Research On Human Behavior Recognition Based On Deep Learning And Radar Micro-Doppler Characteristic

Radar-based human activity recognition technology has many advantages,such as allweather,non-contact,high accuracy,and strong privacy,which can be applied in fields such as intelligent elderly care,human-machine interaction,and intelligent security.Feeding radarpreprocessed micro-Doppler signals into deep learning(DL)networks to achieve human activity classification and recognition is one of the mainstream research directions in the field of radar human activity recognition.This research direction is not only novel and feasible but also has important practical significance and application value.However,existing DL-based recognition models are difficult to achieve a balance between accuracy and lightweight,i.e.,the recognition accuracy of simple networks is limited,while the number of parameters and computation of complex networks are large,which brings difficulties to real-time embedded applications with limited computational resources.This study focuses on the efficient and accurate lightweight human activity recognition problem,based on Frequency Modulated Continuous Wave(FMCW)millimeter-wave radar to carry out research on the human activity recognition network model.The main contents include:(1)Based on the feature representation of human activity in the time-frequency domain,this study proposes an efficient and lightweight human activity recognition model called CLA(1D CNN-LSTM-Attention).CLA decouples the micro-Doppler and temporal features of FMCW millimeter-wave radar preprocessed signals.One-Dimensional Convolutional Neural Network(1D CNN)is used to obtain the micro-Doppler feature representation of the sequence according to the sliding window.Then,the Attention-based Long Short-Term Memory(LSTMAttention)network is used to extract the key temporal features between the micro-Doppler feature sequence and finally complete the recognition of human activity.(2)In the radar micro-Doppler signal processing stage,this study effectively enhances human activity features by using the average cancellation clutter suppression algorithm.Experimental results show that compared with the traditional moving target indication algorithm,the recognition accuracy is improved by approximately 3.7%,and compared with signal processing without clutter suppression algorithm,the recognition accuracy is improved by 6.7%.(3)This study compared the performance indicators of different human activity recognition models on two human activity datasets(a self-established dataset and a public dataset),and the experimental results confirmed that CLA hybrid network balances recognition accuracy and network efficiency.Specifically,CLA achieved nearly 96.9% accuracy on both datasets and had a more lightweight network structure compared to complex networks with similar recognition accuracy.Therefore,the proposed solution in this paper can meet the requirements of efficient and accurate human activity recognition tasks,while having a lightweight network structure,and has great potential for real-time embedded applications.