Robust Control Theory and Design
Robust Control Applications
Predictive Control and Indentification
Signal Processing and Communications
Medical Applications of Information Engineering
Computer Vision and Robotics
Artificial Intelligence and Decision Support
Professor K. Glover
Dr M.C. Smith
Dr G. Vinnicombe
The research activity in control system design methodologies continues to be productive with many extensions and practical applications.
The problem of finding low order H¥ Loop Shaping controllers has been studied in detail and a useful design method has been developed(J147) . This technique has been extended to decentralized and simultaneous control problems(J94). An effective method for dealing with actuator saturation within a robust control framework has also been developed(J95). The development of means of capturing and exploiting the inherent robustness of feedback systems, in order to reduce the apparent conservatism of current design techniques, has continued(J146).
Research effort has continued on input-output approaches to control systems. A generalisation of the gap metric for uncertainty studies in nonlinear feedback systems was developed(J48,J49). An approach to the computation of an induced norm for nonlinear systems was pursued(J123). Results on repetitive control were presented(J75). Research was carried out into passive and active vehicle suspensions. Theoretical work using a mechanical network approach investigated capabilities and limitations of various schemes from a fundamental point of view(J131,J132).
Sampled-data systems are being studied in detail and an effective design procedure was given(J11) for this class of system. The previous work on model order reduction of dynamic systems has been partially generalised to uncertain systems, parameter varying systems and certain nonlinear systems(J7,J14,J152).
Professor K. Glover
A substantial experimental activity has been initiated on the control of internal combustion engines to meet future emissions legislation. This work is sponsored by the Ford Motor Company and the EPSRC and is joint with the IC engines activity describe in Section A. Initially a fully instrumented engine test cell with a dynamic dynamometer is being commissioned for transient testing.
An aircraft model based on the HIRM (High Incidence Research Model) of DERA has been designed and built. This is to be used for the evaluation of novel control laws and has undergone initial wind-tunnel tests using a rapid prototyping environment(J105).
Controller designs for the HIRM were produced for the GARTEUR design challenge (see below) using H¥ loop shaping methods with success(J106,J107).
Professor K. Glover
Dr J.M. Maciejowski
The Control Group has participated in the `Flight Control Design Challenge' mounted by the Group for Aerospace Research and Technology in Europe (GARTEUR). The use of Model Predictive Control in flight control applications was investigated by means of this participation(J61,J62,J84). Software was developed for this investigation, which is also of use in generic investigations of predictive control(J60). The combination of predictive control with more conventional control techniques, in the context of flight control, has also been investigated(J108).
The use of model predictive control as a basis for a generic architecture for fault-tolerant control systems has been proposed, and is being investigated(J83) .
Work on the application of balanced realisations to system identification, which has been pursued for a number of years, has been brought to a conclusion(J13). Some of this work had been done within the European Research Network on System Identification; EU funding of this Network terminated in 1995, but further support under the EU's `TMR' programme has recently been obtained for a further four years (1998-2002).
Several contributions have been made to the theory of the recently discovered and currently popular `subspace' methods for system identification. Several methods have been found of guaranteeing that the models produced by these methods are stable(J15). This is particularly important in applications, such as adaptive control, in which the models are used without operator intervention. Exact conditions on input and output signals have been found, under which subspace algorithms give unique models, and a new 3-stage subspace algorithm has been proposed, which has better statistical properties than the standard algorithms(J16).
Work on the application of computer algebra to dynamic system approximation, which exploits the theory of polynomial ideals and leads to global optima being found, has been published(J54). The computational complexity of this approach is prohibitive for all but the smallest problems, but a more tractable approach, using the same underlying theory, is currently being investigated.
Dr R.J. Richards
As an example of utilising neural networks in applications the control of simple robotics tasks has been developed. An iterative controller capable of self modification during repetitive operation and thus having potential for wide applicability with unknown dynamics has been investigated. However, this requires some gain selection. Value iteration based controllers were evaluated and implemented on a small scale scara robot, the overhead crane optimisation problem, and a gantry robot minimising spillage in liquid transfer. The use of an iterative dynamic programming algorithm for time optimal movement of a flexible payload on a linear gantry robot was also reported(J88). During the above, analogies between the CMAC and iterative methods(J87) and methods of reinforcement learning were utilised(J86).
