Robust Control Theory and Design
Robust Control Applications
Model Predictive Control
Control of Hierarchical and Hybrid Systems
SPEECH, VISION and ROBOTICS
Computer Vision and Robotics
Three Dimensional Ultrasound Imaging
Mechanical Properties of the Newborn Skull
Document Image Processing
LABORATORY for COMMUNICATIONS ENGINEERING
Research in the Information Engineering Division concerns the generation, distribution, analysis and use of information in engineering systems, with the principal interests in system level aspects and processing of signals in contrast to innovation in devices. The information considered include measurements in a flight control systems, speech input to machines, vision, music and image data, and data in ubiquitous communication systems. The generation and transmission of data is expanding at an enormous rate, as is the computer processing power to analyse it; however huge challenges remain to exploit the opportunities of increasingly complex systems.
The research work of the Control Group ranges from theoretical studies which focus on the fundamental limitations in feedback systems and methodologies for control system design, through to very applied work in collaboration with outside organisations. John Lygeros joined the group as a new lecturer in the last year working in the emerging area of hybrid systems.
Professor K. Glover
Dr M.C. Smith
Dr G. Vinnicombe
Research effort has continued on input-output approaches to control systems. A generalisation of the gap metric approach for uncertainty studies and controller design was developed for spatially distributed systems(K131). The ideas of the non-linear gap metric were applied to find a solution of a robust control problem occurring in repetitive control which is not soluble by a linear controller(K111). Work on the problem of vehicle suspension control using ideas from mechanical and electrical network theory contributed to the understanding of the capabilities and limitations of active and passive schemes(K150). The scientific contributions to the control field and intellectual legacy of a distinguished scholar and member of the Cambridge control group, Dr Tom Fuller, who recently passed away, were reviewed and celebrated(K149).
A nonlinear generalization of the nu-gap metric has been obtained(K167). This provides a distance measure for nonlinear systems which captures and exploits the inherent robustness of feedback systems.
Previous work addressing actuator saturation within a robust control framework has been extended to more general substitutions at the plant input. In particular, novel approaches to the problem of switching between feedback controllers have been reported(K125).
Periodically time-varying systems form an interesting class of system where results similar to the time-invariant one can be derived but this generalisation is non-trivial. Sampled-data systems form an important application area of such results. Recent results are presented(K23,K24), which considered the representation of the graphs of several systems and the nu-gap in this context.
Professor K. Glover
A substantial achievement in the period was the successful test flight of a new control methodology on the DERA VAAC Harrier. This controller was designed using recent developments in the control of linear parameter varying (LPV) systems and the approximation of nonlinear systems by quasi-LPV models, and was one of the most successful first flights on this test vehicle. Previous results using similar techniques on a wind-tunnel model of a high incidence research model (HIRM) from DERA were published(K124).
The collaboration between control and the research on combustion engines of Professor Collings continues to produce interesting results on the effectiveness of advanced control in this area (see also section A). Particular results on improved idle-speed control were reported(K60) with significant potential of improved fuel economy by a novel nonlinear controller formulation. The oxygen storage capacity of the three-way automotive catalyst is crucial to its effective operation and the dynamics of this have been studied(K41).
The investigation of thermo-acoustic oscillations in gas turbine is reported in section A. The control of real oscillations has been presented(K34), using both adaptive and robust control for a realistic simulation model.
Dr J.M. Maciejowski
Dr G. Vinnicombe
Research on extending subspace methods to the identification of discrete-time bilinear systems has continued. There appears to be a whole family of such algorithms, each member of it providing a different accuracy/complexity trade-off. All of these algorithms still require very large amounts of storage and computation time, and we have recently succeeded in reducing both of these, but with some accompanying loss of accuracy(K28,K29,K30).
A technique for identification in the nu-gap metric has been reported(K42) as the first stage of a project aimed at the development of a comprehensive framework for the identification of models suitable for control system design.
Dr J.M. Maciejowski
Work on Model Predictive Control (MPC) has contributed to two different aspects, both of which are important for extending the range of applicability of MPC. The first is the question of avoiding infeasibility in constrained MPC problems. We have contributed a specific multi-objective formulation to represent various constraint priorities, which is also applicable to the control of hybrid systems(K100), we have investigated how to obtain exact penalty functions for 'softening' constraints(K101), and we have shown how to reduce the conservativeness of feasibility prediction in a particular scheme proposed for MPC(K140). The second aspect is that of designing the linear mode of operation of MPC, when constraints are inactive. We have shown how to choose MPC parameters such as to reconstruct the behaviour obtained by an H-infinity loop-shaping design in the linear mode(K141). See also www research page URL: http://www-control.eng.cam.ac.uk/jmm/.
