基于低分辨率红外传感器的深度学习动作识别方法

近年来动作识别成为计算机视觉领域的研究热点，不同于针对视频图像进行的研究，本文针对低分辨率红外传感器采集到的温度数据，提出了一种基于此类红外传感器的双流卷积神经网络动作识别方法。空间和时间数据分别以原始温度值的形式同时输入改进的双流卷积神经网络中，最终将空间流网络和时间流网络的概率矢量进行加权融合，得到最终的动作类别。实验结果表明，在手动采集的数据集上，平均识别准确率可达到98.2%，其中弯腰、摔倒和行走动作的识别准确率均达99%，可以有效地对其进行识别。

Deep Learning Method for Action Recognition Based on Low Resolution Infrared Sensors

In recent years, action recognition has become a popular research topic in the field of computer vision. In contrast to research on video or images, this study proposes a two-stream convolution neural network method based on temperature data collected by a low-resolution infrared sensor. The spatial and temporal data were input into the two-stream convolution neural network in the form of collected temperature data, and the class scores of the spatial and temporal stream networks were late weighted and merged to obtain the final action category. The results indicate that the average accuracy of recognition can reach 98.2% on the manually collected dataset and 99% for bending, falling, and walking actions, indicating that the proposed net can recognize actions effectively.