【摘要】 群体行为识别任务中，行为特征具有复杂的时空特性。为了实现有效的行为特征时间编码，本文提出一种融合时间和空间上下文特征的群体行为识别模型。为了分析个体行为特征的时间上下文依赖关系，设计了通道级时间上下文模块，该模块对个体特征的多个通道进行时间平移；分别研究时间延迟移动、时间双向移动、时间循环双向移动的3种策略，并讨论各种策略下通道比例对时间上下文估计的作用。其次，构建了基于融合通道级时间上下文特征的空间图模型，用于对个体空间上下文的编码。该模型使用外观和位置估计初步的个体之间的空间上下文关系，并进一步设计多图策略，来估计多种可能的个体之间的关系。最后，对图模型编码的个体特征，使用个体池化获得群体特征，并使用多层感知器来识别群体行为。本文方法在Volleyball和Collective Activity数据集上优于现有群体行为识别方法，设计的时间上下文特征具有良好个体行为编码能力。更多还原

【Abstract】 In a group activity, individuals have complex spatial-temporal features. To encode the complex spatial-temporal features, the paper proposes a group activity recognition model based on temporal and spatial context features. First, to analyze the temporal context-dependency in individual features, a channel-wise temporal context module is designed, which uses a shift strategy to learn temporal context. Three strategies are studied, including temporal delay shift, temporal bi-direction shift, and temporal recurrent bi-direction shift, and the shift ratio in the shift strategy is also discussed. Second, a spatial graph model based on fusing channel-level temporal context features is constructed to encode the spatial context of the individual. The initial spatial context relation is estimated with both appearance feature and position feature. Furtherly multiple graph strategy is used to represent multiple relations. Finally, temporal pooling is used to aggregate the individual features into group features and multiple layer perceptron is used to predict the group activity. Experimental results in the Volleyball dataset and the Collective Activity dataset show that the proposed method outperforms the state-of-the-art methods. The proposed temporal context features encode the individual features well.更多还原