选取正负样本, 以此来初始化检测器相关参数, 并将目标模板保存起来。

Step3 检测是否有遮挡, 若有, 则更新跟踪器的参数u_t(根据式(8)), 否则执行下一步。

Step4 用改进的L1APG跟踪器进行跟踪, 并对跟踪结果计算其置信度pt; 同时检测器并行处理, 得到目标在下一帧的位置及其置信度pd。

Step5 当pt> δ ₁& pd> δ ₂时(δ ₁、δ ₂分别是跟踪器和检测器设定的阈值, 初始值为0.5), 对比跟踪器和检测器的置信度, 当|pt-pd|> ε (ε 为设置参数, 初始值为0.4)时, 其较大者作为最终的跟踪结果。反之, 对其加权处理, 作为目标位置在下一帧的输出, 并调到步骤Step2。

3 实验

3.1 定性分析

对于部分遮挡情况, 用FaceOcc1作为测试图像序列, 观察图1(a)的跟踪结果:开始的几十帧图像跟踪效果相当, 但随后L1APG跟踪框发生畸变, 无法准确跟踪到目标。传统TLD在目标静止而遮挡运动时发生偏移现象, 但随后又恢复跟踪。在部分遮挡基础上加入目标姿态变化, 用FaceOcc2作为测试图像序列, L1APG用红色框表示, TLD用黄色框表示, 本文算法用蓝色框表示。从图1(b)中可以得出:L1APG在目标遮挡之后跟踪框发生畸变, 并且无法恢复, 导致跟踪失败。传统TLD算法在目标姿态发生变化及目标和背景较为相似时会导致跟踪失败, 直到目标姿态恢复到初始状态时, 检测器才能重新检测到目标。在目标遮挡及姿态变化的基础上, 加入快速运动这一因素来对比跟踪效果。用Football作为测试图像序列, 从图1(c)中可以得出:当有相似目标出现时, L1APG跟踪算法无法区分目标。传统的TLD算法随着目标与相似背景的交叉运动会发生偏移现象, 导致跟踪失败。而本文的算法在以上3种情况下都具有良好的稳定性和健壮性。

	Figure Option View Download New Window
	图1 图像序列跟踪结果Fig.1 Image sequences tracking results

3.2 定量分析

定量分析实验将中心位置误差作为精度标准, 中心误差定义为跟踪目标的中心位置与标准目标位置的中心误差, 采用欧式距离进行比较。从图2的误差结果可以看出, 本文的算法在稳定性和健壮性上都要优于其他两种算法。

	Figure Option View Download New Window
	图2 误差结果图Fig.2 Error result diagrams

4 结束语

提出了一种基于TLD框架的跟踪算法, 通过引入改进的L1APG跟踪器来替换传统TLD中的光流跟踪器。由于L1APG跟踪器具有静态模板, 并加入遮挡检测能自适应地更改系数权重, 以此来抑制漂移问题, 从而在目标遮挡、姿态变化以及快速运动的情况下能很好地跟踪目标, 提高了稳定性, 但实时性略低, 以后研究重点将是在保证准确率的基础上, 尽可能地减少粒子数量, 减小计算量, 提高实时性, 能更快速地跟踪目标。

The authors have declared that no competing interests exist.