Dr J. Lygeros
Hybrid systems are a class of dynamical systems that involve the interaction of continuous and discrete dynamics. This is a fairly new but very active research area, that lies in the intersection of control theory and computer science. Much of the activity in this area has been motivated by its importance to large scale systems involving embedded computation.
During this period research was conducted on the theoretical foundations of hierarchical and hybrid control and on applications to Air Traffic Management and Automated Highway Systems.
Dynamical properties of general hybrid systems, in particular existence of solutions and stability were studied. Implications of the results for the simulation of hybrid systems and path planning for mobile robotics were investigated.
Research was conducted into methods for designing hybrid controllers that meet safety specifications(K153). The work was motivated by applications to Air Traffic Management and Automated Highway Systems(K153).
Extensions of deductive methods from theoretical computer science to the hybrid domain were developed. The motivating application was the verification of on-board collision avoidance systems for aircraft(K114).
Finally, extensions of the hybrid paradigm to include probabilistic dynamics were investigated. Theoretical work was motivated by collision avoidance in air traffic control highway collision avoidance, and fault tolerant automated highway design.
Professor P.J.W. Rayner
Dr W.J. Fitzgerald
Dr N.G. Kingsbury
Dr S.J. Godsill
Dr M.D. Macleod
Dr J. Lasenby
Work this year has continued with Markov Chain Monte Carlo methods, Sequential Markov Chain Monte Carlo methods and Particle filters(K5,K8,K50) applied to a range of different application areas:
1) In nuclear spectroscopy, the spectra contain an unknown number of spectral components with unknown positions and widths and the spectral lines have been convolved with a known point spread function. EP priors(K10,K17) have been used to enable the solution of the model selection problem and excellent results have been obtained. The next phase of the work will apply the approach to real nuclear physics data.
2) Particle filters are being used for the tracking of radar targets in collaboration with DERA - here, both estimation and model order (how many targets) are being investigated. It has been shown that particle filters offer considerable advantages over the standard extended Kalman filter approach - MATLAB demonstrations, working in 'real-time' have been developed. Work is under way, again supported by DERA, for the fusion of multiple sensor data which will be used in various tracking applications(K18,K49,K71,K75).
3) Particle filters are also being applied to various communications applications, for example, BPSK and QAM schemes and it has been shown that the performance significantly outperforms current methods as measured by the bit-error rate. Work is continuing by applying these methods to realistic CDMA schemes(K11,K129).
4) Particle filters have also been applied to speech enhancement and restoration(K163,K165).
A collaboration between the Communications Laboratory and the Signal Processing and Communications Laboratory has investigated stable pre-coders for indoor radio communications.
A general framework for unsupervised image segmentation has been developed and applied, with much success, to the segmentation of colour images and for the retrieval of coloured images from data bases(K91,K92,K93,K94,K95). These methods rely on Bayesian multi-scale random field models and uses Markov chain Monte Carlo techniques and reversible jump methods in particular.
Signal processing in the presence of heavy tailed noise has continued and has investigated beam-forming and communications applications as well as incorporating the time-frequency framework in the presence of alpha stable noise(K6,K7,K9,K19).
Work has continued using extreme value statistics with applications to the insurance industry - the Signal Processing and Communications Laboratory (W.J. Fitzgerald) is organising another Isaac Newton Institute programme on Extreme Value Statistics this summer.
Work is starting concerning the application of Bayesian and Particle filter methods to the area of Quantum Mechanics and Quantum estimation.
Work in on-line inference using Monte Carlo methods ('particle filters') has continued(K4,K13,K14). In particular, new methodology for Monte Carlo smoothing has been developed, in collaboration with researchers at Duke University(K47) which allows for very accurate estimation of non-linear/non-Gaussian state space models through a new class of forward and backward simulation-smoothing techniques(K47,K70). The methods have been applied successfully to degraded speech, audio and EEG datasets. Further new methods for on-line filtering based on intermediate bridging densities, which help to alleviate tracking problems in the presence of abrupt changes, have been presented and a well-received paper has been published on general sequential simulation methodology(K46). For batch-based inference, our earlier Markov chain Monte Carlo methods for inference in symmetric alpha-stable noise, using an exact scale-mixtures representation which eliminates problems of density evaluation or approximation using finite mixtures, have been improved by the introduction of a slice-sampling technique(K69).
