基于Meanshift算法的视频跟踪分析与改进

发布时间：2018-05-17 18:25

本文选题：视频目标跟踪 + Mean　；参考：《太原科技大学》2017年硕士论文

【摘要】：视频目标跟踪是机器视觉领域中一个十分重要且极具挑战的研究方向,在智能视频监控、智能交通、气象分析、远距离测量以及医学图像分析等方面有着广泛的应用。然而在实际应用中,由于目标外观和周围环境不断地变化,令大多数现有跟踪算法在复杂环境下无法实现高性能跟踪。因此,不断提高目标跟踪算法在跟踪过程中的稳定性和准确性是现阶段这一领域的研究热点和重点改进方向。本文主要对动态场景下目标的跟踪问题进行了研究。该问题的关键在于检测和跟踪刚性物体以及非刚性物体的变化,并对目标进行时实跟踪。由于在实际的复杂环境中,目标存在不同尺度的变化,以及相似目标会出现跟踪错误等情况,而基于Mean Shift核函数的方法能够对低维度数据在高维度进行有效区分。因而本文主要阐述Mean Shift算法在目标跟踪中存在的问题。研究了相似目标在复杂环境中出现漂移的问题,提出在核函数的基础上加入了色彩空间距离的方法,实现了色彩距离在高维空间的有效区分,用以解决背景干扰引起的跟踪不准确以及实时性问题,与传统方法相比,能够更精确的进行视频跟踪。针对实际场景中机动目标存在的尺度变化问题,在核函数的基础上融入色彩空间距离,利用巴氏距离来检测目标大小的变化,实现了核函数窗口的自适应性,从而解决了由于跟踪目标大小变化引起的跟踪不精确和不稳定问题。针对视频跟踪算法是否需要时实处理,是否具有可移植性问题,在VS2013下对OpenCV2.4.9进行配置,并完成了系统软件平台的搭建,利用这一平台对本文提出的跟踪算法进行实现及结果分析,通过与传统的Mean Shift算法进行比较,可以分析得出,将色彩空间距离和巴氏距离的优势相结合应用到视频跟踪后,可有效提升在复杂场景下目标尺度变化引起的不准确性。最终本文实现了在复杂背景下高准确度的跟踪,同时做到了核函数窗口大小的自适应调整,通过实验的对比,说明跟踪过程的实时性和准确性得到进一步提升,并在OpenCV平台中稳定运行,为其在实际应用打下了坚实的基础。
[Abstract]:Video target tracking is a very important and challenging research field in the field of machine vision. It is widely used in intelligent video surveillance, intelligent transportation, meteorological analysis, remote measurement and medical image analysis. However, in practical applications, most of the existing tracking algorithms are unable to achieve high performance tracking in complex environments due to the continuous changes in the appearance of the target and the surrounding environment. Therefore, improving the stability and accuracy of target tracking algorithm in the process of tracking is a hot research topic and an important improvement direction in this field. In this paper, the problem of target tracking in dynamic scene is studied. The key to this problem is to detect and track the changes of rigid and non-rigid objects and to track the target in real time. Due to the fact that the target has different scales in the complex environment, and the similar target will have tracking errors, the method based on Mean Shift kernel function can effectively distinguish the low-dimensional data from the high-dimensional data. Therefore, this paper mainly describes the problem of Mean Shift algorithm in target tracking. In this paper, the problem of the drift of similar targets in complex environment is studied, and the method of adding color space distance based on kernel function is put forward to realize the effective distinction of color distance in high dimensional space. In order to solve the problem of inaccuracy and real-time caused by background interference, video tracking is more accurate than traditional methods. Aiming at the problem of scale change of maneuvering target in actual scene, the color space distance is incorporated into the kernel function, and the change of target size is detected by using the pasteurian distance, and the self-adaptability of kernel function window is realized. Thus, the tracking inaccuracy and instability caused by the change of target size are solved. Aiming at whether the video tracking algorithm needs real processing and portability, the OpenCV2.4.9 is configured under VS2013, and the system software platform is built. By using this platform, the tracking algorithm proposed in this paper is implemented and the results are analyzed. By comparing with the traditional Mean Shift algorithm, it can be concluded that the advantages of color space distance and pasteurian distance are applied to video tracking. It can effectively improve the inaccuracy caused by the change of target scale in complex scene. Finally, this paper realizes the tracking with high accuracy in complex background and adaptively adjusts the window size of kernel function. Through the comparison of experiments, it shows that the real-time and accuracy of the tracking process are further improved. And run stably in OpenCV platform, lay a solid foundation for its practical application.
【学位授予单位】：太原科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】