Walking Pattern Identification of FMCW Radar Data based on a Combined CNN and bi-LSTM Approach

Automotive radars could have a pivotal role in human-machine interfaces thanks to their ability to monitor human activity and physiological state in a contactless way. Nowadays, data-driven models, such as deep learning approaches, are the needed cutting-edge technology to achieve appropriate results in the classification of activities. One of the issues in the literature is the identification of walking patterns from the radar acquisition and doing so in real-time would be beneficial to give instant feedback to the user. For this reason, we propose a model composed of a Convolutional Neural Network followed by a bidirectional Long-Short Term Memory for the classification of the range-Doppler-time data obtained from radar acquisition. The approach reaches at least a 90% f1 score for the abnormal walking patterns class observing just one gait cycle or one second of acquisition and it is a perfect anomaly detector with 2.6 seconds of acquisition time. Instead, to achieve good accuracy on the classification of all the classes we need a larger window of observation with an overall accuracy of 89.1% for 8 seconds and 95.6% for the complete acquisition, lasting 12 to 16 seconds.