In the genomics area progress has been made in improved base-calling methods for DNA sequencing, using rigorous physically motivated modelling for the first time(K76). This work has received substantial industrial attention and is likely to be expanded into other related areas next year.
Work has continued on the successful source separation project sponsored by AT&T Cambridge Research Laboratories producing some ground-breaking research in separation of time-varying convolutive mixtures of audio signals, and a patent application has duly been made(K12). Alternative methods for blind source separation have been developed which utilize non-stationarity of the source signals(K1,K2). Non-stationarity has also been used to develop a Bayesian technique for blind de-convolution with particular application to de-reverberation of room acoustics(K82). Further work on the optimal incorporation of perceptual models into audio processing algorithms has led to new and improved rules for suppression of noise in audio recordings(K173). The automatic transcription of polyphonic musical signals has also been advanced by the incorporation of multiple frame data into our earlier algorithms(K168). Finally, optimal enhancement of speech using time-varying state-space models has shown some successful results(K164,K165).
The main focus of image processing research this year has been based on exploring applications of wavelets to a range of image analysis and enhancement problems, although there has also been some work on coding. An earlier coding project on vector quantisation of hierarchical trees of wavelet coefficients has been completed and the results published(K166); and work has continued on the exploration of good ways to encode images, which have been transformed into a redundant wavelet space using the dual-tree complex wavelet transform (DT CWT), developed recently in the Signal Processing and Communications Laboratory(K130).
A particular theme of content-based image retrieval was given added impetus with the award in April 2000 of a European Commission research grant on Methods of Unified Multimedia Information Retrieval (MOUMIR). We are one of seven partners in this project, which draws on our recent work in the area of image segmentation and region classification(K43,K44,K95,K138). Our DT CWT has played a key role in this work, part of which was done in collaboration with Rice University in Houston(K34,K138). We have also been collaborating with INRIA in Sophia Antipolis, France, on the use of complex wavelets for enhancement and segmentation of satellite and aerial images, using an Alliance Grant from the British Council and through the MOUMIR project. Several promising approaches to Bayesian de-convolution and de-noising (enhancement) of images in the complex wavelet domain have been developed and results will be published in the near future.
We have published extensively on the DT CWT this year, both from a somewhat theoretical viewpoint(K102,K103,K104), showing how to design complex wavelet filters with desirable properties, and also from an application viewpoint, demonstrating improved performance for tasks such as texture analysis(K43,K79), image watermarking(K115,K116,K117) and image segmentation and region classification, discussed above. The new Q-shift version of the DT CWT(K103) represents a significant advance in the simplicity and orthogonality of the transform. Work on texture description(K79) was performed in collaboration with the University of California, Santa Barbara.
Work continued on segmentation of both black and white and coloured images, using efficient non-linear multi-scale algorithms(K85,K86).
This past year the main focus of the work on markered optical motion capture(K110) has been to develop robust and efficient tracking techniques when the system has a small number of cameras and therefore suffers from frequent occlusions. A model-based approach uses a detailed articulated model(K108) and multiple Extended Kalman Filter and particle filter algorithms(K133,K134,K135). Work is continuing on how to optimally place markers relative to joints and how to estimate accurate joint positions from the data (skeleton fitting).
Work on geometric algebra (GA)(K107,K109) has included formulations which remove the singularities from the calculation of the inverse kinematics of linked chains with specified joint motions, building efficient articulated skeletal models for tracking and using the rotor formulation of rotations for efficient interpolation(K80,K108). The 5D conformal geometric algebra has also been used as a means of effectively intersecting rays and surfaces in ray tracing and wave-front scattering applications(K106). A course on GA for computer graphics was given by a group from Cambridge, USA, Canada, Holland and Sweden at Siggraph2000.
The ability to perform 3D motion reconstruction and analysis has been used in a number of projects, including classification of gait, investigation of human interaction with moving structures (with the CUED structures group) and evaluation of technique in sport. A collaboration which involves a study of 3D malalignment in the shoulders of post-stroke patients is underway.
Professor S.J. Young
Mr P.C. Woodland
Dr M.J.F. Gales
Dr T.R. Niesler
Dr A.J. Robinson
Work in the speech processing area is focussed mainly on large vocabulary speech recognition and its applications, especially in spoken document retrieval (SDR) and spoken dialogue systems.
