Continuous Human Pose Construction Based On Millimeter Wave Radar Point Cloud

As technology advances,there are increasing demands for more realistic and accurate human pose detection in various fields such as leisure,entertainment,and healthcare.Currently,camera-based body motion information collection has become the mainstream method for capturing joint movements,but this method has problems such as high cost,large data volume,high algorithm complexity,and privacy leakage,and requires strict working conditions such as lighting.To address these issues,a system for human skeletal pose estimation based on millimeter-wave radar point clouds is proposed.The system maps 5D time series point clouds to lower dimensions and uses convolutional neural networks(CNN)to estimate the preliminary joint positions in space.To further optimize the results of joint position estimation,this paper introduces the CNN-LSTM model for the first time,which effectively captures the temporal correlation of human pose,thereby improving the system’s accuracy.Finally,the system can reconstruct 19 joints and bones from the point clouds generated by FMCW radar,achieving high-precision human pose estimation.Experimental results show that the average error of all joint positions is 3.54 cm,indicating good accuracy and performance.To verify the applicability of the model,actual human body postures were collected using a FMCW radar development board and Kinect sensor,and the CNN-LSTM model was trained through data collection and processing into the deep learning network.The 3D coordinates of 19 key skeletal points of the human body were outputted,and the construction of human posture movements with high accuracy was realized,demonstrating the feasibility of the system being mounted on edge devices.This research has significant implications for future applications in leisure,entertainment,and healthcare fields,providing a feasible idea and method for achieving high-precision human pose estimation.