摘要： 在实际场景中，因香烟目标过于微小且特征不明显，现有的目标检测算法难以区分类烟物与香烟，导致吸烟行为识别效果差。提出一种基于弱监督细粒度结构与EfficientDet网络的吸烟行为识别算法。采用Edge Boxes算法检测图像块的特征边缘，通过非极大值抑制对边缘进行筛选，形成候选区域块。构建包含物体级筛选器和局部级筛选器的细粒度两级注意力模型，其中物体级筛选器使用改进的EfficientDet网络滤除候选区域的背景噪声，以分类前景物体及特征较强的候选区域，并在局部级筛选器中使用通道注意力卷积块对候选区域进行聚类，筛选出得分最高的像素块。通过融合物体级筛选器与局部级筛选器得到的结果，以准确识别吸烟行为。在BUU-Smoke数据集上的实验结果表明，该算法的吸烟行为识别准确率为93.10%，误检率为3.6%，并且具有较优的鲁棒性和泛化能力。

关键词: 吸烟行为, EfficientDet网络, 弱监督细粒度, 注意力机制, 行为识别

Abstract: In the actual scene, the cigarette target is too small and the characteristics are not obvious.The existing target detection algorithms are difficult to distinguish similar goals to cigarettes from cigarettes, resulting in poor recognition effect of smoking behavior.This paper proposes a smoking behavior recognition algorithm based on weakly supervised fine-grained structure and EfficientDet network.Edge boxes algorithm is used to detect the characteristic edges of image blocks, and the edges are filtered through non maximum suppression to form candidate region blocks.Fine-grained two-level attention model is builded, including object level filter and local level filter.The object level filter uses the improved EfficientDet network to filter out the background noise of the candidate region, so as to classify the foreground objects and candidate regions with strong features.In the local level filter, the channel attention convolution block is used to cluster the candidate regions and filter out the pixel block with the highest score.The results obtained by fusing object level filter and local level filter are used to accurately identify smoking behavior.The experimental results on BUU-Smoke dataset show that the recognition accuracy of smoking behavior is 93.10% and the false detection rate is 3.6%.It has better robustness and generalization ability.

Key words: smoking behavior, EfficientDet network, weakly supervised fine-grained, attention mechanism, behavior recognition

中图分类号: 

• TP391