目的 视觉目标跟踪是计算机视觉研究的基本问题之一。近年来,基于相关滤波和孪生神经网络的两类判别式目标跟踪方法研究取得了较大进展,但后者计算量过大,完全依赖GPU加速运算。传统相关滤波方法由于滤波模型采用固定更新间隔,难以兼顾快速变化目标和一般目标。针对这一问题,提出一种基于目标外观状态分析的动态模型更新算法,优化计算负载并提高跟踪精度,兼顾缓变目标的鲁棒跟踪和快速变化目标的精确跟踪。方法 本文通过帧间信息计算并提取目标区域图像的光流直方图特征,利用支持向量机进行分类从而判断目标是否处于外观变化状态,随后根据目标类别和目标区域图像的光流主分量幅值动态设置合适的相关滤波器更新间隔。此外,通过在首帧进行前背景分离运算,进一步增强对目标外观表征的学习,提高跟踪精度。 结果 在OTB100基准数据集上同其他6种快速跟踪算法进行对比实验,本文算法的精准度和成功率(AUC)分别为86.4%和64.9%,分别比第二名ECO-HC算法高出1.4%和0.9%。在平面内旋转、遮挡、部分超出视野和光照变化这些极具挑战性的复杂环境下精准度分别比第二名高出3.0%、4.4%、5.2%和6.0%,成功率(AUC)高出1.9%、3.1%、4.9%和4.0%。本文算法在CPU上的运行速度为32.15帧/秒,满足跟踪问题实时性的要求。结论 本文的自适应模型更新算法在优化计算负载的同时取得了更好的跟踪精度,适合工程部署与应用。
Abstract: Objective Visual object tracking is one of the basic problems in computer vision research, which has profound theoretical basis and application value. With wide applications, visual object tracking technology faces increasingly complex environments. Factors such as scale changes, occlusion and illumination variation bring more uncertain interferences to visual tracking. There is still a lot of research space for robust, accurate and fast visual object tracking algorithms. In recent years, two categories of discriminant model methods based on discriminative correlation filter and siamese neural network have achieved higher accuracy and robustness in the tracking problem. However, the tracking methods based on siamese network are limited by the huge computation amount of convolutional neural network (CNN), and can only be performed on high-performance GPUs. The computing requirement seriously affects the application of this kind of methods in practical engineering environment. On the other hand, the tracking methods based on discriminative correlation filter, due to their simple frameworks, can use manually-setting features to learn and update object’s representation, and achieve real-time tracking on a single CPU. This kind of real-time tracking algorithm has been well applied in mobile platforms such as UAVs. Under the traditional correlation filtering framework, updating the correlation filter frame by frame will lead to too large computational load and affect the real-time performance. The sparse model update strategy proposed in recent years simply sets a fixed update interval, which reduces the convergence speed of the tracking model and easily lead to lose track when object changes rapidly. The tracking ability of these two kinds of correlation filtering tracking algorithms cannot meet the increasing application requirements in complex environments. For the correlation filter updating strategy, this paper proposes a dynamic updating algorithm based on appearance representation analysis to optimize computation and improve tracking accuracy. Method Firstly, using optical flow features to estimate the appearance state of the object. We calculate dense optical flow of the predicted target region’s image. When the object is simply shifting, the optical flow’s amplitude of each pixel is quite small and the direction lacks uniform rule because the image of the target area changes little. However, when the object is deforming or being occluded, the deformed part will generate a much larger optical flow, which is different from the common objects. In this paper, optical flow histogram information is extracted by dividing the image into m×n grids. The average optical flow amplitudes and angles of each pixel are counted in each grid to form the histogram feature vector. Then, using a support vector machine to classify feature vectors to estimate object’s current appearance state. After appearance state analysis, counting the optical flow amplitude in the object region of the current frame and constructing a statistical histogram of optical flow amplitude with an interval of 0.5. The updating interval of the filter model is set respectively according to the magnitude of the main optical flow amplitude and the target category, so that the adaptive updating of the correlation filter is realized. In addition, the foreground-background separation operation based on discrete cosine transform in the first frame is used to obtain more accurate labeling information, reduce similar background interference and further optimize the learning of object representation. Result This algorithm is tested on OTB100 data set and compared with ECO-HC, SRDCF, Staple, KCF, DSST and CSK, which are fast tracking algorithms. First of all, on five typical challenging video image sequence, the algorithm in this paper achieved higher tracking overlap through the update interval adaptively setting model. It solved traditional frame by frame update algorithm such as Staple’s overfitting problem, and also solved ECO-HC’s fixed interval sparse update strategy’s limitation that easy to lost rapidly changing object. The comprehensive quantitative analysis results on the whole OTB100 data set showed that the tracking accuracy and success rate of the algorithm proposed in this paper are respectively 86.4% and 64.9%. Compared with other fast-tracking algorithms that can run on a CPU, the tracking accuracy and robustness of our algorithm are both the best. In addition, under some very challenging and complex environments including in-plane rotation, occlusion, out of view and illumination variation, our algorithm’s precision was respectively 3.0%, 4.4%, 5.2% and 6.0% higher than that of the second-place, and success rate was 1.9%, 3.1%, 4.9% and 4.0% higher. In the running speed test on CPU i7-6850k, the FPS of the algorithm in this paper is 32.15, and the computational load is less than the fixed frame-by-frame update algorithm, which meets the real-time requirements of tracking problems. Conclusion This paper proposed a dynamic update correlation filter tracking algorithm based on appearance representation analysis. A series of comparison results showed that the improved algorithm in this paper can take into account both the robust tracking of slow-changing objects and the accurate tracking of fast-changing objects to achieve excellent real-time performance, which is suitable for project deployment and application.