A human posture recognition method for millimeter wave radar sparse point cloud
Yucong Liu, Chao Zhou, Xingming Li, Shanqing Hu · IET conference proceedings. · 2024
Millimeter-wave radar has received considerable attention for its applications in human activity recognition in residential environments due to its insensitivity to lighting conditions and strong privacy features. However, the direct application in posture recognition is hindered by the sparse nature of the point cloud data generated by millimeter -wave radar, along with the presence of considerable noise points. To address these issues, this study proposes an intelligent posture recognition method for sparse point cloud data. Firstly, a multi-frame accumulation technique is employed to generate point cloud data containing the complete posture movements. Subsequently, the DBSCAN algorithm is applied to perform noise reduction on the accumulated data. Finally, a signal-to-noise ratio based weighing approach is utilized to enhance the feature representation of the sample data. To enhance recognition accuracy, this study employs 4-dimensional point cloud data that includes velocity information as input, utilizing an expanded PointNet architecture for posture recognition. Experimental outcomes show that the proposed method effectively reduces noise and achieves a remarkable average recognition accuracy of 96.1% for four distinct postures: falling, hunkering, bowing and jumping.