Human Detection with Range-Doppler Signatures Using 3D Convolutional Neural Networks

Human detection is proposed based on time-varying range- Doppler signatures measured by millimeter-wave FMCW radar using deep recurrent neural networks. Human detection is a significant topic for security, surveillance, and search and rescue. When a target is measured by fast-chirp FMCW radar, a range-Doppler diagram can be constructed in real time. Because the signatures in a range-Doppler diagram are time-varying, we investigated the feasibility of classifying targets using those signatures. We measured five classes-humans, cars, cyclists, dogs, and road clutter-using millimeter-wave FMCW radar. We applied 3D-convolutional neural networks to 3D representations of time-varying signatures and achieved a classification accuracy of 97%, with a human detection rate of 100%.