A Local Anomaly Detection Method for Occluded Targets with Multispectral Point Clouds

Multispectral point clouds (MPC) collected by Unmanned Aerial Vehicle (UAV) simultaneously encompass spectral-spatial information of land covers, making occluded target detection possible. Research on anomaly detection based on UAV-mounted MPCs is limited due to the high dimensionality of unstructured data, which increases the computational load, and anomalies that pollute the global background, leading to high false alarm rates. These factors limit the efficiency and accuracy of detection, particularly for occluded targets. A local anomaly detection method based on supervoxel segmentation has been proposed for MPC data. The false alarm rate is reduced through supervoxel segmentation, and then a spherical model is used for localized background selection to minimize false alarms caused by background contamination from anomalies. Compared to applying advanced global anomaly detectors directly to MPCs, the proposed method achieves higher detection accuracy and is more capable of detecting occluded targets.