An Integrated Spectrogram Classifier-based Hand-Gesture Recognition SoC for an FMCW Radar Sensor
Yongchul Jung, Jong-Ho Kim, Mingeon Shin, Haram Ju, Kang-Il Cho, Young Han Lee, Yunho Jung, Sungho Lee · 2022 IEEE Radar Conference (RadarConf22) · 2022
This work presents a system-on-chip (SoC) for FMCW radar signal processing to implement hand gesture recognition. For smart IoT device applications, a radar signal processing unit with a lightweight convolutional neural network (CNN) is integrated so that a spectrogram classifier can be implemented at the chip level. The spectrogram classifier is integrated with preprocessing blocks of FFTs and an MTI filter in 40nm CMOS process. The network model size of the CNN was set to about 60.1 KiB to make the system suitable for small IoT applications. To handle variations resulting from different hand sizes and gestures, 10,800 training data from 72 participants were utilized with the customized preprocessing. The performance evaluation results showed that the gesture recognition accuracy for the five gestures was measured to be 91.5 % in a user-independent manner.