Underwater Motion Estimation Based on Acoustic Images and Deep Learning

This work develops techniques to estimate the motion of an underwater vehicle by processing acoustic images using deep learning (DL). For this, an underwater sonar simulator based on ray-tracing is designed and implemented. The simulator provides the ground truth data to train and validate proposed techniques. Several DL networks are implemented and compared to identify the most suitable for motion estimation using sonar images. The DL methods showed a much lower computation time and more accurate motion estimates compared to a deterministic algorithm. Further improvements of the DL methods are investigated by preprocessing the data before feeding it to the DL network. One technique converts sonar images into vectors by adding up the pixels in each row. This reduces the size of the DL networks. This technique showed significant reduction in the computation time of up to 10 times compared to techniques that use images. Another preprocessing technique divides the field of view (FoV) of a simulated sonar into four quadrants. An image is generated from each quadrant. This is combined with the vector technique by converting the images into vectors and grouping them together as the input of the DL network. The FoV division approach showed a high accuracy compared to using the whole FoV or different portions of it. Another motion estimation method presented in this work is enabled by full-duplex operation and rather than using images, it is based on DL analysis of time variation of complex-valued channel impulse responses. This technique can significantly reduce the acoustic hardware and processing complexity of the DL network and obtain a higher motion estimation accuracy, compared with techniques based on the processing of sonar images. The navigation accuracy of all the techniques is further illustrated by examples of estimation of complex trajectories using simulated and real data.