Refine
Document Type
- Article (1)
- Conference Proceeding (1)
Language
- English (2)
Has Fulltext
- no (2)
Keywords
- Artificial Intelligence (1)
- Embedded hardware (1)
- Gesture Recognition (1)
- Gesture recognition (1)
- Low-Cost-Radar (1)
- Machine Learning (1)
- Machine learning (1)
- Neural Networks (1)
- Neural networks (1)
- Radar (2)
Institute
Classical signal processing methodologies have been infiltrated by machine learning (ML) approaches for a long time, where the ML approaches are in particular applied when it comes to gesture recognition. In this paper, we investigate naïve gesture recognition methodologies and compare classical and novel machine learning (nML) algorithms. The considered gestures are simple human gestures such as swiping a hand or kicking with a foot. For the sake of comparability, the algorithms are assessed with respect to their true positive rate (TPR), false-positive rate (FPR), their real-time capability together with the required computational power, and their implementability on low-cost hardware. Two different data sets are utilized separately for the training process of the ML algorithms, where both have been recorded by making use of low-cost radar hardware. The results show that all ML approaches are superior to naïve gesture recognition methodologies, e.g., threshold detection. ML algorithms allow almost assured gesture detection. However, our primary contribution is a design approach for scalable neural networks (NNs) that allow such gesture recognition algorithms to be executable on low-cost microcontroller units (MCUs).
Radar based gesture recognition offers great opportunities to increase user-friendliness of countless applications at home, in transportation and for industries. Here, not only data-intensive image and video processing, but also 1D multior single-channel time-series signals are in focus. We examine classical machine learning (ML) approaches and compare them in a reproducible manner. We evaluate the performance of naive methods—such as threshold detection (THD)—and classical ML methods—such as the support vector machine (SVM). The performance is hereby judged by elements such as accuracy, falsepositive rate (FPR), training and prediction time, hardware (HW) requirements and real-time capabilities as well as the size of the classifier. To create the library needed for the given investigation, a two channel continuous wave (CW) modulated radar system with carrier frequency of 10 GHz has been employed. We conclude that naive methods are outperformed by all investigated classical ML methodologies. The results in terms of accuracy and FPR are satisfactory. However, there are large differences between naive and ML methods in terms of HW requirements and real time performance. In conclusion, classical ML methods fulfil the defined requirements satisfactorily, only the real-time performance on low-performance HW is limited due to the required computing power. Thus, the algorithms are a good choice for gesture recognition—of 1D multi- or single-channel time-series signals—if applied correctly.
