Passive towed array sonar systems play an essential role in submarine situational awareness. However, the detection and localisation of sound-emitting objects is a more challenging task compared to their active counterparts due to a lack of immediate range information. By making manoeuvres and changing the bearings at multiple positions, a passive sonar can localise and track the source of the sound. Reinforcement learning is the process of learning an optimal strategy to guide an agent’s actions towards optimising its cumulative reward for a given task. This work evaluates an agent’s ability to control a passive towed array sonar system for optimal source localisation and tracking in the underwater environment, using collision avoidance as a practical example application.

In this paper, we discuss the derivations, implementations and performance of target tracking algorithms for a single-target single-sensor system estimating the state, covariance and existence probability of a target. Given the target exists, we simulate measurements of the target with a given probability of detection, along with false measurements (clutter) parametrised by a clutter density. We evaluate the performance of the algorithms by computing the area under Receiver Operating Characteristic (ROC) curves against a range of clutter density values. We give particular attention to the effectiveness of correctly inferring the presence or absence of the target. We select the Integrated Probabilistic Data Association Filter (IPDAF) and the Integrated Expected Likelihood Particle Filter (IELPF) algorithms, with the IELPF implementing a near-optimal proposal which uses the current scan of measurements as well as a prior proposal for comparison. Simulation results indicate the performance of the IPDAF exceeds that of the preexisting particle filter implementing a prior proposal, but a novel particle filter using a near-optimal proposal and a modest number of particles outperforms the IPDAF.

Manifold embedding techniques have properties that render them attractive candidates to learn a compact and general representation of a three dimensional spatial object. In turn this representation can be used for object recognition through classification. This paper presents a comparative study of several supervised spectral embedding techniques and their relationship with the feature space used to describe the exemplars which act as inputs to an embedding procedure. By concentrating on this aspect, we are able to highlight preferential combinations between feature description and embedding, and we formulate recommendations on the use of such methods for fusing multiple views of an object to recognize it under variable poses.

This paper compares the tracking performance that can be achieved when using a nonlinear drag model for a helicopter, a constant drag motion model, and a baseline constant acceleration model. A particle filter is used for state estimation to address problems associated with nonlinear drag and nonlinear measurements of helicopter pose. We demonstrate that the inclusion of this nonlinear kinematic effect provides improved tracking performance for a manoeuvring target.

This paper proposes a novel manifold embedding method for the automated processing of large varied datasets. The method is based on binary classification, where the embeddings are constructed so as to determine one or more unique features for each class individually from a given dataset. The proposed method is applied to examples of multiclass classification that are relevant for large-scale data processing for surveillance (e.g., face recognition), where the aim is to augment decision making by reducing extremely large sets of data to a manageable level before displaying the selected subset of data to a human operator. The method consists of two stages: Preprocessing and embedding computation. In the embedding computation, adaptive measures of intraclass and interclass information are proposed, based on the concepts of “friend closeness” and “enemy dispersion.” In addition, an indicator for weighted pairwise constraint is proposed to balance the contributions from different classes to the final optimization, in order to better control the relative positions between the important data samples from either the same class (intraclass) or different classes (interclass). The effectiveness of the proposed method is evaluated through comparison with seven existing techniques for embedding learning, using four established databases of faces, consisting of various poses, lighting conditions, and facial expressions, as well as two standard text datasets. The proposed method performs better than these existing techniques, especially for cases with small sets of training data samples.

Using low bit depth images for image processing applications offers a number of advantages over full depth images, reduced data transmission, removal of superfluous detail and improved compressibility with potential reduction in FPGA resource usage for use on-board small UAS platforms. It is demonstrated here that Binary Robust Invariant Scalable Key point (BRISK) descriptors can often be matched directly, without modification, between images with differing number of gray levels (bit-depth). The performance is evaluated within the context of obtaining a sufficient number of control points for an image registration problem. This could be used to determine direct equivalence of an unknown, unaligned, reduced palette image against a known reference image of superior bit-depth.

Collision Avoidance is a critical requirement in Aviation safety. Unmanned Aircraft Systems (UASs) are required to adhere to Rules of the Air, outlined in the UK by the Civil Aviation Authority (CAA), and to be able to resolve any potential collision situations. This work investigates rules governing two common situations; converging traffic, and head-on approach. This paper considers the implementation of a generic flight planner using simulated UAS models. Upon detection of a potential collision, an appropriate collision avoidance maneuver is calculated and executed, with the aircraft reverting to their desired path afterwards.

Projection techniques are frequently used as the principal means for the implementation of feature extraction and dimensionality reduction for machine learning applications. A well established and broad class of such projection techniques is the projection pursuit (PP). Its core design parameter is a projection index, which is the driving force in obtaining the transformation function via optimization, and represents in an explicit or implicit way the user's perception of the useful information contained within the datasets. This paper seeks to address the problem related to the design of PP index functions for the linear feature extraction case. We achieve this using an evolutionary search framework, capable of building new indices to fit the properties of the available datasets. The high expressive power of this framework is sustained by a rich set of function primitives. The performance of several PP indices previously proposed by human experts is compared with these automatically generated indices for the task of classification, and results show a decrease in the classification errors.

This paper proposes a novel algorithm for function approximation that extends the standard generalized regression neural network. Instead of a single bandwidth for all the kernels, we employ a multiple-bandwidth configuration. However, unlike previous works that use clustering of the training data for the reduction of the number of bandwidths, we propose a distinct scheme that manages a dramatic bandwidth reduction while preserving the required model complexity. In this scheme, the algorithm partitions the training patterns to groups, where all patterns within each group share the same bandwidth. Grouping relies on the analysis of the local nearest neighbor distance information around the patterns and the principal component analysis with fuzzy clustering. Furthermore, we use a hybrid optimization procedure combining a very efficient variant of the particle swarm optimizer and a quasi-Newton method for global optimization and locally optimal fine-tuning of the network bandwidths. Training is based on the minimization of a flexible adaptation of the leave-one-out validation error that enhances the network generalization. We test the proposed algorithm with real and synthetic datasets, and results show that it exhibits competitive regression performance compared to other techniques.

The existing method for sampling a modulated signal is the same as that for a non-modulated message signal. Therefore, the quadrature sampling rate for a modulated signal must be larger than its bandwidth. However, a new algorithm, which can recover completely a DC-free message signal from an FM signal at a sampling rate less than its bandwidth, is presented. The new sampling rate limit is also discussed. The investigation shows that the current sampling theorem does not present an optimal sampling rate for recovering the message signal from an FM signal and further modification of the theorem might be needed.