Pose-Graph Semantic SLAM for Underwater Targets Detection and Localization

This paper presents a Semantic Simultaneous Localization And Mapping (SLAM) strategy for the navigation of Autonomous Underwater Vehicles (AUVs). Every frame acquired from an onboard camera is processed by an Automatic Target Recognition (ATR) module that outputs the bounding box of each identified object. This information is then used to improve the accuracy of the localization of both the vehicle and the landmarks found, in a pose-graph framework. A distance-based criterion is employed for performing the data association process, while the back-end optimization leverages also measurements coming from a Doppler Velocity Log (DVL), a depth sensor, an Inertial Measurements Unit (IMU) and a Fibre-Optic Gyroscope (FOG). This strategy has been tested both in a simulated environment and in a real one, using data collected during the RAMI 2023 competition held in La Spezia, Italy, and it has shown promising results.