The problem of combined motion and force control has been reassessed, tracking reference trajectories in unconstrained directions and regulating force in constrained directions in the presence of model uncertainty and unknown stiffness at the arm/environment interface. A single controller with sliding mode control has been used for both domains coupled with an adaptive algorithm during the contact phase.
Dr. P.J.W. Rayner
Dr W.J. Fitzgerald
Dr S.J. Godsill
Dr N.G. Kingsbury
Dr J. Lasenby
Dr M.D. Macleod
Work has continued, in collaboration with Schlumberger Cambridge Research, to develop optimal methods for mud pulse telemetry in order to increase the bit rate for real time monitoring of `down-hole data' for oil well drilling applications. This has resulted in novel digital modem algorithms that are currently being implemented and tested in real systems(J140,J141).
Applications of time-frequency and wavelet methods have been applied to neurophysiological data applied to the correlation of neuronal firing and muscular activity. Significant improvements over short time Fourier methods have been obtained(J23).
Work is continuing with numerical Bayesian methods and sequential MCMC methods are being applied to several areas of signal processing, (mobile) communication channels in particular. (Professor M. West from the Department of Statistics and Decision Sciences, Duke University, visited the group for a month to collaborate on this work). Techniques have been developed for the segmentation of textured images(J3) and for the modelling and classification of oil paintings from the structure of the cracks in the paint(J144,J145).
A new approach to dendrochronology has been developed and applied, with good results, to the dating of wood from violins(J57).
Bayesian methods have been applied to the problems of cyclostationarity and this work is being supported by the Government Communications Headquarters (GCHQ).
Model selection methods continue to be investigated(J113) and methods based on predictive densities have been shown to yield good results(J39,J133,J134,J135,J156). New techniques have been developed for incorporating prior parameter information in the general linear model(J56). Final results from a project on the self-structuring of artificial neural networks have been published(J47).
New methods for dealing with sea clutter in radar systems have been investigated in collaboration with DERA (Malvern) and results have been published(J104).
The methods developed for changepoint detection are being applied to many areas and the problem of ion channel changepoint detection has shown some interesting results(J38,J136).
Work continues on the application of statistical image processing techniques to the restoration of degraded manuscripts and similar techniques have been applied to the identification of malignant cells using fluorescent microscopy(J137).
This year has seen new developments of work in processing and modelling of non-Gaussian time series, with application mostly to acoustical signals such as speech and music(J50,J51,J53). In this work a methodology has been devised for estimation of signals corrupted by heavy-tailed noise distributions such as the Student-t or alpha-stable distributions, which are of physical significance in many real-world noise environments. Markov chain Monte Carlo (MCMC) simulation methods have been used for carrying out sophisticated Bayesian computations within highly non-Gaussian environments. Further work has looked at the problem of model uncertainty in non-linear time series model selection, using novel indicator variable methods to determine which terms in a highly over-parameterised complex model are relevant to the observed data(J142). This year has also seen the start of industrial collaborations with Olivetti, who will fund an extensive two year project in source separation techniques for the multimedia office, and with Sibelius Software Ltd who will fund work in automatic transcription of musical signals.
A research programme in the general area of non-linear and non-stationary signal processing has been instigated and progress has been made with problems concerning non-Gaussian noise and how to develop optimal signal processing algorithms for such environments(J68,J69,J70,J97). Some initial results on time-varying autoregressive models have been published(J112,J114). Preparations continue for the Isaac Newton Institute programme on Nonlinear and Nonstationary Signal Processing.
The research project on the restoration of degraded image sequences has been extended by the European Union. Nonlinear techniques have been developed for interpolation of missing data in images(J2) and methods developed for the registration of line scan in jittered video sequences(J67). Bayesian methodology has been applied to a number of problems in robust detection, estimation and interpolation of degraded image sequence data(J52,J66).
