Person Identification Method Based on PointNet++ and Adversarial Network for mmWave Radar

As 3D point-cloud has the ability to present the contour of an object clearly, it provides more spatial information for person identification (PI) task. Aiming at the improvements on quality of point-cloud and distribution of features, an innovative treatment method for point-cloud and a novel network structure are investigated in this paper. Firstly, spatiotemporal feature of point-cloud is enhanced by implementing dual-stage density-based spatial clustering of applications with noise (DST-DBSCAN) method, which can filter most invalid points and decrease the sparsity of point-cloud. After that, the optimized point-cloud is input into neural network, which contains three parts for feature extraction, classification and feature optimization. Specifically, PointNet++ is adopted to extract features and realize PI recognition. In addition, an adversarial network is designed for optimizing feature distribution of point-clouds by encouraging the feature extractor of PointNet++ to generate features of the same person as similar as possible. Experimental results demonstrate that the proposed method can improve the accuracy by 3.77% than original PointNet++ network with raw data.