Fall Detection Based on Parallel 2DCNN-CBAM With Radar Multidomain Representations

With the rapid development of population aging, fall detection based on the radar manifests great application value in the field of medicine and health. Focusing on the problem of the false alarm and missing alarm during the detection, this article proposes a fall detection method based on the parallel 2-D convolutional neural network and convolutional block attention module (2DCNN-CBAM) with radar multidomain representations. Specifically, the means makes use of pulse compression, Doppler-fast Fourier transform (FFT), capon beamforming, and target status update to form four types of representations in different domains from the radar echo signal for the purpose of constituting the multidomain representations, and then, the images are sent into the parallel 2DCNN-CBAM for recognition. According to the network structure in this article, the parallel 2DCNN is utilized to extract the characteristics from the corresponding type of representation independently, and then, the extracted features are screened and optimized by the CBAM that is connected to the corresponding 2DCNN in series. Finally, the features in representations are fused to realize the multidomain characterization of the target behavior, which is able to enhance the classification rate. Experimental results illustrate that the proposed method can effectively decrease the probability of false alarm and missing alarm. Besides, the proposed methodology improves the average accuracy of five typical human behaviors by at least 4.2% and promotes the classification rate of fall by at least 2.2% compared with other contrastive methods.