An invited paper was given to open the 1st. European Conference on Signal Analysis and Prediction(J115).
Four members of the Signal Processing and Communications Group received La Nomination d'Honneur 1997 in a competition organised by Fondation ALTRAN pour l'Innovation on the "Preservation of the Humanity Memory". From the 250 participants only 7 received the award which was presented by President of the French Senate. The work submitted for the competition was based on the research in the restoration of degraded manuscripts, film and audio using signal and image processing techniques. Abstracts of the submissions will be published in 1998.
Projects on methods of modifying standard image and video compression methods, so that they are suitable for use over noisy and error-prone transmission channels, have been completed. The basic error resilient entropy code (EREC) algorithm has been successfully applied to MPEG-II standard video bit-streams(J116,J138) to enhance the resilience to transmission errors while requiring no increase in transmitted data rate or bandwidth. Work is continuing(J65) on applying the EREC algorithm to more advanced compression schemes using wavelets.
Two other current projects are investigating a new form of complex wavelet transform which provides an efficient and versatile tool for image analysis and synthesis in areas such as noise reduction, image enhancement, texture processing and image registration.
Motion estimation for video signals continues to be a key feature of several projects. In particular, wavelet techniques have been extended to use filter banks with complex coefficients so that object motion can be measured to high accuracy within a multi resolution framework(J64). Enhancements to these methods have been developed to provide greater accuracy for colour image sequences(J85). Two other projects on motion are studying (a) the problem of optimal object segmentation methods based on motion throughout a sequence of many frames(J37), and (b) methods of improved motion compensation based on adaptive triangular tessellation of image frames(J9). By combining these techniques into a composite motion estimation and compensation model, it is hoped to provide significantly improved image quality for low bit-rate compression applications.
In the area of extracting information from multiple camera views, progress has been made on completely characterising the structure of the tensors relating matching points in three views(J5,J72) and in deriving new invariants over multiple views using these tensors(J6,J72) . The Royal Society Discussion Meeting on Geometry in Computer Vision, co-organized by Cambridge University Engineering Department, Oxford and Sussex took place in July 1997 and was a great success, attracting around 200 participants on each of the two days. As a front-end for feature extraction, considerable progress has been made in the area of image segmentation. An efficient segmentation scheme has been developed which is based on the classical creep and merge techniques but overcomes many of the problems traditionally associated with such methods(J4). Work into the use of geometric algebra as a tool for encoding structure into matrices and tensors has continued(J73,J74) and geometric algebra has also been used as a means of teaching the fairly advanced concepts in the engineering mathematics involved in some areas of mechanics(G24).
Further results concerning efficient digital filter implementation have been published(J35,J36). A short project was completed concerning modelling of the signals from the accelerometer signals used to activate airbags(J91).
The research in the general area of wireless communications continues and a new variable-rate technique for frequency-hop CDMA has been reported(J89). Further results have been obtained on multi-media transmission in fibre-optic LANs(J90).
Professor S.J. Young
Dr A.J. Robinson
Mr P.C. Woodland
As in previous years, the main focus of the work in Speech has been on recognition. Previously the emphasis has been on dictated or read speech, however, current research is aimed at dealing with the problems of transcribing so-called "found speech" which is speech produced without any prior intention by the speaker to be transcribed. Examples of this are broadcast news programmes, court proceedings, debates, conversations, etc.
Since 1992, the Speech Group has been associated with the ARPA programme and it has continued to take part in its annual evaluations. In 1996, the ARPA evaluation was concerned with transcribing broadcast television and radio news programmes. The CUED group submitted two systems: the HTK HMM-based system and the ABBOT neural network hybrid system. The HTK system had the lowest error rate on the main focus task and the second lowest overall(J153,J154). The ABBOT system also fared well coming sixth overall.
Several new funded research projects began in the last year. A new collaborative EPSRC project between the Computer Laboratory and the Engineering Department began on the topic of Multimedia Document Retrieval. This project will build on the earlier Video Mail Retrieval project(J10,J40,J158). It aims to integrate large vocabulary transcription of video and audio documents with information retrieval techniques in order to provide content-based retrieval of audio and video databases.