In terms of core development, work has been published on a variety of methods for improving the performance of HMM-based speech recognition systems. Early pioneering work by the group on the use of maximum likelihood linear regression for speaker adaptation was recognised by a best paper award from Computer, Speech and Language and an invited review of the state of the art on speaker adaptation technology has been presented(K175). New techniques for very rapid speaker adaptation schemes have been developed(K61). These allow systems to adapt to an individual speaker using only a single sentence of data. Other rapid adaptation schemes were used in the IBM Broadcast News Transcription system. Techniques for improving both the feature sets used for speech recognition, via maximum likelihood projections(K62), and the form of state distribution modelling have been studied(K64). Schemes for decreasing the computational load of these techniques have also been developed(K3).
In parallel with the core work on HMM-based speech recognition, a number of alternative approaches are being studied. The overall goal is to find models which more closely match human speech production yet retain the ability to be trained automatically. The approaches currently being studied include explicit articulatory models trained from X-ray data(K20) and parallel combinations of HMMs.
Work on the statistical modelling of spoken dialogues continues(K179) with particular emphasis on user simulation(K143,K144). Other application related work includes the completion of EC supported work on the application of speech recognition to language learning(K172).
The EPSRC sponsored Multimedia Document Retrieval project (in collaboration with the Computer Laboratory) is focussed on indexing and retrieving audio documents that are created from a stream of audio, such as recordings of news broadcasts. The MDR project is using the HTK broadcast news system to automatically create transcriptions and information retrieval techniques to index the data and return portions relevant to queries. The basic approach including the integration of statistical and knowledge-based techniques was described(K90). This work was evaluated at the TREC-8 conference and the Cambridge system was shown to yield state-of-the-art results(K88). Recent work has focused on the indexing and retrieval of complete broadcast news shows with no prior information regarding story boundary location(K87) as well as investigating the effect of query items that do not appear in the speech recognition vocabulary on retrieval performance(K176). The project has also developed techniques for audio search which can be used for the detection of advertisements in the broadcast news stream(K89). A demonstration system has been produced by the project which can retrieve data from automatically transcribed news broadcasts from the Internet(K158,K159). In 1999 CUED hosted a workshop on "Accessing information in spoken audio", which was well attended and resulted in a special issue of the Speech Communication Journal. In addition, the issue contained two papers from the group(K90,K132).
The GCHQ sponsored project on the Automatic Recognition of Conversational Telephone speech has continued with excellent progress being made. In particular work on discriminative training, developed by a PhD student and partly funded by the project has produced significant improvements in performance(K127,K177,K178). This work has demonstrated for the first time that for the largest and most difficult speech recognition tasks discriminative training can be beneficial, which has lead to widespread interest in the techniques developed. Another significant component of recent improvements was the use of posterior probability decoding during speech recognition search and in system combination(K57,K58). These improvements, and many others, led to the CUED HTK system yielding the lowest word error rate by a significant margin in the NIST-sponsored March 2000 evaluation of research systems for the automatic transcription of telephone conversations(K77).
Professor R. Cipolla
Research in computer vision and robotics has continued to develop new theories for recovering the three-dimensional shape of visible surfaces from images taken from arbitrary viewpoints with uncalibrated cameras with important applications in 3D model acquisition, visual tracking and the guidance of robots(K36,K37,K38).
Novel contributions have been made in the analysis of curved surfaces (where the dominant image feature is the silhouette or outline). It is well-known that it is possible to reconstruct the shape of curved surfaces from the family of outlines obtained by looking at them from different but known viewpoints(K38). We have also been able to show that viewer motion can also be recovered from the envelope of consecutive outlines - the frontier(K142).
A particularly simple and elegant solution was found for a special type of motion in which an object is placed on a turntable which is rotated in front of a stationary camera. A novel algorithm was introduced which exploited the symmetry in the envelope of outlines swept out by the rotating surface(K121,K122,K174). This technique uses a single curve tracked over the image sequence and has been successfully used to recover the shape of an arbitrary object from an uncalibrated camera.
Research in the visual guidance of robots has continued(K51,K52) and resulted in the development of real-time 2 and 3-dimensional tracking system which exploits Lie algebraic techniques and robust methods(K33,K52). This has been incorporated within the control loop of a robot manipulator which can be taught positions in the co-ordinate frame of its workpiece by presenting it with images from target views.
These tracking techniques have also been applied to tracking articulated objects including humans(K53) and to motion segmentation and layering of video sequences. In the latter Bayesian statistical methods are used to separate the motion seen in a video sequence into two components and to interpret each of the regions within the image as belonging to one of those motions(K151,K152).
