基于多特征融合与多级多模式分类的人体动作识别技术研究

The major challenge of human action recognition is to deal with all kinds of variabilities, such as different viewpoints, different acting speeds and styles, different genders and sizes. To address these concerns, the research is carried out from three aspects as action description, action modeling and action recognition. (1) For action description, action description algorithm based on multi-features fusion is studied. Complementary static and motion features are extracted and fused by exploiting local descriptors and holistic features as well as temporal and spatial features. Thus the influence of variable factors is reduced and the robustness of feature representation is improved. (2) For action modeling, non-linear SVM (support vector machine) action model learning algorithm based on MKL (multiple kernel learning) is firstly studied. Then additional action model learning algorithm based on transfer learning is studied to improve the extensibility of action models. (3) For action recognition, coarse action classification algorithm based on template matching of key poses is firstly studied to address the key poses mined from the action sequence. Then fine action classification algorithm based on non-linear MKL-based SVM is studied to address the spatiotemporal specialty in the sub video of key frames. In addition, action classification algorithm based on fuzzy logic is studied to address the uncertainty problem during action recognition. Therefore, the overall performance of action recognition can be improved by the above approaches of multi-features fusion, multi-levels and multi-patterns classification. Smooth execution of the project is helpful to provide more effective foundation for activity analysis, activity understanding and other intelligentized applications.

视角变化、动作执行速度和人体结构差异等可变因素是造成人体动作识别困难的关键所在，本项目从动作描述、动作建模和动作识别三方面展开研究：（1）在动作描述方面，研究基于多特征融合的动作描述算法，从整体和局部、时间和空间等不同层次提取具有互补性的静态与动态动作特征，减少可变因素的影响，提高特征描述的鲁棒性；（2）在动作建模方面，研究基于MKL的非线性SVM动作模型学习算法，在此基础上研究基于迁移学习的动作模型学习算法，提高动作模型的可扩展性；（3）在动作识别方面，针对关键帧图像中的关键姿态，研究基于关键姿态模板匹配的动作粗分类算法；针对关键帧子视频中的时空特性，研究基于MKL的非线性SVM动作细分类算法；同时，针对动作识别中存在的不确定性，研究基于模糊逻辑的动作分类算法，从而通过多级多模式的分类方法提高动作识别算法的整体性能。项目的顺利开展将会为行为分析和理解等其他智能化的应用提供更有效的依据。

动作识别技术在很多领域都有着广泛而重要的应用，能为行为分析和理解等其他智能化的应用提供更加有效的依据，项目从基于多层次特征提取与融合的动作描述、基于机器学习的动作建模以及基于多级多模式分类的动作识别三方面展开研究。将相关问题抽象为N最短路径优化问题，提出了一种基于GA的智能优化算法；对传统的问题描述进行了扩展，在分析传统多重标号算法的基础上，提出了一种改进的理论严密的多重标号算法；结合蚁群算法的信息正反馈思想，提出了一种基于人工免疫的N最短路径检索算法，提高了求解大规模问题的性能；提出了一种基于Bayesian网络的目标类型识别算法，利用曲线拐点检测方法对运动模式段进行划分，提取得到多运动特征参数后据此建立目标类型识别的贝叶斯网络模型；为降低计算复杂度、提高算法的执行效率，提出了一种基于V-最优直方图的KNN分类算法；针对传统使用Newton方法进行损失函数优化的多类逻辑回归分类器计算开销较大，且迭代过程不稳定，提出了一种基于BFGS的多类逻辑回归分类算法；并将BFGS用于对多类线性SVM进行优化，以更少的运行时间，取得了与非线性SVM相当的分类性能。