参考文献

View Option

[1]	Beaudry C, Péteri R, Mascarilla L. Action recognition in videos using frequency analysis of critical point trajectories[C]∥2014 IEEE International Conference on Image Processing (ICIP), Paris, France, 2014: 1445-1449. [本文引用:1]
[2]	Dondo D G, Redolfi J A, Griffa M, et al. Target tracking system using multiple cameras and Bayesian estimation[J]. IEEE Latin America Transactions, 2016, 14(6): 2713-2718. [本文引用:1]
[3]	郝志成, 高文. 多模跟踪技术在轮式侦察车图像处理器的应用[J]. 中国光学, 2011, 4(5): 480-488. Hao Zhi-cheng, Gao Wen. Application of multi-pattarn tracking in wheel type scout car image processor[J]. Chinese Optics, 2011, 4(5): 480-488. [本文引用:1]
[4]	熊晶莹, 戴明. 适应移动智能设备的目标跟踪器[J]. 光学精密工程, 2017, 25(12): 3152-3159. Xiong Jing-ying, Dai Ming. Design of tracker for mobile smart devices[J]. Optics and Precision Engineering, 2017, 25(12): 3152-3159. [本文引用:1]
[5]	Buehler P, Everingham M, Huttenlocher D P, et al. Long term arm and hand tracking for continuous sign language TV broadcasts[C]∥Proceedings of the 19th British Machine Vision Conference, Leeds, UK, 2008: 1105-1114. [本文引用:1]
[6]	Xu J, Li J. Moving target tracking algorithm based on scale invariant optical flow method[C]∥Information Science and Control Engineering (ICISCE), Beijing, China, 2016: 468-472. [本文引用:1]
[7]	杨欣, 费树岷, 李刚, 等. 基于复杂特征融合的改进mean shift目标跟踪[J]. 控制与决策, 2014(7): 1297-1300. Yang Xin, Fei Shu-min, Li Gang, et al. Improved mean shift target tracking based on complex feature fusion[J]. Control and Decision, 2014(7): 1297-1300. [本文引用:1]
[8]	王田, 刘伟宁, 韩广良, 等. 基于改进MeanShift的目标跟踪算法[J]. 液晶与显示, 2012, 27(3): 396-400. Wang Tian, Liu Wei-ning, Han Guang-liang, et al. Target tracking algorithm based on improved Meanshift[J]. Chinese Journal of Liquid Crystals and Displays, 2012, 27(3): 396-400. [本文引用:1]
[9]	许佳佳. 基于MeanShift算法的航空影像联合分割[J]. 液晶与显示, 2014, 29(4): 586-591. Xu Jia-jia. Joint segmentation using MeanShift algorithm for high resolution aerial images[J]. Chinese Journal of Liquid Crystals and Displays, 2014, 29(4): 586-591. [本文引用:1]
[10]	夏楠, 邱天爽, 李景春, 等. 一种卡尔曼滤波与粒子滤波相结合的非线性滤波算法[J]. 电子学报, 2013, 41(1): 148-152. Xia Nan, Qiu Tian-shuang, Li Jing-chun, et al. A nonlinear filtering algorithm combining Kalman filter with particle filter[J]. Journal of Electronics, 2013, 41(1): 148-152. [本文引用:1]
[11]	杜超, 刘伟宁, 刘恋. 一种基于卡尔曼滤波及粒子滤波的目标跟踪算法[J]. 液晶与显示, 2011, 26(3): 384-389. Du Chao, Liu Wei-ning, Liu Lian. Target tracking algorithm based on Kalman filter and particle filter[J]. Chinese Journal of Liquid Crystals and Displays, 2011, 26(3): 384-389. [本文引用:1]
[12]	Mei X, Ling H. Robust visual track-ing using ℓ 1 minimization[C]∥IEEE International Conference on Computer Vision, Xi'an, China, 2009: 1436-1443. [本文引用:1]
[13]	张亚红, 杨欣, 沈雷, 等. 基于视觉显著特征的自适应目标跟踪[J]. 吉林大学学报: 信息科学版, 2015, 33(2): 195-200. Zhang Ya-hong, Yang Xin, Shen Lei, et al. Adaptive target tracking based on salient features of vision[J]. Journal of Jilin University (Information Science Edition), 2015, 33(2) : 195-200. [本文引用:1]
[14]	李静宇, 王延杰. 基于子空间的目标跟踪算法研究[J]. 液晶与显示, 2014, 29(4): 617-622. Li Jing-yu, Wang Yan-jie. Subspace based target tracking algorithm[J]. Chinese Journal of Liquid Crystals and Displays, 2014, 29(4): 617-622. [本文引用:1]
[15]	Zhang K, Zhang L, Yang M H. Real-time Compressive Tracking[M]. Florence, Italy: Springer, 2012: 864-877. [本文引用:1]
[16]	杨欣, 刘加, 周鹏宇, 等. 基于多特征融合的粒子滤波自适应目标跟踪算法[J]. 吉林大学学报: 工学版, 2015, 45(2): 533-539. Yang Xin, Liu Jia, Zhou Peng-yu, et al. Particle filter adaptive target tracking algorithm based on multi-featurefusion[J]. Journal of Jilin University(Engineering and Technology Edition), 2015, 45(2): 533-539. [本文引用:1]
[17]	王暐, 王春平, 李军, 等. 特征融合和模型自适应更新相结合的相关滤波目标跟踪[J]. 光学精密工程, 2016, 24(8): 2059-2066. Wang Wei, Wang Chun-ping, Li Jun, et al. Correlation filter tracking based on feature fusing and model adaptive updating[J]. Optics and Precision Engineering, 2016, 24(8): 2059-2066. [本文引用:1]
[18]	Kalal Z, Mikolajczyk K, Matas J. Tracking-learning-detection[J]. IEEE Trans Pattern Anal Mach Intell, 2012, 34(7): 1409-1422. [本文引用:2]
[19]	Kalal Z, Mikolajczyk K, Matas J. Forward-backward error: automatic detection of tracking failures[C]∥International Conference on Pattern Recognition, Japan, 2010: 2756-2759. [本文引用:1]
[20]	周鑫, 钱秋朦, 叶永强, 等. 改进后的 TLD 视频目标跟踪方法[J]. 中国图象图形学报, 2013, 18(9): 1115-1123. Zhou Xin, Qian Qiu-meng, Ye Yong-qiang, et al. The improved TLD video target tracking method[J]. Journal of Image and Graphics, 2013, 18(9): 1115-1123. [本文引用:1]

