Autonomous 3D Semantic Mapping of Coral Reefs

This paper presents the first-ever approach for autonomous 3D semantic mapping of coral reefs. The position of corals in 3D coordinates and the type of the coral are presented in such a3D semantic map. The intended application of this work is coral reef health monitoring, as the current assessment is based entirely on direct or indirect human observation. The proposed system joins a convolutional neural network (CNN) with a direct visual odometry approach and a correlation filter based tracker, Kernelized Correlation Filter (KCF), to identify the different coral species detected. In addition to the coral classification, the 3D position of each coral is identified producing a semantic map of the observed reef. Each coral is identified once and tracked to prevent a recount. The number of different coral species encountered in two separate traversed areas is reported. Furthermore, the shape and size of a coral can be extracted from the 3D reconstruction enabling the extraction of volumetric data for subsequent studies. Experimental results from the coral reefs of Barbados verify the robustness and accuracy of the proposed approach.