Real-Time Gesture Recognition with Shallow Convolutional Neural Networks Employing an Ultra Low Cost Radar System

Ultra-low-cost radar hardware (HW) in combination with low-cost processing units is investigated in order to create and evaluate a holistic ultra-low-cost gesture recognition system. We study the real-time performance of novel machine learning (nML) methods: neural networks (NNs), in particular shallow architectures of convolutional neural networks (CNNs). The real-time performance of each approach is judged by computational complexity, prediction time, accuracy, and false-positive rate (FPR). As HW, a two-channel radar system with continuous wave (CW) modulation at a carrier frequency of 10 GHz has been employed throughout the investigations. The algorithms are designed, trained, evaluated, and juxtaposed. The results show that the classification process on low-cost HW is feasible and allows to achieve accuracies of 97.9% and FPRs of 1.72%, all of which with a response time of less than 180 ms.