基于多任务分类的吸烟行为检测

为了及时检测吸烟行为,准确做出状态判断,提出了一种基于多任务分类的吸烟行为检测算法。该算法融合多任务卷积神经网络、级联回归和残差网络,通过多任务卷积神经网络算法和基于梯度提高学习的回归树方法(RET级联回归)快速定位嘴部感兴趣区域(ROI);在此基础上,采用残差网络对ROI内目标进行检测和状态识别。实验结果表明,该算法可以准确检测到吸烟行为的发生并做出状态判断,准确率可以达到87. 5%。

Smoking Detection Algorithm Based on Multitask Classification

In order to detect the smoking behavior in time and accurately judge the state. A smoking behavior detection algorithm based multi-task classification was proposed. The algorithm integrates multi-task convolution neural network, ensemble of regression trees cascade regression and depth residual network, quickly and accurately locates the region of interest through multi-task convolutional neural network algorithm and ERT cascade regression. Based on this, detect targets in the region of interest and identify status using deep residual network. The experimental results showed that the algorithm can accurately detect the occurrence of smoking behavior and make state judgments, the accuracy rate can reach 87.5%.