A sensor network that connects multiple 3D image sensor devices using light detection and ranging (LIDAR) is effective for collecting spatial information to detect potential risks associated with people's movements. However, due to the large volume and high frame rate of point cloud data obtained by LIDAR devices, under the condition of strictly limited bandwidth, some part of the data can go missing when detection is performed from point cloud data, causing a degradation of the accuracy. As the communication traffic volume of other applications increases, the communication available bandwidth for spatial information applications becomes insufficient. This paper proposes a scheme of spatial feature-based prioritization for the transmission of point cloud data with strictly limited communication bandwidth. It estimates how much each piece of point cloud data contributes to the accuracy improvement of a machine learning task for spatial prediction and assigns a higher score to those with a larger contribution. These scores prioritize the control for the transmission of point cloud data from sensor devices to an edge server. Feature selection with machine learning is helpful for estimating data importance. To improve the estimation accuracy, the proposed scheme complements missing entries when performing feature selection. We developed an indoor experimental environment and found that the prediction accuracy was sufficiently maintained compared to two benchmark schemes, random and heuristic, with limited communication bandwidth. Comparison with the perturbation method, which is a conventional method of feature selection, demonstrated the effectiveness to complement missing values in the proposed scheme.
ASJC Scopus subject areas