2014

0.0

... 传统的跟踪方法可分为两类:基于跟踪器的跟踪算法和基于检测器的跟踪算法,前者主要在于目标特征的选取,代表性的特征包括目标的质点^[1]、目标的轮廓^[2,3,4]、链式模型^[5]和光流^[6] ...

2016

0.0

2011

0.0

. 2011, 4(5):480-488 DOI:doi:10.3969/j.issn.2095-1531.2011.05.012

Application of multi-pattarn tracking in wheel type scout car image processor

多模跟踪技术在轮式侦察车图像处理器的应用

Hao Zhi-cheng , Gao Wen.

郝志成, 高文

As variety illumination, deformation and rotation of targets and many other complicated conditions are always hard to be cracked in target tracking, this paper researches and improves particle filtering and Scale Invariant Feature Transform(SIFT) algorithms. By combined with the two algorithms, it proposes a multi-pattern tracking technique, which adopts a particle filter to predict target position and then chooses SIFT character matching to get the accurate position of an object. This target tracking is quite robust and has applied to an image processor in a artillery wheel type scout car. Experiments has been performed in many aspects and results prove it is accetable and effective.

为解决目标旋转形变、遮挡、光照变化等目标跟踪的难题,对粒子滤波和尺度不变特征变换(SIFT)算法进行了改进,结合两种算法提出了决策主导模式的多模跟踪技术。该技术采用粒子滤波预测目标位置进行粗定位,SIFT特征匹配进行精定位的方法,在解决上述难题上有很好的鲁棒性。将该技术应用于轮式侦察车图像处理器,并进行了各种实验验证,结果证明了提出算法的有效性。

2017

0.0

. 2017, 25(12):3152-3159

Design of tracker for mobile smart devices

适应移动智能设备的目标跟踪器

Xiong Jing-ying , Dai Ming.

