基于压缩感知的高精度实时视觉跟踪方法研究

项目名称： 基于压缩感知的高精度实时视觉跟踪方法研究

项目编号： No.61501509

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 无线电电子学、电信技术

项目作者： 胡磊

作者单位： 中国人民解放军陆军工程大学

项目金额： 18万元

中文摘要： 作为计算机视觉领域的经典问题之一，视觉跟踪具有重要的理论与应用价值。由于视频数据规模巨大，且目标外观易于发生变化，视觉跟踪也是计算机视觉领域最富挑战性的问题之一。本项目研究基于压缩感知的视觉跟踪方法，通过压缩感知的欠采样能力减小由大规模视频数据带来的沉重计算与存储负担，同时利用稀疏恢复对目标外观进行建模，提高跟踪器对于外观变化的鲁棒性。具体地，本项目研究适应于视觉跟踪需求的压缩感知测量矩阵的设计与实现方法，由此提高测量矩阵兼顾计算复杂度与特征稳定性的能力。其次，本项目发展目标外观的联合稀疏表示与词典优化模型。基于该模型，通过在稀疏恢复过程中动态地更新稀疏表示词典来处理目标的外观变化，从而消除或减轻由外观变化引起的跟踪精度恶化。同时，本项目在贝叶斯推理框架下求解基于上述模型的联合稀疏恢复与词典优化问题，从而避免因算法参数人工设置不准确而引起的算法精度下降，同时提高跟踪过程的自动化水平。

中文关键词： 视觉跟踪；压缩感知；稀疏恢复；词典优化；贝叶斯推理

英文摘要： As one of the classical problems in computer vision, visual tracking is of great significance in both theory and applications. Since video data is usually huge, and meanwhile the target appearance is prone to change during the tracking period, visual tracking is also one of the most challenging problems in computer vision. This program focuses on the visual tracking method based on compressed sensing (CS). By exploiting the undersampling ability of CS, it aims to reduce the heavy burden in computation and memory incurred by extremely huge video data; meanwhile, by modeling the target appearance using sparse recovery, it tries to improve the robustness of the tracker to appearance variations. Specifically, the program first investigates the generation and implementation method of the CS measurement matrices that are tailored to the requirements of visual tracking, thereby endowing the measurement matrix with the ability to taking account of both computational complexity and feature stability. Secondly, the program develops the model that can achieve joint sparse representation and dictionary optimization for target appearance. Based on this model, it deals with target appearance variations via dynamically updating the sparsifying dictionary during sparse recovery. As a consequence, the degradation of tracking accuracy caused by appearance variations can be eliminated or mitigated. Furthermore, under the framework of Bayesian inference, the program solves the joint sparse recovery and dictionary optimization problem corresponding to the above model. By doing so, it can avoid the degradation of tracking accuracy incurred by inaccurate setting of algorithmic parameters, and at the same time strengthen the automation characteristic of the tracking procedure.

英文关键词： visual tracking;compressed sensing;sparse recovery;dictionary refinement;Bayesian inference