基于视频序列的目标跟踪算法优化

发布时间：2018-08-14 18:20

【摘要】：视频序列的目标跟踪是计算机视觉领域的一个关键技术,是高级机器视觉的研究基础,有着广泛的应用。由于目标外观变化、复杂的环境和目标尺度变化等干扰的存在,现有的目标跟踪算法在实际的应用中仍存在跟踪漂移、跟踪目标丢失等问题。针对这些问题,本文对基于区域特征的均值漂移跟踪算法和压缩跟踪算法进行研究和优化以提高算法的高效性,主要工作如下：在均值漂移算法的基础上,针对均值漂移跟踪算法易受相似颜色的背景干扰的问题,提出了一种双权值算法,结合背景削弱权值与中心加权权值,使算法对目标的建模更加准确,提高了算法的抗干扰能力；针对均值漂移跟踪算法缺少目标模板更新模块的问题,提出了一种模板更新策略,解决了原始算法在跟踪过程中受遮挡、光照、形变等不确定因素的影响时导致的跟踪漂移现象,保证了跟踪的稳定性；为解决均值漂移跟踪算法跟踪窗大小固定不变,在长时间跟踪中导致目标定位偏差甚至跟踪失败的问题,利用CAMShift算法中的尺寸更新方法,通过反向投影图的零阶矩和一阶矩更新跟踪窗尺寸,有效地适应了跟踪过程中目标尺寸的变化。在压缩感知理论的基础上,针对遮挡、外观变化和目标移动速度对压缩跟踪算法的影响,提出一种样本选择方法和学习因子更新策略。根据正样本与目标中心的距离,分配正样本不同的权值以增强样本的差异性,同时通过前后两帧图像的巴氏系数关系实时更新模型参数的学习因子,提高了算法的跟踪稳定性。实验表明,优化的算法在满足实时性的基础上增强了跟踪的有效性。
[Abstract]:Video sequence target tracking is a key technology in the field of computer vision. It is the research foundation of advanced machine vision and has a wide range of applications. Aiming at these problems, this paper studies and optimizes the mean shift tracking algorithm based on region feature and the compression tracking algorithm to improve the efficiency of the algorithm. The main work is as follows: Based on the mean shift algorithm, aiming at the problem that the mean shift tracking algorithm is susceptible to the background interference of similar colors, a new algorithm is proposed. Dual-weight algorithm combines background weakening weight and central weighting weight to make the algorithm more accurate in modeling the target and improve the anti-jamming ability of the algorithm. Tracking drift caused by uncertain factors such as deformation guarantees the stability of tracking. In order to solve the problem that the tracking window size of mean shift tracking algorithm is fixed, which leads to target positioning deviation and even tracking failure in long-term tracking, the size updating method of CAMShit algorithm is used to solve the problem by using the back projection graph. Zero-order moment and first-order moment update the tracking window size, which effectively adapts to the change of target size in the tracking process. Based on compressed sensing theory, a sample selection method and a learning factor update strategy are proposed according to the influence of occlusion, appearance change and moving speed of target on the compressed tracking algorithm. Distance is allocated to different weights of positive samples to enhance the diversity of samples. The learning factors of model parameters are updated in real-time by the relationship between the Barkhausen coefficients of the previous two frames to improve the tracking stability of the algorithm.
【学位授予单位】：中北大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】