Department of Engineering - Annual Report 1997/98

Information Engineering

CONTROL SYSTEMS
Robust Control Theory and Design
Predictive Control and Identification
Robust Control Applications
Control Robotics

SIGNAL PROCESSING

SPEECH, VISION and ROBOTICS
Speech Processing
Neural Networks
Medical Applications of Information Engineering
3D Ultrasound
Computer Vision and Robotics

LABORATORY for COMMUNICATIONS ENGINEERING

References

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 innovations in devices. The information considered varies from measurements in a flight control system, speech input to machines, vision and image data, and data in ubiquitous communication systems. A significant development this year has been the establishment of the Laboratory for Communications Engineering within the Division.

Control Systems

Research in Control Systems has continued across a broad range of areas from applications in the automotive and aerospace industries to theoretical and methodological developments in control system analysis, modelling and design.

Robust Control Theory and Design

The effectiveness of the H-infinity Loop Shaping method for the design of robust controllers is increasingly being recognised and further developments have been reported(K110).

Research effort has continued on input-output approaches to control systems. A generalisation of the gap metric for uncertainty studies in non-linear feedback systems was further developed(K41). A study of the robustness properties of sliding mode control schemes was carried out(K39,K40). Theoretical results on repetitive control, interpreted in an H-infinity framework, were presented(K85) and experimental results of this work applied to the control of the radial loop in a CD player were presented(K84).

Distance measures which aim to capture and exploit the inherent robustness of feedback systems have been developed further, and preliminary results have been obtained on generalisations of the nu-gap metric which are suitable for the study of nonlinear systems.

Work has continued on means of addressing actuator saturation within a robust control framework, and preliminary results obtained for unstable systems, when global closed-loop stability cannot be guaranteed but local robustness is achievable.

The study of sampled-data systems has continued, with the development of a numerically robust algorithm for H-infinity synthesis(K15) and a frequency-domain analysis tool for design(K14). System representations that are notionally equivalent to coprime factors were shown to exist for linear, periodically time-varying, continuous-time systems(K16).

A class of nonlinear dynamic systems which include a repeated nonlinearity (such as neural networks) has been analysed and results on sufficient conditions for stability and performance have been presented(K21). Distance measures which aim to capture and exploit the inherent robustness of feedback systems have been developed further, and preliminary results have been obtained on generalisations of the nu-gap metric which are suitable for the study of nonlinear systems. Work has continued on means of addressing actuator saturation within a robust control framework, and preliminary results obtained for unstable systems, when global closed-loop stability cannot be guaranteed but local robustness is achievable.

The synthesis of controllers of reduced dynamic order is a challenging problem and an effective approach has been presented(K42) using results from H-infinity theory.

Predictive Control and Identification

Work on the application of Model Predictive Control to high-bandwidth applications, in particular to flight control, has continued, with particular emphasis on providing fault-tolerant control systems(K59,K60,K61,K62,K91). Some of this work was sponsored by DERA. A course of lectures was delivered at Delft University of Technology and has been published(K92).

Work on new variants of `subspace methods' of system identification has continued. A 3-block algorithm has been proposed, which gives unbiased results for a wider range of conditions than comparable previously proposed algorithms(K22).

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 continued, and a much more efficient approach has been found than the one previously developed by us(K55). This work has been done in collaboration with the Free University of Amsterdam and with Delft University of Technology.

New techniques for the identification of models suitable for control systems design have been developed. A particularly promising approach to this problem aims to ensure that the identified model is close to the true system as measured by the nu-gap distance.

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 EPSRC and the Ford Motor Company and is a multidisciplinary project with activity as reported in Section A and with the Department of Chemistry. A fully instrumented dynamic dynamometer with rapid prototyping facilities is now fully operational and supporting this work.

A new programme of research on flight control for Short Take-Off/Vertical Landing aircraft has been initiated with sponsorship from DERA, Bedford.

Combustion oscillations that can occur in gas turbines are reported in Section A and initial work on the stabilisation of such oscillations has been reported(K20).

Research continued in the general area of robotics and control, work being completed on the project of Iterative Robot Control. A framework for iterative controller analysis and reduction in design complexity was developed. Value-iteration based controllers were implemented on low order, time optimal small-scale robotic tasks. A novel robust trajectory/force controller based on a combination of sliding mode and adaptive control techniques has been developed and demonstrated. The controller is constructed to track reference trajectories in constrained directions and regulate force in constrained directions within a single controller. A broader consideration of analytical aspects of performance weights in H-infinity frameworks has followed.

