Exploiting Time Series for Human Activity Recognition with Radar

This is the undergraduate final year project. In this project, the radar-based human activity recognition (HAR) problem is analysed in details. Specifically, recurrent neural networks (RNNs) that exploit time series for classification are employed. In the first part of experiments, raw radar data are processed in the frequency domain, generating range profiles and micro-Doppler spectrograms, which reveal the change of range and speed with time. In the second part, LSTM and Bi-LSTM architectures are used to build RNN models that are able to preserve long time dependencies. Results show that applying Bi-LSTM architecture to micro-Doppler spectrograms is the most effective method, achieving an F1 score of over 91%. The Bi-LSTM model outperforms the LSTM model by 6% on micro-Doppler spectrograms, proving that the use of future information contributes to the improvement of performance.

Report and Code of this project can be viewed by clicking the corresponding buttons above.