| Gesture recognition in human-computer interaction has become a hot topic in the fields of computer science and artificial intelligence.Traditional gesture recognition methods include those based on sensors,vision,and wireless signals.Each of these methods has its own advantages and limitations,with vision-based methods performing better in terms of accuracy but requiring a large amount of training data and computing resources.Gesture recognition based on millimeter-wave radar is not affected by factors such as lighting and obstruction,making it suitable for a variety of environmental scenarios and achieving higher recognition accuracy for more precise gesture recognition.This article focuses on gesture recognition algorithm research using the IWR1642millimeter-wave radar development board,with specific details as follows: Firstly,the radar gesture signal is analyzed and processed.The appropriate radar parameters for gesture movement are calculated by analyzing the millimeter wave radar signal structure and measurement algorithm,and the signal is collected to extract gesture data.Then,the signal data is analyzed in the time and frequency domains,and a frequency-domain approach is proposed to filter out hardware interference and clutter signals other than gestures within the measurement range by designing an algorithm that removes the DC component in the distance and velocity dimensions.To further extract dynamic gesture targets,a S-CFAR algorithm is proposed by combining the advantages and disadvantages of CA-CFAR and OS-CFAR algorithms,effectively extracting gesture targets in the RDM.Secondly,two different methods are studied to design two sets of features.Based on the RDM processed by S-CFAR,time-coupled processing is performed,and the distance-time and velocity-time graphs are separated.On the basis of FFT in the distance dimension,the micro-Doppler characteristics of each gesture are extracted through STFT to form the first set of features.To balance spectral energy and improve sample imbalance,two methods,frame-by-frame normalization and data augmentation,are designed to solve the problem.Finally,the angle information is extracted using the 3D-FFT algorithm,and a radar data cube is constructed to form the second set of features based on the three sections of the Range-Angle-Doppler matrix.Lastly,research on convolutional neural networks with optimized structures is conducted to achieve gesture recognition.Firstly,to address the shortcomings of traditional neural networks,two optimizations,residual structure and separable convolution,are studied to design an improved single-channel neural network,with detailed design details for each part.Secondly,with six sets of features,the network is optimized through multiple experiments,and each dataset is tested.Finally,a multi-branch neural network is designed to extract multiple features,and the network is tested on the two sets of features,achieving an accuracy of 96% and 97.8%,proving the superiority and effectiveness of the network on complex datasets with small and multiple samples. |
