Learning Dynamic Processes from a Range-Doppler Map Time Series with LSTM Networks

This paper proposes a Deep Learning approach for microwave frequency based dynamic classification tasks using a time series of Range-Doppler maps. Besides the range and velocity space, the temporal space is learned by using recurrent connections in a neural network (LSTM: Long Short-Term Memory). The Range-Doppler maps are recorded with a 77 GHz chirp-sequence radar sensor within a cooking sensor application to prevent overboil. The proposed combination of convolutional and recurrent network is verified with a dynamic cooking process dataset which shows real-life conditions in a kitchen. The achieved accuracy is 98.02% over six classes. The network's ability to learn temporal features is investigated by resetting the LSTM's internal state vector at different time steps.