基于知识迁移的跨领域人体动作识别

Recognizing human activities in video has become one of the most amazing applications of pattern recognition and computer vision. One fundamental assumption in traditional learning is that training and testing data are sampled from an identical distribution. However, this assumption is not always valid. For example, when the data for one learning task (called the target domain) are limited and we want to use the data from the auxiliary domain to improve the performance of the leaning task at hand. Therefore, this project study the problem of human action recognition from a new perspective which aims to present a novel cross-domain recognition model for human activities by transfer learning. The main works included three aspects: (1) developing a novel cross-dataset human action recognition model which employing the probabilistic topic model to explore the cross-dataset action knowledge representation. (2) proposing a novel model for recognizing human actions by topic model from different views by view knowledge transfer which based on the high-level semantic knowledge. (3) combining cross-dataset and cross-view human action recognition strategies to boost a novel architecture for the cross-domain transfer learning by a supervised topic model.

人体动作识别是当前模式识别和计算机视觉领域研究的热点。传统的动作识别假设训练和测试样本来自相同的分布,然而这个假设在很多情况下并不成立。例如一个识别系统很难获得足够多的标记样本来训练分类模型。为有效利用多种动作视频数据，提高动作识别的性能，本项目借助迁移学习的思想,拟深入探索跨领域(Cross-Domain)的动作识别理论，主要研究：(1) 跨库(Cross-Dataset)动作识别方法，深入分析动作数据库的底层特征，将概率主题模型引入到跨库识别中，研究数据库中高层语义知识的表示和传递。（2）跨视角(Cross-View)的动作识别方法，深入探索不同视角下动作模式的本质联系，将概率主题模型引入跨视角的识别任务中，构建可在不同视角间迁移的高层知识表示。（3）研究基于主题模型的迁移学习理论，引入监督形式的概率主题模型，将跨库动作识别和跨视角动作识别整合到一个统一的学习框架。

人体动作识别是当前计算机视觉和模式识别领域研究的热点。本项目利用迁移学习的思想，提出了一个主题迁移模型(topic transfer model)用于跨视角的动作识别。借助源视角视频和目标视角视频，学习一个迁移模型，利用这个模型来实现对目标视角下视频的分类。具体方法是在源视角下训练一个主题模型，将反应源视角的语义信息传递到目标视角中，然后在目标视角中训练一个主题模型，实现跨视角的动作表示，利用分类器进行动作的训练和分类。.重要结果有：.1) 利用主题模型中的主题(topic)来表示视觉词的更高层语义。.2) 提出基于主题的转移模型，在源视角和目标视角下的学习可以采用相同的模型进行学习。.3) 建立跨视角识别中可迁移特征学习与分类器学习之间的联系。提出一种监督的主题迁移学习模型，在源视角,下，利用监督的主题模型(supervised topic model)，来学习源视角下的动作类型。