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

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