Human Activity And Identity Multi-task Recognition Based On Micro-doppler Radar

In recent years,with the increase of Doppler radar accuracy and cost reduction,human recognition based on radar has gradually become a hot topic in theoretical research and applied research.Human activity and identity recognition based on microDoppler radar have very broad application prospects in military,disaster rescue,medical monitoring and other fields.By analyzing the existing human recognition algorithms based on micro-Doppler radar,this paper finds that there are limitations to the recognition methods for modeling different recognition tasks.Such methods ignore supervisory information between related tasks and can easily lead to overfitting on a single recognition task.Based on the potential correlation between radar human activity recognition and identity recognition,this paper introduces the idea of multi-task learning into deep learning model to promote feature sharing between related tasks.Therefore,the recognition accuracy and generalization performance of the recognition model have been improved.The main research work includes:(1)A multi-task recognition model based on convolutional neural network for human activity and identity is proposed.Aiming at the problem of insufficient supervision information in single task recognition,construct multi-task feature sharing layers and task specific layers,introduce space-based attention mechanism on the structure,re-correct the generated feature maps at various levels.Multi-scale structure is used for feature fusion to obtain more robust multi-task sharing features.Finally,multi-task recognition is implemented on the task specific layers.In order to improve the generalization performance of the model,the joint loss function consisting of central loss,mean square error loss and cross entropy loss is designed to optimize the multitask recognition model,which is beneficial to the sample to form a more compact class distribution in the shared feature space.(2)Aiming at the time series of radar time-frequency spectrums,a human activity and identity multi-task recognition model combining convolutional neural network and long short term memory network is proposed.First,automatic learning of image features for each time period is implemented based on a convolutional neural network.Then,the obtained features are organized in chronological order as input of each time step of the long short term memory network,and the shared features of multi-task recognition are obtained through time series analysis.On this basis,the time-based attention mechanism is introduced.By learning the attention weight parameters,the importance degree of the shared feature information outputted by each time step is obtained for each recognition task,and the more accurate representation of the specific task is realized.The contrast experiments on the radar measured data set show that the method can complete the human activity recognition task and identity recognition task more effectively.The generalization analysis is carried out on the datasets with different training data volume,noise levels and short-time Fourier transform window lengths.Experiments show that the proposed method has good generalization performance on the radar measured data sets under various conditions.