Perception research

Perception is the process of constructing an internal representation of the world from real-time sensory information. Perception by autonomous systems, in unstructured dynamic worlds, is one of the outstanding and significant research challenges in building intelligent systems. The main outcomes of the perception research program are the development of innovative new methods and algorithms able to generate and interpret the information-rich representations of complex environments required for the next generation autonomous systems.

Research in Perception is categorised into the following four areas:

1. Sensors and sensing

We have significant expertise in the development and use of different types of sensors for autonomous systems including; vision, infra-red and other electrooptic (EO) arrays, mm-wave radar and acoustic sensors. Sensor research focuses on:

  • The development of novel sensors for autonomous systems including noise radar, low-cost hyperspectral imagers, and MEMs-based vehicle sensors.
  • Sensor system architectures including the use of embedded FPGAs for image and signal processing, modular hardware and interface design, and
  • The development of integrated multi-sensor systems, especially common-mode radar and EO sensors, and tight-coupled GPS-inertial systems.

2. Representations

Effective representations of the environment lie at the heart of the perception challenge. We are recognised innovators in the development of sensor-derived representations for autonomous agents. Research on representations focuses on:

  • Efficient modelling and manipulation of parametric environment models.
  • The use of tessellation and compression methods, especially avelet models, in describing and understanding complex outdoor scenes, and
  • Integrated representations of non-geometric properties such as texture, colour and dynamic object descriptions.

3. Uncertainty

The representation and management of information and uncertainty is inherent in perception with real sensors in real environments. We play a leading international role in this field. Research in uncertainty focuses on:

  • Probabilistic sensor modelling incorporating full density models of the physics of observation and detection.
  • Efficient Bayesian engines for the manipulation and management of different probability representations, and
  • Algorithms for compression, abstraction and refinement of probabilistic data, focusing particularly on the use of information measures. These three areas represent a real departure from current methods and approaches in defining a fundamental basis for management of perceptual uncertainty.

4. Data Fusion

Data fusion is the process of putting together information obtained from many heterogeneous sensors, on many platforms, into a single composite picture of the environment. We have a considerable track record in data fusion for both military and robotic applications, especially in information-theoretic decentralised data fusion methods pioneered in the group. Data fusion research focuses on:

  • The development of novel sensor-centric data fusion models appropriate to fusion of disparate range and image data.
  • Fusion of complex, information-rich, probabilistic representations of geometric and physical properties, and
  • Development of endogenous data fusion and communication algorithms appropriate to large-scale distributed sensor networks.