基于EfficientNet模型的毫米波雷达人体行为识别

FMCW Radar Human Action Recognition System Based on EfficientNet Model

针对基于传统光学摄像头实现人体行为识别系统所带来的隐私暴露,易受光照、遮挡的影响等问题,设计基于EfficientNet模型的FMCW雷达人体行为识别系统。首先对FMCW雷达实测数据采用距离多普勒算法构建每一帧距离-速度图像;接着采用逐帧积累的方法延长观测时间窗口,构建覆盖整个动作过程的距离-速度轨迹;最后采用改进预训练的EfficientNet模型对不同人体行为进行识别。实验结果表明,在5秒观测时间窗口内,改进预训练的EfficientNet-B4模型对已知个体和未知个体9种行为识别准确率达到99.3%与98.2%,均高于传统机器学习方法及经典深度学习方法,进一步缩短观测时间窗口至2.5秒,改进预训练的EfficientNet-B4模型对已知个体和未知个体的9种行为识别准确率仍能达到96.7%与95.4%。除此之外,在5秒观测时间窗口内,所提方法对已知个体和未知个体的9种行为识别准确率比常见利用时间-速度提取行为参数的方法分别提高了3.5%与4.9%,缩短观测时间窗口至2.5秒,所提方法准确率提高了4.2%与4.8%,可见所提方法可以有效地提升FMCW雷达人体行为识别的准确率,且模型的泛化能力较强。

Since using traditional optical camera to realize human action recognition system will bring some problems,such as privacy exposure,easy to be affected by light and occlusion,a FMCW radar human action recognition system based on EfficientNet model is designed.Firstly,range Doppler algorithm is used to construct range-velocity images of each frame from the measured data of FMCW radar.Then the range velocity trajectory covering the whole movement process is constructed by using the method of frame-by-frame accumulation to extend the observation time window.Finally,the improved pre-training EfficientNet model is used to recognize different human actions.The experimental results show that within the 5-second observation time window,the accuracy of the improved pre-training EfficientNet-B4 model for identifying nine actions of known and unknown individuals is 99.3%and 98.2%,which are both higher than that of traditional machine learning methods and classical deep learning methods.When the observation time window was further shortened to 2.5 seconds,the recognition accuracy of the improved pre trained efficientnet-b4 model for 9 behaviors of known and unknown individuals could still reach 96.7%and 95.4%.In addition,within the 5-second observation time window,the accuracy of the proposed method for identifying 9 actions of known and unknown individuals is 3.5%and 4.9%higher than the common methods of extracting behavior parameters by time-velocity.When the observation time window is shortened to 2.5 seconds,the accuracy of the proposed method is improved by 4.2%and 4.8%.It can be seen that the proposed method can effectively improve the accuracy of FMCW radar human action recognition,and the model has strong generalization ability.