In addition, new projects began on transcribing conversational speech (funded by GCHQ) and on transcribing speech corrupted by noise (funded by Siemens).
PhD research continues in the areas of articulatory modelling(J8); interactive spoken language education(J150,J151); pronunciation modelling and adaptation across speaker accents(J58,J59); class-based statistical language modelling(J98,J99,J100) and grammatical inference(J130).
Other active research is investigating the development of improved acoustic modelling techniques(J41,J44,J103); noise compensation(J43); discriminative training of HMMs(J143) and multilingual speech recognition systems(J157).
One particular feature of the HTK HMM-based speech recognition system is the use of the maximum likelihood linear regression (MLLR) technique for adaptation of HMM parameters to both the current speaker and acoustic environment. The original version of this technique was developed in the Department in 1994 and is and has now become a standard component of many state-of-the-art speech recognition systems worldwide. The MLLR technique has continued to be developed(J42,J45,J46,J155) and combined with other adaptation techniques(J111). Alternative techniques for rapid adaptation of large HMM systems have also been pursued(J1).
Work continues on the ABBOT speech recognition system. This constituted research into the optimal combination of connectionist acoustic models(J26,J27,J148,J149) and long term language modelling(J22). The system participated in the ARPA evaluations of very large vocabulary speech recognition(J63) and specifically the recognition of broadcast material(J24,J25).
Speech enhancement work was carried out using autoregressive hidden Markov models for small vocabulary tasks(J77,J78,J79) which is now being adapted to large vocabulary tasks.
One very low bit rate (1200-2400 kbps) speech coding project has been completed and the result will soon be commercially distributed(J129). Research continues into speech coding at even lower bit rates (600-1200 bps) using segmental techniques(J163,J164,J165).
Dr M. Niranjan
Work on Neural Networks applied to nonstationary signal processing problems continued in two areas(J55,J96,J117) and steady progress has been made on an EPSRC funded project to monitor liver transplant patients(J92,J93).
Further, work has advanced on modelling financial data especially addressing the problem of approximate methods for options pricing(J101,J102).
Dr M. Niranjan was involved as co-organiser of a six month programme on Machine Learning and Neural Networks at the Isaac Newton Institute for Mathematical Sciences (July-December 97), and was Chairman of an IEE International Conference on Neural Networks (July 1997).
Dr RW Prager
Dr M Niranjan
The European Community funded research into ways of using neural networks and other statistical techniques to predict risk in pregnancy(J80,J81,J82). The EuroPunch project was recently completed and the QAMC project is due to finish at the end of 1997. The Department's main partner in these projects is the University of Cambridge Department of Obstetrics and Gynaecology.
The study of the response of liver transplant patients to immunosuppressive drugs continues through the project, "Neural Computing for Nonstationary Medical Signal Processing". It is funded by the EPSRC and involves collaboration with the University of Cambridge Department of Clinical Biochemistry and the Transplant Unit at Papworth Hospital(J92,J93).
Dr RW Prager
Dr AH Gee
Research continues on various aspects of 3D diagnostic ultrasound. New techniques have been developed for the accurate acquisition of free-hand 3D ultrasound data(J109,J110). Further research has shown how the quality of the 3D data can be improved by registration and spatial compounding(J118,J119,J120,J121,J122,J139).
Many clinical applications require accurate segmentation of structures within the 3D ultrasound data set. With this in mind, new algorithms have been developed for geometrical(J31,J33,J34,J76) and statistical(J128) segmentation.
The segmented data can then be reconstructed to facilitate 3D visualisation and accurate volume measurement(J30,J32,J71).
Work continues on the SOLUS-3D project, focused on the development of 3D diagnostic ultrasound systems for Obstetrics and Gynaecology. This project is a collaborative venture between the Cambridge Departments of Engineering, Radiology, and Obstetrics & Gynaecology, as well as other medical and engineering sites across Europe. It is funded by the European Community.
Dr R Cipolla
Dr AH Gee
Research in computer vision and robotics has continued on the development of new theories for recovering the three-dimensional shape of visible surfaces from images taken from arbitrary viewpoints with uncalibrated cameras.
