Human activity classification based on micro-Doppler signatures using an artificial neural network

An ANN has been proposed to classify human activities from their micro-Doppler signatures. Data were collected using a Doppler radar for 12 human subjects performing seven activities to construct the training data set. Six features from Doppler signatures were captured in the spectrogram. Validation tests based on the features resulted in an 82.7% and 87.8% classification accuracy for two different validation scenarios. This result shows that it is quite feasible to recognize the different human activities using micro-Doppler information. Several issues still need to be further addressed. In this study, we used measurement data for the training process. The features can be affected by the characteristics of the particular radar used, such as I-Q imbalance, polarization and Rx-Tx locations. Therefore, the trained ANN could lead to high error when it is used to classify data measured from another sensor. Our study is only applicable when the human approaches the radar head-on. Data from other aspects should be included in the testing. Also, we used a 3-second time-window for the features extraction. If the human activity changes during the window duration, classification error may increase. A method to extract features within a shorter time duration needs further research.