Automated behavior recognition using continuous-wave Doppler radar and neural networks

Continuous-wave Doppler radar (CWDR) signals were used to automatically differentiate the overt waking behaviors of exploratory locomotion (EL), grooming (GR) and behavioral stillness (BS) in the rat. CWDR signals were acquired during EL, GR and BS in 2 second epochs. RMS values and spectral estimates of the CWDR epochs were computed and used to train a multilayer feedforward neural network to discriminate the behaviors. 126 epochs of empirical data were collected and 96 of these were used for training the various neural networks. The remaining 30 epochs (10 of each behavioral class) were used to test the predictive ability of the trained networks. Four 3-layer network architectures were tested differing only by the number of “hidden” units (2, 3, 4 or 5). Otherwise, all networks had 5 input units (coded for 4 spectral bands plus RMS amplitude) and 3 output units (coded for the 3 overt behaviors). These networks were able to learn from 87% to 96% of the training set. However, only one model (3 hidden units) was able to predict overt rat behavior with better than 85% accuracy from the test set. These results demonstrate the feasibility of using CWDR signals to distinguish various forms of overt behavior in the rat. It also validates the usefulness of a combined power-spectral/neural-net approach to recognize complex electrical signals