Human Activity Recognition Based on 4-Domain Radar Deep Transfer Learning

We demonstrate the improvement of theradar-based human activity recognition using the combination of four datadomains: time-frequency, time-range, range-Doppler and, for the first time,time-angle domain. Six different activities are observed from nine subjectsusing frequency-modulated continuous-wave millimeter-wave radar. Each domainoffers additional information to the classification process. The classificationresults of four deep convolutional neural networks are then combined using theJoint Probability Mass Function method to achieve a combined classificationaccuracy of 100%. The proposed system also demonstrates similar performance inrecognizing activities from participants not involved in training the network.To the best of our knowledge, this is the first work that demonstrates theutilization of four data domains to address the radar-based human activityrecognition problem.