Adaptive compression of geospatial LiDAR data and images for metal clutter detection

Abstract

The work is devoted to solving the urgent scientific and practical problem of automated detection of chaotic metal clutter (fragments of reinforcement, man-made debris) in areas of destruction and natural disasters using multimodal data from unmanned aerial vehicles. The work proposes the use of adaptive compression of input spatial-spectral image data and lidar for further intelligent analysis using the architecture neural network during data preprocessing. To reduce the computational load on the onboard systems of unmanned aerial vehicles and minimize background redundancy, the use of an algorithm for primary statistical filtering is proposed. The algorithm combines threshold filtering based on the three-sigma rule for cutting off random geometric noise of the depth map and the quantile filtering method for localization of spectral anomalies (peaks of the intensity of laser radiation reflection or metallic sheen). This allowed to select informative areas containing a high density of potentially metallic objects, and to cut off homogeneous areas of the underlying surface (soil, concrete slabs) before the neural network recognition stage. Experimental studies of the method were conducted on the basis of synthetic images of chaotic debris generated in a simulation environment. The simulation results demonstrated that the sequential implementation of threshold and spatial-spectral filtering by quantiles at the data preprocessing stage allows to reduce the volume of data processed by the neural network by ninety percent while maintaining an acceptable accuracy of debris segmentation. These results can serve as the basis for creating effective on-board systems for monitoring emergency zones in real time.