Sidescan sonar images are 2D representations of the seabed. The pixel location encodes distance from the sonar and along track coordinate. Thus one dimension is lacking for generating bathymetric maps from sidescan. The intensities of the return signals do, however, contain some information about this missing dimension. Just as shading gives clues to depth in camera images, these intensities can be used to estimate bathymetric profiles. The authors investigate the feasibility of using data driven methods to do this estimation. They include quantitative evaluations of two pixel-to-pixel convolutional neural networks trained as standard regression networks and using conditional generative adversarial network loss functions. Some interesting conclusions are presented as to when to use each training method.

Pipeline inspection is a very human intensive taskand automation could improve efficiencies significantly. We propose a system that could allow an autonomous underwater vehicle (AUV), to detect pipeline damage in a stream of images.Our classifiers were based on transfer learning from pre-trained convolutional neural networks (CNN). This allows us to achieve good results despite relatively few training examples of damage. We test the approach using data from an actual pipeline inspection.

This paper studies the problem of detection and tracking of general objects with semistatic dynamics observed by a mobile robot moving in a large environment. A key problem is that due to the environment scale, the robot can only observe a subset of the objects at any given time. Since some time passes between observations of objects in different places, the objects might be moved when the robot is not there. We propose a model for this movement in which the objects typically only move locally, but with some small probability they jump longer distances through what we call global motion. For filtering, we decompose the posterior over local and global movements into two linked processes. The posterior over the global movements and measurement associations is sampled, while we track the local movement analytically using Kalman filters. This novel filter is evaluated on point cloud data gathered autonomously by a mobile robot over an extended period of time. We show that tracking jumping objects is feasible, and that the proposed probabilistic treatment outperforms previous methods when applied to real world data. The key to efficient probabilistic tracking in this scenario is focused sampling of the object posteriors.

In this letter, we present a deep learning-based network, GCNv2, for generation of keypoints and descriptors. GCNv2 is built on our previous method, GCN, a network trained for 3D projective geometry. GCNv2 is designed with a binary descriptor vector as the ORB feature so that it can easily replace ORB in systems such as ORB-SLAM2. GCNv2 significantly improves the computational efficiency over GCN that was only able to run on desktop hardware. We show how a modified version of ORBSLAM2 using GCNv2 features runs on a Jetson TX2, an embedded low-power platform. Experimental results show that GCNv2 retains comparable accuracy as GCN and that it is robust enough to use for control of a flying drone. Source code is available at: https://github.com/jiexiong2016/GCNv2_SLAM.

Highly automated driving systems are required to make robust decisions in many complex driving environments, such as urban intersections with high traffic. In order to make as informed and safe decisions as possible, it is necessary for the system to be able to predict the future maneuvers and positions of other traffic agents, as well as to provide information about the uncertainty in the prediction to the decision making module. While Bayesian approaches are a natural way of modeling uncertainty, recently deep learning-based methods have emerged to address this need as well. However, balancing the computational and system complexity, while also taking into account agent interactions and uncertainties, remains a difficult task. The work presented in this paper proposes a method of producing predictions of other traffic agents' trajectories in intersections with a singular Deep Learning module, while incorporating uncertainty and the interactions between traffic participants. The accuracy of the generated predictions is tested on a simulated intersection with a high level of interaction between agents, and different methods of incorporating uncertainty are compared. Preliminary results show that the CVAE-based method produces qualitatively and quantitatively better measurements of uncertainty and manage to more accurately assign probability to the future occupied space of traffic agents.

In this letter, we proposed a new deep learning based dense monocular simultaneous localization and mapping (SLAM) method. Compared to existing methods, the proposed framework constructs a dense three-dimensional (3-D) model via a sparse to dense mapping using learned surface normals. With single view learned depth estimation as prior for monocular visual odometry, we obtain both accurate positioning and high-quality depth reconstruction. The depth and normal are predicted by a single network trained in a tightly coupled manner. Experimental results show that our method significantly improves the performance of visual tracking and depth prediction in comparison to the state-of-the-art in deep monocular dense SLAM.

On-the-fly registration of overlapping multibeam images is important for path planning by AUVs performing underwater surveys. In order to meet specification on such things as survey accuracy, coverage and density, precise corrections to the AUV trajectory while underway are required. There are fast methods for aligning point clouds that have been developed for robots. We compare several state of the art methods to align point clouds of large, unstructured, sub-aquatic areas to build a global map. We first collect the multibeam point clouds into smaller submaps that are then aligned using variations of the ICP algorithm. This alignment step can be applied if the error in AUV pose is small. It would be the final step in correcting a larger error on loop closing where a place recognition and a rough alignment would precede it. In the case of a lawn mower pattern survey it would be making more continuous corrections to small errors in the overlap between parallel lines. In this work we compare different methods for registration in order to determine the most suitable one for underwater terrain mapping. To do so, we benchmark the current state of the art solutions according to an error metrics and show the results.

On-the-fly registration of overlapping multi-beam images is important for path planning by AUVs per-forming underwater surveys. In order to meet specificationon such things as survey accuracy, coverage and density,precise corrections to the AUV trajectory while underwayare required. There are fast methods for aligning pointclouds that have been developed for robots. We compareseveral state of the art methods to align point clouds oflarge, unstructured, sub-aquatic areas to build a globalmap. We first collect the multibeam point clouds intosmaller submaps that are then aligned using variationsof the ICP algorithm. This alignment step can be appliedif the error in AUV pose is small. It would be the finalstep in correcting a larger error on loop closing where aplace recognition and a rough alignment would precedeit. In the case of a lawn mower pattern survey it would bemaking more continuous corrections to small errors in theoverlap between parallel lines. In this work we comparedifferent methods for registration in order to determinethe most suitable one for underwater terrain mapping. Todo so, we benchmark the current state of the art solutionsaccording to an error metrics and show the results.

Mobile robot navigation in complex and dynamic environments is a challenging but important problem. Reinforcement learning approaches fail to solve these tasks efficiently due to reward sparsities, temporal complexities and high-dimensionality of sensorimotor spaces which are inherent in such problems. We present a novel approach to train action policies to acquire navigation skills for wheel-legged robots using deep reinforcement learning. The policy maps height-map image observations to motor commands to navigate to a target position while avoiding obstacles. We propose to acquire the multifaceted navigation skill by learning and exploiting a number of manageable navigation behaviors. We also introduce a domain randomization technique to improve the versatility of the training samples. We demonstrate experimentally a significant improvement in terms of data-efficiency, success rate, robustness against irrelevant sensory data, and also the quality of the maneuver skills.

In this paper, we propose a new learning scheme for generating geometric correspondences to be used for visual odometry. A convolutional neural network (CNN) combined with a recurrent neural network (RNN) are trained together to detect the location of keypoints as well as to generate corresponding descriptors in one unified structure. The network is optimized by warping points from source frame to reference frame, with a rigid body transform. Essentially, learning from warping. The overall training is focused on movements of the camera rather than movements within the image, which leads to better consistency in the matching and ultimately better motion estimation. Experimental results show that the proposed method achieves better results than both related deep learning and hand crafted methods. Furthermore, as a demonstration of the promise of our method we use a naive SLAM implementation based on these keypoints and get a performance on par with ORB-SLAM.