Abstract: With the rapid development of artificial intelligence, human abnormal behaviour recognition based on computer vision has received great attention and is widely used in fields such as intelligent security. For violations such as people smoking in important places such as petrol stations and drivers talking on the phone while driving, this paper proposes a hybrid attentional mechanism for abnormal behaviour recognition. The introduced convolutional block attention module is used to focus on the salient features of the input objects, and the input information is finely allocated and processed to highlight important information while weakening irrelevant information. To enhance the feature mining capability of the network model and the interactivity of the network, the proposed convolutional feature extraction module is used to further extract the high-level semantic features of the recognised objects and fuse them with the low-level detailed features to achieve multi-scale feature interaction. In addition, in order to reduce the loss caused by mislabelling during network training, this paper uses label smoothing to correct the cross-entropy loss function to drive the learning process of the model. Experimental results show that the proposed model outperforms current mainstream networks and can effectively detect anomalous behaviour.

Key words: abnormal behavior detection, attention mechanism, convolution block attention module, convolution feature extraction module, label smoothing

摘要： 随着人工智能的快速发展，基于计算机视觉的人体异常行为识别受到极大的关注，并被广泛应用到智能安防等领域。针对人们在加油站等重要场所抽烟以及司机驾驶途中打电话等违规行为，提出一种混合注意力机制的异常行为识别方法。利用引入的卷积块注意力模块重点关注输入对象的显著性特征，并对输入信息进行精细化的分配和处理，在突出重要信息的同时弱化无关信息。为提升网络模型的特征挖掘能力及增强网络的信息交互性，利用提出的卷积特征提取模块进一步提取识别对象的高层语义特征，并将其与低层细节特征进行融合以达到多尺度特征交互的目的。此外，为了减少网络训练过程中错误标签造成的损失，采用标签平滑对交叉熵损失函数进行修正以此来驱动模型的学习过程。实验结果表明，所提出的模型优于当前的主流网络，可有效检测出异常行为。

关键词: 异常行为检测, 注意力机制, 卷积块注意力模块, 卷积特征提取模块, 标签平滑