基于深度信息的核相关滤波目标跟踪算法研究

发布时间：2018-03-12 18:25

本文选题：目标跟踪　切入点：深度信息　出处：《哈尔滨工业大学》2017年硕士论文　论文类型：学位论文

【摘要】：视觉目标跟踪(Visual Object Tracking)在智能监控、人机交互、无人驾驶、虚拟现实等领域有非常重要的应用价值,一直以来都是计算机视觉领域的研究热点。近几年,相关滤波跟踪算法和基于深度学习的跟踪算法的提出显著提升了跟踪速度和精度。但是在目标被遮挡、目标尺度变化或处于复杂背景等情况下,尤其是目标被长时间或者严重遮挡的时候,如何准确跟踪目标仍是困扰着研究者们的难题。遮挡是三维世界投影到二维平面的信息损失导致的,因此出现一些使用深度信息的二维或三维跟踪方法,借助三维空间结构来帮助解决这些难题,并取得了一定的进展。由于近几年深度传感器精度的提升和价格的下降,使得深度信息的获取变得容易。此外还有无人机、机器人、无人驾驶汽车等越来越多的设备携带有深度传感器,因此在与这些设备相关的应用场景下使用深度信息协助目标跟踪有重要的研究和应用价值。目前基于深度信息的跟踪算法分为两类,其中二维跟踪方法不能有效的使用深度信息,没有把深度信息与已有的跟踪算法深度融合。而三维跟踪方法由于缺乏比较成熟的三维特征提取技术,目标的三维表观模型并不鲁棒,进而影响其跟踪效果。本文针对上述问题提出使用自适应量化的深度信息,根据不同场景建立相适应的分层结构,一方面过滤前景和背景信息减少跟踪的干扰因素,结合成熟的图像特征提取技术,包括HOG特征和颜色属性直方图特征等,建立目标鲁棒的目标表观模型;另一方面这样的分层结构简化了深度信息的使用方法,使得处理目标尺度变化以及检测遮挡更为容易。在分层结构的基础上,提出在取样之前完成目标尺度估计的策略,以及快速检测遮挡的策略。结合核相关滤波(Kernel Correlation Filter)跟踪算法实现了使用二维表观模型在空间结构下的跟踪方法,能够有效应对遮挡和处理目标尺度变化。本文参加普林斯顿跟踪测评,该测评的数据集有100个跟踪视频序列,包含多种目标、多种遮挡情况和多个不同场景。最终实现的跟踪算法在基于RGB-D图像分组中排名第四,验证了所提算法的有效性。
[Abstract]:Visual Object tracking has very important application value in the fields of intelligent monitoring, human-computer interaction, driverless, virtual reality and so on, and has always been the research hotspot in the field of computer vision. Correlation filter tracking algorithm and depth learning based tracking algorithm can improve the tracking speed and precision significantly. However, when the target is occluded, the target scale changes or is in a complex background, etc. Especially when the target is occluded for a long time or severely, how to track the target accurately is still a difficult problem for researchers. Occlusion is caused by the loss of information projected into the two-dimensional plane by the three-dimensional world. Therefore, some 2D or 3D tracking methods using depth information have been developed to help solve these problems with the help of three-dimensional spatial structure, and some progress has been made. It makes it easier to access depth information. And more and more devices, such as drones, robots, driverless cars, are carrying depth sensors. Therefore, using depth information to assist target tracking in application scenarios related to these devices has important research and application value. At present, depth information based tracking algorithms can be divided into two categories. The 2D tracking method can not use depth information effectively and does not fuse the depth information with the existing tracking algorithms. However, the 3D tracking method lacks the mature 3D feature extraction technology. The 3D apparent model of the target is not robust, which affects the tracking effect. In this paper, the adaptive quantization depth information is used to establish the appropriate hierarchical structure according to different scenes. On the one hand, filtering foreground and background information to reduce the interference factors of tracking, combined with mature image feature extraction technology, including HOG features and color attributes histogram features, to establish a robust target model; On the other hand, this kind of layered structure simplifies the use of depth information, makes it easier to process the change of target scale and detect occlusion. On the basis of stratified structure, the strategy of completing target scale estimation before sampling is proposed. Combined with Kernel Correlation filter (Kernel Correlation filter) tracking algorithm, the tracking method using two-dimensional apparent model in spatial structure is realized. This paper participates in the Princeton tracking Evaluation, which has 100 tracking video sequences, including a variety of targets. Finally, the tracking algorithm is ranked 4th in the image grouping based on RGB-D, which verifies the effectiveness of the proposed algorithm.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】