Random Forest Based Adaptive DBSCAN for Reducing Noise in mmWave Radar Point Clouds

Abstract

The application of millimeter-wave(mmWave) radar in human activity recognition has attracted significant attention because of its insensitivity to ambient lighting and privacy concerns. Millimeter-wave radar will be transformed into a point cloud as an input. This paper proposes a system that dynamically adjusts DBSCAN to handle point cloud noise. It consists of four main components: random forest to determine minPts, KNN to compute Eps, Cluster Merging, and Noise Re-judgment to optimize the point cloud clustering problem in human activity recognition. The proposed new method makes the dynamic selection of the two parameters minPts and Eps of DBSCAN, as well as the design of a new clustering method for human point cloud clustering based on the property of local sparsity of point cloud of the human body taken by radar. We build a dataset based on a millimeter-wave radar of 10 volunteers. Based on this dataset using the Rand Index and Purity assessment, our proposed method has higher accuracy than other methods, reaching average accuracy of 88.52% and 84.86%, respectively, almost 18% higher than other methods.