【摘要】 人体行为检测旨在精准定位视频中人员的空间位置，并有效识别其执行的各类动作。针对因动作变化细微以及不同动作表征相似所导致的行为检测不准确的问题，提出了卷积时间注意力与多尺度信息学习的时空动作检测网络（CTAMIL-Net）。设计一种多尺度信息学习2D网络（MSILN），采用异尺度卷积层从关键帧中抽取不同粒度的特征信息，并通过注意力特征融合模块（AFF）自适应地聚合高级语义与低级细节特征，以捕获目标外观特征，感知动作的细微变化；提出一种3D卷积时间注意力（3D-CTA），分别从帧间、帧内捕获短-长时时间依赖关系，并利用3D主干网络（3D-ResNext-101）挖掘更具判别性的目标运动特征，共同提升对相似动作的区分度；采用特征融合检测头将2D外观特征和3D运动特征语义交互并得出动作检测结果。实验结果表明，该算法与现有先进方法在公共动作检测数据集UCF24、AVA和自建热不适动作数据集TDA上均取得了较好的性能结果，能够准确检测出热环境中的热不适动作，如扇风、穿衣脱衣等，验证了所提方法在以热不适动作检测为典型应用场景中的有效性，为复杂环境下的人体行为检测提供了新的解决方案。更多还原

【Abstract】 Human behavior detection aims to precisely locate individuals in videos and effectively recognize the various actions that they perform. Aiming at the problem of inaccurate behavior detection caused by subtle action changes and the similarity of different action representations, this paper proposes a convolutional temporal attention and multiscale information learning for thermal discomfort spatio-temporal action detection networks(CTAMIL-Net). Firstly, a multiscale information learning 2D network(MSILN)is designed to extract feature information of different granularities from keyframes using a heteroscale convolutional layer, and adaptively aggregates high-level semantic and low-level detail features via an attention feature fusion module(AFF)to capture target appearance features and perceive subtle changes in movements. Secondly, 3D convolutional temporal attention(3D-CTA)is proposed that captures short-long time temporal dependencies from inter-frame and intra-frame, respectively, and utilizes a 3D backbone network(3D-ResNext-101)to mine more discriminative target motion features, which together enhance the differentiation of similar actions. Finally, a feature fusion detector head is employed to semantically interact the 2D appearance features with the 3D motion features and derive the action detection results. The experimental results demonstrate that the algorithm presented in this paper,along with existing advanced methods, achieves superior performance on the public action detection datasets UCF24,AVA, and the self-constructed thermal discomfort action(TDA)dataset. It accurately detects thermal discomfort actions in hot environments, such as fanning oneself, putting on or taking off clothes, etc. This validates the effectiveness of the proposed method in typical application scenarios involving thermal discomfort action detection. Furthermore, it provides a novel solution for human behavior detection in complex environments.更多还原