This work package involves research into hardware acceleration of state-of-the-art algorithms generated in WP1-5 or elsewhere in the signal processing literature, with a focus on real-time implementation.
E_WP 6.1 Efficient parallelization of Sensing Processing
We have addressed challenges #29 (Reducing Size, Weight and Power through efficient processing) and #27 (Accreditable machine learning). A SAR (synthetic aperture radar) algorithm for fast back-projection, developed during previous UDRC work, has been implemented on multicore CPU and GPU architectures. This has reduced the computation time required to generate high-resolution SAR images by several thousand times over naïve back-projection, a significant improvement.
As part of these challenges, we also look at object detection and classification tasks in vision, SAR and sonar scenarios. Expanding on previous work in this area, we study the application of entropy/uncertainty measurements to make automatic detections of potential targets more reliable, and evaluate the computational requirements of doing so. Using machine learning algorithms to generate probabilistic classifications (SVMs, Adaboost, etc.) can produce detection scores which are under- or over-confident. Other ML algorithms such as Gaussian Processes can correct this. By combining these two approaches we obtain fast detections and extract image regions which we are less confident about: these can then be analysed by more compute-expensive algorithms or passed to a human operator for final classification. This approach is more efficient than running slower, accurate algorithms over entire images.
E_WP 6.2 Implementation of Distributed Signal Processing Algorithms
In this work package, we consider the problem of using distributed signal processing across multiple platforms to detect and localise RF signals. In such a distributed network of heterogeneous processing platforms used for multi-sensor tracking and situational awareness, limited energy is available for both on-platform computation and inter-platform communication. We will analyse the most efficient trade-offs between these two tasks and explore power reduction, timely computations and ways of sharing processing between nodes.
E_WP 6.3 Algorithm/computation resource management
This sub-package deals with management of computational resources in multi-sensor systems. We will focus on optimal assignment of computational resources to sensor data and how this may change based on situational priorities, e.g. concentrating on one region or distributing processing resources across sensor modalities, including LiDAR and video.