基于稀疏表示的多源目标融合跟踪方法研究

发布时间：2018-12-16 13:54

【摘要】：目标跟踪是计算机视觉领域的主要研究方向之—,在视频监控、军事制导、无人驾驶、人机交互等领域得到了广泛应用,深受研究者们的广泛关注。作为目标跟踪技术的一个重要分支,多源目标跟踪是通过将来自多个传感器的图像数据进行联合来完成目标跟踪的。由于它利用了各传感数据的冗余互补特性,因此能获得比单个传感器更好的跟踪性能。红外与可见光图像目标的融合跟踪是被研究最多的一种,如何高效准确的在传感器中呈现跟踪目标并分析出运动状态的变化,从而获得对实际应用有意义的信息是目前多源跟踪亟待解决的问题。针对这些问题,本文开展了基于稀疏表示的多源目标融合跟踪方法研究。主要工作如下:1、提出了基于L1-APG红外与可见光目标融合跟踪的算法。首先,该方法将稀疏表示引入融合跟踪中,建立了红外与可见光目标的联合稀疏表示模型;然后,以它们的联合重构误差最小为目标来构建L1最优化问题。并采用APG算法来求解L1问题;最后,利用最小二乘边界误差来减少粒子重采样的次数,达到降低整个算法时间复杂度的目的,从而实现了实时融合跟踪。2、提出了基于遮挡检测的红外与可见光目标融合跟踪算法。通过对目标图像进行稀疏表示来描述目标的外观模型,并在稀疏表示的基础上引入跟踪和识别同时进行的方法。为了解决目标模板更新过程中所存在的被遮挡的跟踪结果被不当地添加到参考模板集合中的问题,本文建立了遮挡检测模型来计算被遮挡区域的大小,并根据遮挡面积大小使用协同学习的方法来更新参考模型,从而降低遮挡因素对跟踪结果的影响。对多组红外与可见光图像序列对的测试结果表明,本文所提出的两种跟踪方法在处理目标交汇,目标旋转,光照变化,以及目标遮挡等方面都表现良好。
[Abstract]:Target tracking is one of the main research directions in the field of computer vision. It has been widely used in video surveillance, military guidance, unmanned driving, human-computer interaction and so on. As an important branch of target tracking technology, multi-source target tracking is accomplished by combining image data from multiple sensors. Because it takes advantage of the redundant and complementary characteristics of each sensor data, it can achieve better tracking performance than a single sensor. The fusion tracking of infrared and visible image targets is one of the most studied, how to efficiently and accurately present the tracking target in the sensor and analyze the change of the moving state. Therefore, obtaining meaningful information for practical applications is an urgent problem to be solved in multi-source tracking. In order to solve these problems, a sparse representation based multi-source target fusion and tracking method is proposed in this paper. The main work is as follows: 1. A fusion and tracking algorithm based on L1-APG infrared and visible light target is proposed. Firstly, the sparse representation is introduced into fusion tracking, and the joint sparse representation model of infrared and visible targets is established, and then the L1 optimization problem is constructed with the minimum joint reconstruction error as the target. APG algorithm is used to solve L1 problem. Finally, the least square boundary error is used to reduce the times of particle resampling, and the time complexity of the whole algorithm is reduced, and the real-time fusion tracking is realized. An infrared and visible target fusion and tracking algorithm based on occlusion detection is proposed. The appearance model of the target is described by sparse representation of the target image, and a simultaneous tracking and recognition method is introduced based on the sparse representation. In order to solve the problem that the occluded tracking results are improperly added to the reference template set during the target template updating process, a occlusion detection model is established to calculate the size of the occluded region. According to the size of occlusion area, the reference model is updated by using cooperative learning method, so as to reduce the influence of occlusion factors on the tracking results. The test results of multiple infrared and visible image sequences show that the two tracking methods presented in this paper have a good performance in dealing with the intersection of targets, the rotation of targets, the variation of illumination and the occlusion of targets.
【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】