熊晶莹, 戴明

To apply augmented reality technology in mobile smart devices, a novel target tracker for the embedded system of a smart mobile device was proposed. In the stage of feature description, the binary feature was segmented rapidly and discriminative binary descriptors were established by fast brightness segment. In the stage of feature matching, a sparse searching template was proposed to improve the execution efficiency of the tracking algorithm. Initial target information was stored in a static library established by the tracker and the change of target appearance was stored in a dynamic library, and the target location was tracked by comparing search information and the templates in libraries. The execution time of the tracker was compared, and the results show that sparse search templates significantly improve the execution time of the algorithm at maintaining a higher tracking accuracy, and setting the search radius to 10 is able to meet the real-time requirements of trackers. The effective capacity of the tracker was compared also. And the results indicate that the tracking error of DBRISK are 16%, 28% and 29% decrease than those of the original BRISK (Binary Robust Invariant Scalable Keypoints) under three different search radii, which means the proposed method significantly improves the accuracy of tracking method, and is applicable to the limited computing power smart mobile devices.

2008

0.0

2016

0.0

... 在此背景下,研究者提出了大量的解决方案,例如:Mean-shift^[7,8,9]、粒子滤波^[10,11]、L1跟踪器^[12]、视觉显著性特征^[13,14]、实时压缩感知跟踪 (Real-time compressive tracking)^[15] ...

2012

0.0

. 2012, 27(3):396-400

Target tracking algorithm based on improved Meanshift

基于改进MeanShift的目标跟踪算法

Wang Tian , Liu Wei-ning , Han Guang-liang

王田, 刘伟宁, 韩广良

2014

0.0

. 2014, 29(4):586-591 DOI:doi:10.3788/YJYXS20142904.0586

Joint segmentation using MeanShift algorithm for high resolution aerial images

基于MeanShift算法的航空影像联合分割

Xu Jia-jia.

许佳佳

In order to achieve exact and consistent segmentation result using two-phase high resolution aerial images of the same area which contain some changed information,a joint segmentation method is proposed. First,the two-phase images are manipulated using Guassian smooth filter to suppress the object's internal diversity and avoid over-segmentation. Then principal component analysis is carried out on the combined image contained all bands of two-phase images to eliminate the redundant information. And the first principal component will be as gray component. Finally,texture is extracted from two-phase origin images as texture component,and meanshift algorithm will be implemented on the image combined gray component with two-phase texture components. Experimental results indicate that the proposed method can use the image information effectively and save much processing time. It can obtain better segmentation result.

为了实现在包含部分变化信息的相同区域不同时相的高分辨率航空影像上获得准确一致的分割结果，提出了两时相高分辨率航空影像联合分割方法。首先，对两时相影像进行高斯平滑处理，减小地物的内部差异以避免过分割；然后将两时相影像进行波段组合，采用主成分分析法剔除冗余数据，取其第一主分量作为灰度分量；最后对原始影像进行纹理分析，获得两时相影像的纹理信息作为纹理分量并与灰度分量组合在一起进行MeanShift分割。实验对比结果表明，该方法能够有效利用数据，节省处理时间，获得了较好的分割结果。

2013

0.0

. 2013, 41(1):148-152 DOI:doi:10.3969/j.issn.0372-2112.2013.01.026

A nonlinear filtering algorithm combining Kalman filter with particle filter

一种卡尔曼滤波与粒子滤波相结合的非线性滤波算法

Xia Nan , Qiu Tian-shuang , Li Jing-chun

夏楠, 邱天爽, 李景春

A nonlinear filtering algorithm is proposed based on the Kalman filter and the particle filter.The method can provide significant performance for dynamic nonlinear system which is consist of linear state equation and nonlinear measurement equation.Firstly,the particle filter is utilized for initial estimation of the state variables,and then the Kalman filter is performed.The Cramer-Rao Bound is derived for the nonlinear model.Computation complexity analysis and numerical simulations demonstrate that the proposed algorithm has the same complexity as the standard particle filter,but the estimation accuracy is higher than the standard particle filter and the extended Kalman filter.The estimation error is even lower than the Cramer-Rao Bound of the system model.

