Unmanned surface vehicles are becoming a promising solution to many challenges in maritime enviroment, such as exploration of inaccessible areas, surveillance and maritime logistics. Estimation of the USV's near-surrounding distance could potentially be used as a source of additional information for autonomous navigation and for other challenges mentioned above. There is currently no research evaluating how state-of-the-art methods behave in maritime environment. As part of the thesis, we train the DPT model used for monocular depth estimation from colour images of maritime environment, analyse the results and propose new appropirate variations. The MODD2 dataset is used for training and the proposed methods are evaluated with the MaSTR1325 dataset. Both contain images of general maritime environment captured by an USV. Since such vessels are usually equipped with IMUs and both datasets contain time-synchronized IMU measurements in addition to colour images, we propose two additional models that use this added information. Since the two datasets do not contain ground truth depth images, but do contain stereo image pairs, we generated the reference values for training with the CREStereo model which predicts depth images from the stereo image pairs. All proposed models outperform the baseline DPT model in maritime environment.
|
