Channel Knowledge Maps Construction Based on Point Cloud Environment Information

This work focuses on utilizing point cloud data to enhance the accuracy of channel impulse response (CIR) predictions in channel knowledge maps (CKMs). We propose a method, named Point Selector, which uses a set of co-foci ellipsoids to identify reflection and scattering objects in the point cloud. Then, we employ a PointNet neural network that directly utilizes the point cloud from Point Selector to predict the CIR. To validate our method, we employ a multi-view scene reconstruction technique to obtain the point cloud and collect CIR measurements in the real world. Experiments reveal that the proposed method identifies reflection and scattering objects more accurately, and significantly improves CIR prediction to achieve a root mean squared error of 2.88 dB compared to 8.68 dB by using traditional ray tracing methods.