提出一种基于卡尔曼滤波与粒子滤波的非线性滤波算法.这种方法对于状态变量服从线性变化而观测方程为非线性的动态系统模型具有显著的效果.首先使用粒子滤波对状态变量进行初估计,然后对估计结果进行卡尔曼滤波,另外推导出该系统模型下状态变量估计误差的克拉美劳下界.通过计算复杂度分析及仿真实验验证,表明新方法与标准粒子滤波算法复杂度相当,但参数估计精度要高于标准粒子滤波以及扩展卡尔曼滤波算法,估计误差甚至要低于系统模型的克拉美劳下界.

2011

0.0

. 2011, 26(3):384-389

Target tracking algorithm based on Kalman filter and particle filter

一种基于卡尔曼滤波及粒子滤波的目标跟踪算法

Du Chao , Liu Wei-ning , Liu Lian.

杜超, 刘伟宁, 刘恋

2009

0.0

2015

0.0

. 2015, 33(2):195-200 DOI:doi:10.3969/j.issn.1671-5896.2015.02.013

Adaptive target tracking based on salient features of vision

基于视觉显著特征的自适应目标跟踪

Zhang Ya-hong , Yang Xin , Shen Lei

张亚红, 杨欣, 沈雷

In order to solve the singular describing of candidate target in object tracking, an algorithm based on adaptive feature fusion of visual significant feature is proposed. The quaternion model of target composed of intensity, motion, color and the change of color feature is built, and the PQFT (Phase spectrum of Quaternion Fourier Transform) is used to extract saliency map. According to similarity coefficient, the fusion weights are adaptively adjusted and the color feature is combined to ensure accurate tracking. The experimental results demonstrate that the approach can effectively overcome the part occlusion and the interference of background, realizing the accurate tracking under the case of complex background.

为解决运动目标跟踪过程中候选目标信息描述单一的问题, 提出一种基于视觉显著性特征融合的自适应目标跟踪算法。提取目标颜色、颜色的变化、强度和运动信息构建目标四元数模型, 采用相位谱重建算法检测目标的显著图(Saliency Map), 并根据特征相似度大小自适应调整权值, 融合视觉显著性特征和颜色特征实现目标跟踪。实验结果表明, 该算法能有效克服部分遮挡和背景融合干扰, 从而实现复杂背景下目标的准确跟踪。

2014

0.0

. 2014, 29(4):617-622 DOI:doi:10.3788/YJYXS20142904.0617

Subspace based target tracking algorithm

基于子空间的目标跟踪算法研究

Li Jing-yu , Wang Yan-jie.

李静宇, 王延杰

In terminal guidance tracking phase,the missile's flight attitude,the distance from missile to the target and target's moving status and appearance change a lot. Using a single fixed template cannot achieve stable tracking. In this paper,a new subspace-based moving target tracking algorithm is proposed. First,a set of orthogonal sparse subspace feature vectors are used to describe target model. Then,using incremental method to continuously update the subspace model to accommodate the target changes in appearance which is caused by intrinsic and extrinsic factors,therefore,the tracking accuracy is improved. Finally,importance sampling algorithm and maximum likelihood estimation are applied to solve complex optimization problems. Experimental results show that when the camera and the background suffer larger relative motion and target status changed dramatically,the algorithm can still be able to achieve sustained and stable tracking,and the average tracking error keeps less than 10 pixels. It can satisfy the system requirements of stability and robustness for terminal guidance tracking.

末制导跟踪阶段，导弹的飞行姿态，弹体与目标的距离以及目标自身的运动姿态和形态均会发生较大的变化，采用单一固定模板无法实现稳定跟踪。本文提出一种新的基于子空间的运动目标跟踪算法，首先采用一组正交的稀疏子空间特征向量表示目标模型，然后采用增量方法不断更新子空间模型，以适应由于目标内在和外在因素所造成的在外观上的变化，从而提高跟踪精度；采用重要性采样算法以及最大似然估计，解决复杂的优化问题。实验结果表明，当摄像机与背景发生较大相对运动以及目标姿态发生剧烈变化时，仍然能够实现对目标的持续稳定跟踪，平均跟踪误差小于10个像素。基本满足末制导跟踪系统的稳定性和鲁棒性等要求。

