Drones have revolutionized the field of aerial imaging, mapping, and disaster recovery. However, drone deployment in low light conditions is limited by the image quality produced by their on-board cameras. In this paper, we present a learning architecture for improving 3D reconstruction in low-light conditions by finding features in a burst. Our approach enhances visual reconstruction by detecting and describing high-quality true features and less spurious features in low signal-to-noise ratio images. We demonstrate that our method is capable of handling challenging scenes in millilux illumination, making it a significant step towards drones operating at night and in extremely low-light applications such as underground mining and search and rescue operations.

This paper presents an autonomous approach to tree detection and segmentation from high resolution airborne LiDAR pointclouds, such as those collected from a UAV, that utilises region-based CNN and 3D-CNN deep learning algorithms. Trees are first detected in 2D before individual trees are further characterised in 3D. If the number of training examples for a site is low, it is shown to be beneficial to transfer a segmentation network learnt from a different site with more training data and fine-tune it. The algorithm was validated using airborne laser scanning over two different commercial pine plantations. The results show that the proposed approach performs favourably in comparison to other methods for tree detection and segmentation.

Recovering information from contrast-limited, SNR-limited, color-attenuated images in a scattering media is of paramount importance for the autonomous functioning of robotic agents. The task is challenging due to the transient state of the medium, unknown medium parameters and in many cases the need for fully autonomous operation. This work presents a target-less, calibration-less method for restoring underwater light field images and requires no explicit model of the medium. The method adopts a light-field imaging approach to capture, model and compensate for backscatter in the scene leading to the recovery of high- fidelity images. The proposed method for backscatter compensation is validated against other state-of-the-art methods and is demonstrated to yield superior image quality.

This paper presents a system for crowdsourcing saliency interest points for robotically gathered 3D maps rendered on smartphones and tablets. An app was created that is capable of interactively rendering 3D reconstructions gathered with an Autonomous Underwater Vehicle. Through hundreds of thousands of logged user interactions with the models we attempt to data-mine salient interest points. To this end we propose two models for calculating saliency from human interaction with the data. The first uses the view frustum of the camera to track the amount of time points are on screen. The second treats the camera's path as a time series and uses a Hidden Markov model to learn the classification of salient and non-salient points. To provide a comparison to existing techniques, several traditional visual saliency approaches are applied to orthographic views of the models' photo-texturing. The results of all approaches are validated with human attention ground truth gathered using a remote gaze-tracking system that recorded the locations of the person's attention while exploring the models.

In this paper we detail a new calibration method for determining the spatial sensitivity of the field-of-view of a spectrometer as well as a method for determining the spatial co-registration of a co-mounted spectrometer and an RGB camera. The spatial sensitivity of the field-of-view of a spectrometer is not necessarily round nor even, and knowledge of this spatial sensitivity is essential for using the spectrometer in an imaging capacity. The combination of the spectrometer and RGB camera provides the capability to produce image data with both high spatial and spectral resolution, which is useful in classification and characterisation for remote sensing and mapping operations.

This paper presents recent developments in data processing of multi-year repeat survey imagery and precision automatic registration for monitoring long-term changes in benthic marine habitats such as coral reefs and kelp forests. Three different methods are presented and compared for precision alignment of imagery maps collected over a range of time-scales from 12 hours to two years between dives. The first method uses Scale Invariant Feature Transform (SIFT) features computed over imagery mosaics to compute the relative translational offset between repeat dives. The second method employs scan-optimisation using the bathymetry generated via structure-from-motion thus capturing more stable features in the environment, lending itself to larger timescale registration. The third method uses mutual information optimisation to register imagery maps, providing robustness to changes in the colour and brightness of objects in an underwater scene across multiple years. Results are presented from field data collected using an Autonomous Underwater Vehicle (AUV) in sites across the Australian coast between 2009 and 2011.

