基于多特征融合的粒子滤波目标跟踪算法的研究

发布时间：2018-04-23 15:37

本文选题：目标跟踪 + 粒子滤波　；参考：《吉林大学》2017年硕士论文

【摘要】：计算机视觉包含众多领域,运动目标跟踪已经成为该领域备受欢迎的研究方向和研究热点。应用范围和应用领域非常之多,例如,智能监控,人机交互,机器人导航,流量控制,生物医疗诊断等。到目前为止,虽然国内外专家学者已经提出很多经典的算法并且在此基础上的改进算法,但是这些目标跟踪算法在实际的应用中仍然面临着巨大的挑战,比较常见的有光照变化,遮挡,目标突然移动所导致的非线性形变,背景与跟踪目标相似度很高以及复杂背景下的噪声干扰等因素对目标跟踪算法速度和准确性的影响。可能造成难以估计的后果。考虑遇到的这些问题,实现一种能够综合适应复杂场景的目标跟踪算法还是比较艰巨和困难的。目前,粒子滤波是解决非高斯非线性跟踪问题的最佳方法。该算法能够很好地适应多种外界干扰因素的影响,能够最大程度上保证目标跟踪的准确性和鲁棒性。(1)然而经典的粒子滤波算法采用单一特征来描述待跟踪的目标,在跟踪过程中容易受到光照,遮挡,目标与背景相似等因素的干扰。针对这个问题,本文将多种目标特征融合到粒子滤波中用于目标跟踪。在粒子滤波框架下,综合考虑各种特征对不同干扰因素的鲁棒性和准确性,选择合适的特征提取算法融合其中。运用动态计算的方法计算不同特征对目标和背景的区分度和稳定性,自动选择区分度高,稳定性好的特征来表征目标,形成多特征融合目标模型。使用度量特征的不确定性方法来动态调整所提取的目标特征所占的比重。利用粒子滤波算法本身的颜色特征适应目标的形变和尺度的变化,利用边缘特征信息来适应背景的变化以及利用局部二值模式(LBP)特征与图像帧的灰度信息相结合来适应光照的变化,使得改进后的算法稳定性和准确性更好。(2)尽管上面所提出的算法能够很好地适应大多数外界环境因素的变化。但是跟踪视频中出现目标遮挡时,容易出现目标跟踪偏移或者目标跟踪丢失问题。因此本文将目标的上下文信息融入粒子滤波框架中来解决目标遮挡的问题。由于连续帧图像中待跟踪目标与其周围局部场景的上下文关系是保持相似的。因此可以通过分析上一帧中目标的上下文信息来估计下一帧中目标的位置。在目标存在遮挡时实现快速、鲁棒性跟踪。最后,在公共图像数据集上对本文提出的两种目标跟踪算法与其他相关目标跟踪算法的跟踪性能进行对比。通过对实验结果进行分析,得出结论:在应对光照变化,遮挡,形变,目标与背景相似以及噪声干扰方面,提出的改进算法很大程度上提升了跟踪准确性和鲁棒性。综上所述,本文的算法研究可以极大地丰富目标跟踪领域的理论研究并且能够满足部分计算机视觉领域的现实应用需求。
[Abstract]:Computer vision includes many fields, and moving target tracking has become a popular research direction and research hotspot in this field. There are many applications, such as intelligent monitoring, human-computer interaction, robot navigation, flow control, biomedical diagnosis and so on. Up to now, although experts and scholars at home and abroad have proposed a lot of classical algorithms and improved algorithms on this basis, these target tracking algorithms are still facing enormous challenges in practical applications. Some common factors, such as illumination variation, occlusion, nonlinear deformation caused by sudden moving of target, high similarity between background and tracking object and noise interference in complex background, affect the speed and accuracy of target tracking algorithm. It can have incalculable consequences. Considering these problems, it is difficult and difficult to implement a target tracking algorithm which can adapt to complex scenes. At present, particle filter is the best method to solve non-Gao Si nonlinear tracking problem. The algorithm can adapt to the influence of many kinds of external interference factors and ensure the accuracy and robustness of target tracking to the greatest extent. However, the classical particle filter algorithm uses a single feature to describe the target to be tracked. In the process of tracking, it is easy to be interfered by illumination, occlusion, similarity between target and background, and so on. In order to solve this problem, a variety of target features are fused to particle filter for target tracking. In the framework of particle filter, considering the robustness and accuracy of various features to different interference factors, the appropriate feature extraction algorithm is selected. The method of dynamic calculation is used to calculate the discrimination and stability of different features to target and background. The features with high degree of discrimination and good stability are automatically selected to represent the target and form a multi-feature fusion target model. The uncertainty method of measuring features is used to dynamically adjust the proportion of extracted target features. The color feature of particle filter algorithm is used to adapt to the deformation and scale change of the target. The edge feature information is used to adapt to the background change, and the local binary mode LBP) feature is combined with the gray level information of the image frame to adapt to the illumination change. It makes the improved algorithm more stable and accurate.) although the proposed algorithm can well adapt to the changes of most external environmental factors. However, when the target occlusion occurs in the tracking video, the target tracking offset or the target tracking loss is easy to occur. Therefore, the contextual information of the target is incorporated into the particle filter framework to solve the occlusion problem. Because the contextual relationship between the target to be tracked and the local scene around it is similar in successive frame images. Therefore, the location of the target in the previous frame can be estimated by analyzing the context information of the target in the previous frame. Fast and robust tracking is achieved when there is occlusion. Finally, the performance of the two target tracking algorithms proposed in this paper is compared with that of other related target tracking algorithms on the common image data set. Through the analysis of the experimental results, it is concluded that the proposed improved algorithm improves the tracking accuracy and robustness to a great extent in dealing with changes of illumination, occlusion, deformation, similarity between target and background, and noise interference. To sum up, the algorithm research in this paper can greatly enrich the theoretical research in the field of target tracking and can meet the practical application requirements of some computer vision fields.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】