Novel contributions have been made in the analysis of curved surfaces (where the dominant image feature is the silhouette or profile). 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. We have also been able to show that viewer motion can be recovered from the envelope of consecutive contour generators - the fronter(J21,J126). These theories have been successfully implemented, thus allowing the registration of images taken from different viewpoints and the reconstruction of arbitrarily complex shapes. The special case of singular profiles(J18) (often exploited by artists in sketches of curved surfaces) has now been solved and awaits implementation.
Progress continues in developing new techniques to match image curves over arbitrary viewpoints for extracting symmetry axes and recognition. Novel geometric invariants under affine and projective transformation groups were developed which to do not suffer from the problems of noise sensitivity of differential invariants and the occlusion problems of global moment invariants(J124). A quasi-invariant parameterisation of image curves which approximates group invariant arc-length with lower spatial derivatives has been shown to be extremely powerful in detecting symmetry in natural and textured images and efficiently recognising arbitrarily complex curves under arbitrary viewpoints and occlusion.
Visual motion, as perceived by a camera mounted on a robot moving relative to a scene, can be used to compute the robot motion and the structure of the scene and aid navigation. Simple qualitative cues such as time to contact and relative surface orientation could be reliably extracted from the image divergence estimated from the temporal evolution of the apparent area of a closed contour(J17). The method was extended to image texture, abundant in natural images and used to guide a robot manipulator in welding and motion planning tasks(J28,J29). We are now working on techniques to automatically segment images based on properties of the texture(J4).
Stereo vision is often used to recover 3D position but requires correspondence of image features and calibration of position of cameras. Our research has aimed to find simpler, calibration-free cues to surface position and geometry. A simple but more robust approximation to stereo has been proposed exploiting cues present in orthographic projection only(J12,J19,J20). A hand-eye coordination system has been developed to guide a robot manipulator to pick up unfamiliar objects in unstructured environments with uncalibrated stereo cameras. By detecting and tracking a human hand it is possible to allow the user to point at an object of interest and guide a robotic manipulator to pick it up(J19).
Work continues on the design of novel man-machine interfaces. Pointing and face gestures can provide more natural ways of communicating with computers and machines. Automatic techniques for localising faces(J159,J160,J161) and a realtime visual tracking algorithm(J12,J125) are being exploited in hands-off computer and robot interfaces for physically handicapped people as well as for an interface to 3D television displays. Projects in rehabilitation engineering have led to the design of an interactive robotic visual inspection system which is undergoing extensive user-trials(C21,C23).
Collaboration continues with the Olivetti and Oracle Research laboratories, Panasonic, Hewlett Packard, Toshiba and Xerox. Collaboration with the Xerox Research Centre in Cambridge has led to a new project on document image processing. The aim is to develop techniques to scan documents using over-the-desk CCD cameras. A paper describing preliminary results(J162) was awarded the Industry Prize at the 1997 British Machine Vision Conference. The prize is awarded for the best paper describing work with significant industrial potential. The Science Prize was awarded for the reconstruction of surfaces from uncalibrated views(J126).
Dr T. Holden
Work is progressing in applying Artificial Intelligence and Decision Support techniques to large-scale industrial systems, in particular to energy and safety-critical situations. Work with BP and Honeywell has led to the development of information and management structures for large, complex information systems for improving performance and reducing risk through better informed operators.
One outcome of this work has been the installation and full-time operational use of a decision-support system installed in a BP operational plant over the last year. By the use of neural network techniques, novel data-mining techniques have been developed to extract and make beneficial use of the large accumulation of knowledge built up over this time.
Related research work has concerned knowledge-sharing in a corporate environment. A key concern in safety-critical or otherwise risk-prevalent organisations is the management of the human, communication, informational and knowledge factors that influence the quality of decisions. One aspect of this has been the development of methods for supporting technology and skills transfer between a fully developed plant in the UK and a newly-installed one in Malaysia. Particular insight was gained into the how to transfer knowledge across cultures.
[Information Engineering References] [Table of Contents]