Work continues on building photorealistic models of architectural scenes from uncalibrated photographs. Novel algorithms have been developed to calibrate the cameras from information in the images alone(K118,K119) and to find correspondences automatically(K21,K32,K45,K154) and to recover the geometry(K22). A Windows 2000 application called PhotoBuilder(K136) has been developed which allows a non-expert user to interactively build a 3D model from arbitrary photographs. This can be downloaded from the group's web-site (http://svr-www.eng.cam.ac.uk/research/vision/). Preliminary results are extremely promising and a model of Jesus College has already been built.
Dr A.H. Gee
Dr R.W. Prager
Work on on-line learning and regularisation for non-stationary signal processing has reached a successful conclusion(K73,K74,K75). Further work on the development of unsupervised algorithms for independent component analysis has been completed(K78). Linear statistical techniques have been compared with neural algorithms as predictors of risk in liver transplant applications(K120).
Dr A.H. Gee
Dr R.W. Prager
The Stradx 3D diagnostic ultrasound system, already established as one of the best of its kind, is under continual development. New volume rendering facilities(K65) allow clinicians to obtain X-ray-like images of bony structures. This can clarify certain features and assist the diagnostic process.
Many clinical applications involve the measurement of the volume of structures within the body. Our pioneering work in this area has been extended to cope with large structures which do not fit entirely within a single ultrasound scan(K156,K157,K158). The algorithm used to interpolate surfaces through contours has been extended to interpolate intermediate surfaces in a morphing sequence(K155): this has applications in the medical imaging of moving objects, and also in computer animation in general.
A new graphical user interface (GUI) has been developed(K25,K26), to make 3D ultrasound more accessible to clinicians. Other GUI innovations(K161,K162) make the Stradx system better suited to telemedicine applications.
Initial experiments have shown how 3D ultrasound can be achieved without the need to attach an external position sensor to the probe: instead, the positions are deduced by analysing subtle changes in the images' speckle content(K128). Radial basis function algorithms have been evaluated for interpolating sparse 3D ultrasound data(K137). This work has led to a better understanding of image artefacts caused by different interpolation strategies.
A new EPSRC-funded project, to develop an extremely high resolution 3D ultrasound system and investigate its clinical utility, commenced in late 2000. The Stradx system continues to be updated and distributed freely on the web from http://svr-www.eng.cam.ac.uk/~rwp/stradx/. The software is in use in other international research laboratories.
Dr R.W. Prager
A study of the deformation of the foetal skull under the forces acting in the birth canal has recently been completed(K105).
Dr A.H. Gee
Collaboration with the Xerox Research Centre in Cambridge continues. A new Xerox-funded project, to develop strategies for storing and retrieving document images, will commence in 2001.
Professor A. Hopper
Dr J. Bates
Dr I.J. Wassell
"Sentient Computing" is a long-term research theme of the LCE. The Royal Society published the text of the Clifford Patterson Lecture given by Professor Hopper in the previous year which summarizes the research(K83). Location technology and location-aware systems and applications are key to this and a distributed location system has been developed and tested(K171).
Work on thin client architectures has spanned systems used for collaborative working(K58,K59) and also in situations where there is a control function in the network loop(K54,K55).
An EPSRC-funded project is developing a middleware platform that integrates the location and contextual information provided by sensors with multimedia data streaming mechanisms to assist the construction of collaborative applications in medical and other environments(K123).
Dr I.J. Wassell has been conducting a number of research projects in the area of Broadband Wireless Access systems. These include Combined Coding and Modulation(K112,K148), Orthogonal Frequency Division Multiplex (OFDM) systems(K84), space/time coding, interference cancellation and Dynamic Channel Allocation/Selection in fixed wireless systems.
Collaborative research projects in Broadband Wireless Access systems have continued with Adaptive Broadband Ltd (ABL) and have commenced with a new partner, namely Cambridge Broadband Ltd (CBL). Both companies are developing Broadband Wireless Access systems for wireless local loop (WLL) applications. In the WLL application point-to-multipoint configurations are employed with over-the-air data rates in excess of 25 Mbit/s. A result of the collaboration with ABL has been the development of fast training (low overhead), yet hardware efficient equalisers that are required to combat multi-path interference whilst keeping packet turnaround time low(K145,K146,K147).
In addition, the LCE in collaboration with AT&T Laboratories, Cambridge and ABL has deployed a Wireless Broadband demonstration/trial network in Cambridge(K113). This network permitted multimedia applications, networking issues and air interface issues to be investigated in a real system for the first time. A demonstration of the system was given at the Cambridge University Engineering Department 125 Anniversary Events held on the 13th and 14th July 2000.
Information Engineering References
[Table of Contents]
Last modified: September 2001