Radar-based human gesture recognition

Abstract

Radar technology is becoming popular in modern industrial settings due to its ability to create intelligent systems that enhance productivity and ensure safety. It's particularly promising in human gesture recognition, which allows radar sensors to detect human worker movements and interpret them with deep learning technics for more efficient and accurate robotic arm performance. In this project, we developed a radar-based human gesture recognition system for robotic arms in a factory. We first collected raw radar data from three volunteers using TI’s AW1642 FMCW radar. Then we designed a data processing pipeline, which converted the radar raw data into Micro-Doppler feature maps. A hand gesture classification dataset was built with more than 4000 data. Finally, a convolutional neural network (CNN) model was proposed to perform classification on the dataset which reached an average accuracy of 92.79% on 7 hand gesture classes. To improve the performance on unseen people, we modified the current model and proposed an end-to-end complex-valued CNN model, which accepts our complex radar data as inputs. The proposed complex network increased the classification accuracy up to 2% to 4% when testing on different scenarios.