LSTM Based Classification of Targets using FMCW Radar Signals

According to the World Health Organization(WHO), every year around 20-50 million people are injured from road traffic accidents. Most of the injuries are among vulnerable pedestrians, cyclists, and motorcyclists. Autonomous vehicles (AVs) seem to be the perfect solution to this problem. Radar sensors in AVs is an effective sensor since it simultaneously measures speed and range while being robust in bad weather conditions. In this work first, a dataset which contains 300 spectrograms are created by simulating a 24GHz FMCW radar signals. In a 2D simulation environment, a single radar is placed to the origin, and other objects of varying parameters ( e.g height, heading, speed) are placed in this rectangular area. Then, the features are extracted from the Micro-Doppler patterns on the spectrogram images and trained on Long Short Term Memory Networks(LSTMs). The average accuracy and F1 score of the proposed method on the test set is 95% which outperforms some existing methods.