基于联合直方图表示的多源目标融合跟踪方法研究

发布时间：2018-05-30 22:35

本文选题：直方图表示 + 均值漂移　；参考：《广西师范大学》2017年硕士论文

【摘要】：目标跟踪是完成视觉监控、人机交互、车辆导航等诸多视频场景分析和理解任务的基础,已有大量的跟踪方法被报道,可将这些方法大致分成单光谱跟踪和多光谱跟踪两大类。与单光谱目标跟踪系统相比,多光谱目标跟踪系统在生存能力、时空覆盖范围、可信度等方面都具有明显的优势,因而被广泛研究,其中最具代表的是红外与可见光的融合跟踪。红外传感器是通过检测目标辐射的热能差异来形成影像,因此在恶劣的光照环境下要比可见光成像更好,但它无法捕获目标对象的颜色和纹理特征。可见光传感器虽然无法感知温度,但在处理多个热目标交汇时,它通常要优于红外传感器,特别当目标对象间有显著的颜色和纹理差异时。因此通过联合它们的数据,能获得比使用单个传感器更好的跟踪性能。本文从目标表示和目标搜索两个方面出发,提出了以下两种基于联合直方图表示的红外与可见光目标融合算法:1.基于联合直方图表示的红外与可见光目标核跟踪算法。首先,以直方图为特征表示模型,分别计算给定候选状态下红外图像块的颜色直方图和可见光图像块的颜色直方图,并采用巴氏系数分别计算红外图像块的颜色直方图与其目标模板间的相似度,以及可见光图像块的颜色直方图与其目标模板间的相似度;然后,将两个相似度进行加权组合以形成目标函数来;接着,对目标函数进行多变量泰勒展开,得到其线性逼近式,并通过最大化该逼近式推导出一个从当前候选状态到新候选状态的目标状态转移关系式;最后,根据该转移关系式,使用均值漂移程序来递归地获得目标在当前帧中的最终状态。2.基于联合直方图表示的红外与可见光目标粒子滤波跟踪算法。首先,以前一帧的跟踪结果为初始状态,采用六参数的仿射变换模型产生高斯随机采样粒子集;然后,分别计算给定采样粒子对应的红外图像块的颜色直方图和可见光图像块的颜色直方图,并计算它们各自与其目标模板间的相似度;接着,将两相似度的加权组合作为粒子滤波跟踪器的观测似然,并启动粒子滤波跟踪程序得到其后验概率,对其余粒子重复上面的步骤,得到所有粒子的后验概率;最后,以所有粒子与其后验概率乘积的期望作为目标在当前帧中的最终状态。该算法克服了基于核的融合跟踪方法在迭代时候容易陷入局部最优的缺点,能较好地处理目标的缩放、旋转和形变等仿射运动变化。对多组的红外与可见光图像序列对的测试表明,基于核的融合跟踪算法具有较高的实时性,而基于粒子滤波的融合跟踪方法能胜任目标的仿射运动变化。两种融合跟踪方法在处理遮挡、光照变化以及目标交汇等方面都展现出较好的性能。
[Abstract]:Target tracking is the basis of many video scene analysis and understanding tasks, such as visual monitoring, human-computer interaction, vehicle navigation and so on. A large number of tracking methods have been reported. These methods can be divided into two categories: single-spectral tracking and multi-spectral tracking. Compared with single spectral target tracking system, multispectral target tracking system has obvious advantages in survivability, space-time coverage and reliability, so it has been widely studied. One of the most representative is infrared and visible light fusion tracking. Infrared sensor can produce image by detecting the thermal energy difference of target radiation, so it is better than visible light imaging in bad light environment, but it can not capture the color and texture features of target object. Although the visible light sensor can not sense the temperature, it is usually superior to the infrared sensor in dealing with the intersection of multiple thermal targets, especially when there are significant differences in color and texture between the target objects. Therefore, by combining their data, better tracking performance can be achieved than using a single sensor. In this paper, we propose two fusion algorithms of infrared and visible targets based on joint histogram representation, which are composed of two aspects: target representation and target search. Infrared and visible target core tracking algorithm based on joint histogram representation. Firstly, the color histogram of infrared image block and the color histogram of visible light image block are calculated under given candidate state by using histogram as the characteristic representation model. The similarity between the color histogram of the infrared image block and its target template and the similarity between the color histogram of the visible image block and the target template are calculated respectively by using the pasteurian coefficient; then, the similarity between the color histogram of the infrared image block and its target template is calculated. The two similarity degrees are weighted together to form the objective function, and then, the multivariable Taylor expansion of the objective function is carried out, and the linear approximation of the objective function is obtained. A target state transition formula from the current candidate state to the new candidate state is derived by maximizing the approximation. Use the mean-shift program to recursively obtain the final state of the target in the current frame. Particle filter tracking algorithm for infrared and visible targets based on joint histogram representation. First, the tracking result of the previous frame is initial state, and the affine transformation model with six parameters is used to generate the random sampling particle set of Gao Si. The color histogram of the infrared image block corresponding to the given sample particle and the color histogram of the visible image block are calculated respectively, and the similarity between them and their target template is calculated. The weighted combination of two similarity degrees is used as the observation likelihood of particle filter tracker, and the posterior probability is obtained by starting the particle filter tracking program, then repeating the above steps to the other particles to obtain the posterior probability of all particles. The expectation of the product of all particles and the posterior probability is taken as the final state of the target in the current frame. The algorithm overcomes the shortcoming that the kernel-based fusion tracking method is easy to fall into local optimum in iteration, and it can deal with the affine motion changes such as scaling, rotation and deformation of the target. The test of multiple infrared and visible image sequences shows that the fusion tracking algorithm based on kernel has high real-time performance, while the fusion tracking method based on particle filter is capable of changing the affine motion of the target. The two fusion tracking methods show good performance in the processing of occlusion, illumination change and target intersection.
【学位授予单位】：广西师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】