2012

0.0

2015

0.0

. 2015, 45(2):533-539

Particle filter adaptive target tracking algorithm based on multi-featurefusion

基于多特征融合的粒子滤波自适应目标跟踪算法

Yang Xin , Liu Jia , Zhou Peng-yu

杨欣, 刘加, 周鹏宇

... 多特征融合粒子滤波算法^[16,17]对目标遮挡以及背景混淆等诸多问题具有较高的鲁棒性,但当视频序列的尺寸较大或者背景较复杂需要数量较多的粒子时,算法处理的时间较长,会导致实时性变差 ...

2016

0.0

. 2016, 24(8):2059-2066

Correlation filter tracking based on feature fusing and model adaptive updating

特征融合和模型自适应更新相结合的相关滤波目标跟踪

Wang Wei , Wang Chun-ping , Li Jun

王暐, 王春平, 李军

2012

0.0

... 为实现持久的目标跟踪,Kalal等^[18]提出了一种新的TLD (Tracking-learning-detection)跟踪算法,采用跟踪器与检测器并行处理的框架,并通过学习模块不断更新检测器相应参数值,最终由整合模块输出跟踪结果 ...

... 1 TLD跟踪算法传统TLD算法中跟踪模块主要是用中值流跟踪算法^[18],中值流跟踪算法就是在光流法的基础上加了个限制,一个#cod#x0201c ...

2010

0.0

... 如过跟踪器跟踪失败,由检测器修正跟踪器的目标模型^[19] ...

2013

0.0

. 2013, 18(9):1115-1123 DOI:10.11834/jig.20130908

The improved TLD video target tracking method

TLD 视频目标跟踪方法

Zhou Xin , Qian Qiu-meng , Ye Yong-qiang

周鑫, 钱秋朦, 叶永强

As an effective visual target tracking algorithm, the tracking-learning-detection(TLD) has drawn wide attention around the world. In this paper, we propose an improved TLD visual target tracking algorithm, which is obtained by making several improvements based on the original TLD algorithm. The improvements include modifying local tracker placement as well as local tracker failure predicting method for the tracker of TLD to improve the precision and robustness of the tracker; employing Kalman filter in the detector of TLD for estimating the location of the target to reduce the scanning region of the detector and improve the speed of the detector; adding Markov model based target moving direction predictor in the detector of TLD to increase the discretion for targets with similar appearance. Experiments have been conducted to compare the performance of the original TLD and improved TLD. The experimental results show that, compared with the original TLD, the improved TLD has more accurate tracking precision, faster tracking speed and stronger distinguishing ability for targets with similar appearance in the scene.

TLD(tracking-learning-detection)是近期受到广泛关注的一种有效的视频目标跟踪算法。在原始TLD的基础上,对其进行改进,改进包括:在TLD的跟踪器中对其局部跟踪器的布置和局部跟踪器的跟踪成败预测方法进行改进,提高跟踪器的跟踪精度和鲁棒性;在TLD的检测器中引入基于Kalman滤波器的当前帧目标所在区域预估,缩小了检测器的检测范围,提高了检测器处理速度;在TLD的检测器中加入基于马尔可夫模型的方向预测器,增强了检测器对相似目标的辨识能力。通过实验对原始TLD和改进后的TLD进行了比较,实验结果显示改进后的TLD算法较原始TLD算法具备更高的跟踪精度和更快的处理速度,而且增强了对场景中相似目标的辨识能力。

... 根据上述分析,本文提出了一种改进的TLD目标跟踪算法,用改进L1APG(L1 tracker using accelerated proximal gradient approach)^[20]跟踪器来取代传统TLD算法中基于光流法的跟踪器,同时加入遮挡检测判断,设定遮挡阈值,并根据遮挡程度实时更新目标模板和背景模板系数权重,进一步解决遮挡问题,从而抑制漂移问题 ...