基于背景感知与稀疏表达模型的目标跟踪

Object tracking has broad application prospects on following areas, such as intelligent video surveillance, intelligent traffic control, robot control, unmanned aerial vehicles driving and navigation. Most of existing tracking methods cannot satisfy real world applications in terms of object representation, background suppression and occlusion. To address these problems, this project introduces background-aware and sparse representation models. First, according to online classification and multi-kernel learning, it develops a tracking algorithm based on ideal-kernel guided multi-kernel learning. To promote the robustness on occlusion, it introduces structure information into sparse representation, and based on this, it proposes a tracking algorithm via image-guide structure sparse representation. Finally, combining with background aware and sparse representation models, it constructs a tracking algorithm based on local discriminative kernel sparse representation. In addition, this project will develop an experimental object tracking system based on the proposed tracking algorithms. Meanwhile, the proposed algorithm will be tested on this system to improve their accuracy and robustness.

目标跟踪在智能视频监控、智能交通控制、机器人控制、无人机驾驶与导航、体感人机交互等领域具有广阔的应用前景。现有的跟踪方法在目标表示、背景抑制和遮挡鲁棒等方面还不能满足实际应用。针对上述问题，本项目引入背景感知以及稀疏表达模型。本项目首先以在线分类和多核学习为基础，重点研究基于理想核指导多核学习的目标跟踪算法。为了提高跟踪算法对遮挡的鲁棒性，本项目将结构信息引入稀疏表达模型，提出基于图像指导结构稀疏表达的目标跟踪算法。最后，将背景感知和稀疏表达结合起来，构建基于局部鉴别核稀疏表达模型的目标跟踪算法。另外，本项目以研制的目标跟踪算法为基础，设计目标跟踪原型系统。同时，利用该系统验证本项目所提的跟踪算法，进一步完善和提高算法的准确性和鲁棒性。

本项目研究内容主要是以背景感知、稀疏表达以及鉴别学习等理论研究为基础，进而应用于视觉目标跟踪问题。首先，提出了基于理想核指导多核学习的目标跟踪算法，从而自适应地融合多特征（包括背景感知）。其次，提出了结构稀疏表达的目标跟踪算法，从而处理目标外观变化、相似背景干扰和遮挡等主要的跟踪难题。在次，提出了多个最大间隔的多类分类器模型以及深度聚类模型，为基于鉴别学习的目标跟踪提供算法和理论基础。最后，以视觉目标跟踪相关理论和算法为基础，实现“基于交通监控系统的车辆检测及分类”演示系统。

内容获取失败，请点击重试