We derive a probabilistic framework for robust, realtime, visual tracking of multiple previously unseen objects from a moving camera. This framework models the discrete depth ordering of the objects being tracked in the scene. The method uses the observed image data to compute a posterior over the objects' poses, shapes and relative depths. The poses are group transformations, the shapes are implicit contours represented using level-sets and the relative depths give the discrete depth ordering of the objects. All nuisance variables are marginalised out at the pixel-level resulting in a pixel-wise posterior, as opposed to a pixel-wise likelihood, and we show using quantitative results that this provides increased resilience to noise. We also demonstrate how motion models can be incorporated within the same probabilistic framework and show how this enables the system to track complete occlusions. The effectiveness of our method is demonstrated on a variety of challenging video sequences.

We present a hybrid SLAM system for marine environments that combines cubic splines to represent the trajectories of dynamic objects, point features to represent stationary objects and an occupancy grid to represent land masses. This hybrid representation enables SLAM to be applied in environments with moving objects, where solutions using point features alone are computationally prohibitive or where dense objects e.g. landmasses can not be represented correctly using point features. Estimation is achieved using a sliding window framework with reversible data-association and reversible model-selection. Our main contributions are: (i) a hybrid representation of the environment; (ii) occupancy grid fusion is continually refined for the duration of the sliding window; (iii) the trajectories of dynamic objects are represented using cubic splines and (iv) radar scans are re-rendered at a sub-scan resolution to compensate for the egomotion during the scan acquisition period. We show that the continual refinement of the occupancy grid greatly improves the quality of the resultant map, leading to a better estimate of the egomotion and therefore better estimates of the trajectories of dynamic objects. We also demonstrate that the use of cubic splines to represent trajectories has two major advantages: (i) the state space is compressed i.e. many vehicle poses can be represented using a single spline section and (ii) the trajectory becomes continuous and so fusing information from asynchronous sensors running at multiple frequencies becomes trivial. The efficacy of our system is demonstrated using real marine radar data, showing that it can successfully estimate the positions/velocities of objects and landmasses observed during a typical voyage on a small boat.

We describe progress in the design, build and test of a mechatronic system capable of visually tracking objects at sea from a moving platform, such as a boat. The mechanism comprises three controllable degrees of freedom that can carry a small camera as payload. The camera position is stabilised using inertial sensing, and the stabilised images are processed using a published colour-based tracking algorithm to achieve visual pursuit of the target [1], [3]. We describe novel improvements to the tracker that enhance its performance in the case of tracking a target at sea. The system is demonstrated on real footage and its performance assessed.