This paper introduces an unsupervised graph cut based object segmentation algorithm, ShadowCut, for robotic aerial surveillance applications. By exploiting the spatial setting of the aerial imagery, ShadowCut algorithm differs from state-of-the-art object segmentation algorithms ([1] [2] [3] [4] [5]) by not requiring a large number of labelled training data set, nor constant user interaction ([6] [7] [8]). In this paper it is shown that, by combining robotic navigation data and a shadow model, it is possible to provide these seed labels with a probabilistic sampling model for object segmentation in aerial imagery. Experiments were performed on aerial data sets consisting of data collected in outback Australia with an aerial robotic platform during an ecological surveillance mission, and aerial images with various natural targets from Google Earth. The segmentation results from the unsupervised ShadowCut algorithm are shown to be comparable with those from supervised graph cut algorithms.

This paper presents a distributed algorithm for performing joint localisation of a team of robots. The mobile robots have heterogeneous sensing capabilities, with some having high quality inertial and exteroceptive sensing, while others have only low quality sensing or none at all. By sharing information, a combined estimate of all robot poses is obtained. Inter-robot range-bearing measurements provide the mechanism for transferring pose information from well-localised vehicles to those less capable. In our proposed formulation, high frequency egocentric data (e.g., odometry, IMU, GPS) is fused locally on each platform. This is the distributed part of the algorithm. Inter-robot measurements, and accompanying state estimates, are communicated to a central server, which generates an optimal minimum mean-squared estimate of all robot poses. This server is easily duplicated for full redundant decentralisation. Communication and computation are efficient due to the sparseness properties of the information-form Gaussian representation. A team of three indoor mobile robots equipped with lasers, odometry and inertial sensing provides experimental verification of the algorithms effectiveness in combining location information.

This paper presents and compares two different approaches to integrating sensor information from an Inertial Measuring Unit (IMU), Global Positioning System (GPS) receiver and monocular vision camera mounted to a low-flying Unmanned Aerial Vehicle (UAV) for building large-scale 3D terrain reconstructions. Both approaches utilise a statistically optimal bundle adjustment formulation that incorporates Vision, IMU and GPS observations into the map and pose optimisation process. Our first approach employs a novel pose-only formulation that optimises relative camera poses based on vision feature matches between frames, while incorporating IMU and GPS information. Our approach is related to, but differs from existing pose-graph techniques by formulating a set of 1D epipolar constraints for features matched between two camera frames, rather than minimising 4D feature re-projection errors, or marginalising feature states. We compare results of the method to a second approach which estimates both 3D features and poses together, using airborne vision, IMU and GPS data collected in an ecology mapping application. The results demonstrate a reduction in the computational complexity during optimisation for the pose-only approach, while producing equivalent accuracy in the reconstructed 3D terrain map.

This paper presents a framework for integrating sensor information from an inertial measuring unit (IMU), Global Positioning System (GPS) receiver and monocular vision camera mounted to a low-flying unmanned aerial vehicle (UAV) for building large-scale terrain reconstructions. Our method seeks to integrate all of the sensor information using a statistically optimal non-linear least squares smoothing algorithm to estimate vehicle poses simultaneously to a dense point feature map of the terrain. A visualisation of the terrain structure is then created by building a textured mesh-surface from the estimated point features. The resulting terrain reconstruction can be used for a range of environmental monitoring missions such as invasive plant detection and biomass mapping.

In this paper we present an observability analysis for a mobile robot performing SLAM with a single monocular camera. The aim is to get a better understanding of the well known intuitive behavior of these systems, such as the need for triangulation to features from different positions in order to get accurate relative pose estimates. The characterisation of the unobservable directions is made using the nullspace basis of the stripped observability matrix. This allows us to identify which vehicle motions are required to maximise the number of observable states in the system, which in turn affects accuracy in the estimation process. The analysis is performed by modelling the system in the continuous time domain as piecewise constant. Simulation results using an extended information filter are shown to verify the results of the observability analysis.

This paper presents an intelligent, on-line guidance scheme for maximising vehicle navigation and snap information for an uninhabited aerial vehicle while operating over unknown terrain. The tasks of localisation and mapping are performed concurrently using the well known extended Kalman filter implementation of the simultaneous localisation and mapping (SLAM) algorithm. A guidance scheme is formulated using vehicle actions in order to maximise the quality of the resulting vehicle pose estimates and environment map by maximising the entropic information in vehicle and map estimates. Results are presented using a high-fidelity six-degree of freedom simulation of an aerial vehicle using inertial navigation and a range/hearing sensor.