Signal Processing

This year a 6 month programme on Non-linear and Non-Stationary Signal Processing was held and organised by Dr Fitzgerald, at the Isaac Newton Institute for Mathematical Sciences between July and December. The classical theory of signal processing is based on models which are stationary, linear and in many cases also Gaussian. Recent advances in time series and the theory of signal processing have drawn attention to many new models and methods. There has been a parallel growth in the applications of signal processing in many modern areas of modern engineering and into other areas such as financial time series, the environmental sciences, physiology, etc. In many instances, methods have developed in an ad hoc way, being designed for specific applications rather than fitted into a general framework. The purpose of this programme was to bring together statisticians, engineers and other researchers who use signal processing methodology to unify existing methods and to identify areas of research where new methodology is required. This programme involved many members of the group, both giving talks and interacting with the international invitees present at the Newton Institute.

Research has continued with Bayesian approaches to Cyclo-stationarity(K99). State space Bayesian methods have also been applied to radar issues(K107,K108,K109). Results have been published on a class of Bayesian parameter estimation problems which can be solved analytically using a Volterra series representation(K57).

There exist many techniques for optimal signal processing in a stationary Gaussian noise environment but many real-world situations are non-Gaussian. A class of probability density functions which allow modelling of a wide range of noise processes is the alpha-stable distribution. Techniques have been developed for near-optimal filtering and detection of signals in impulsive noise modelled by such a process(K70,K71,K72,K73).

The past 10 years of research work in digital audio restoration methods has come to fruition with the publication in September 1998 of a Springer Verlag research text(K45) and a chapter in a Kluwer book on digital audio(K47). Ongoing research in these areas has studied MCMC methods for noise reduction(K5,K44,K46,K47), polyphonic musical pitch extraction(K150), non-linear signal modelling and model uncertainty(K148,K146) and digital video restoration(K69).

New projects have been initiated with Olivetti-Oracle Research Laboratory, Cambridge on Signal Source Separation Methods, with the EPSRC on Non-linear Audio Signal Modelling and on Sequential Simulation Methods. This year has seen the acceptance of simulation-based Bayesian methods for signal processing, and this has been heralded by tutorial/special sessions at ICASSP 98, EUSIPCO 98 and HOS 99, all of which were jointly organised and participated in by members of the group. In addition, an invited talk was given at the prestigious Valencia meeting on Bayesian statistics(K36).

The results of a major research effort in the area of motion picture restoration has been published in a research text by Springer Verlag(K67) and the culmination of a EU funded joint research programme on motion picture restoration was the public demonstration of a real-time digital video restoration system at EXPO98 in Lisbon.

The segmentation of images into constituent objects is an important problem in many potential applications of image processing and Bayesian methods have been developed which are based on Markov random field descriptions of object texture(K4). Methods developed for the reconstruction of missing areas in monochrome images have been extended to colour images(K2).

Work on error-resilient compression of images and video sequences is continuing. The basic aims of the work have been the development of methods to code images which are virtually as efficient as conventional methods, but which do not suffer from the problem of severe error propagation, which affects most standard compression methods(K117). In particular a project to trans-code MPEG-2 video signals into error resilient form has been completed and been found to work well over simulated noisy mobile radio links and congested ATM networks(K142,K143). The project is now continuing with the development of error-resilient coding for the latest progressive wavelet compression methods which use embedded zero trees and set partitioning into hierarchical trees.

The final results of a project to investigate methods of measuring the visibility of coding distortions in video sequences have been reported(K154). This project built on results from an earlier project which attempted to model the sensitivity of the human visual system to various types of coding distortion for still images. The new work has extended the reaction-time measurement method to video sequences and produced some interesting new parameters for temporal characteristics of the human visual system. Three projects on motion estimation and modelling for video sequences have just been completed. The first(K34) has developed techniques of recovering the layer representation of a moving scene, so that all objects in a scene can be allocated to a small number of layers, each of which is accurately modelled by a single set of affine motion parameters. The Viterbi algorithm is used to achieve near-optimal performance. The second motion project(K11,K12) has developed affine motion models using Delaunay triangulation of frames within a sequence. These models are shown to give useful reductions of coding artefacts when compared to more conventional block-based translational motion models. Fast search algorithms have been developed for motion estimation with these models. The final motion project(K97) uses complex wavelets to perform an efficient hierarchical estimate of motion between consecutive frames in a video sequence, such that a variable resolution motion field can be generated, dependent on the level of detail in different parts of the frames. This work has also been applied to coloured frames(K98).

To extend this work on complex wavelets to a wider range of image processing applications, we have recently developed the dual tree complex wavelet transform (DT CWT). This provides a perfectly invertible transform, which has many of the desirable properties of transforms that are only approximately invertible, such as the Gabor transform, but avoids their disadvantages. The desirable properties of the DT CWT include shift invariance, good directional selectivity, and efficiency of computation. Applications of the DT CWT include image restoration and enhancement(K64) and texture analysis and synthesis(K65). A number of other application areas of the new transform, which are being pursued currently, include video compression and image watermarking.

Work on motion capture and analysis has continued. A motion capture unit consisting of two cameras, soon to be updated to three, is now fully operational and has been used to capture data from swimmers, runners and musicians. From this data, 3D animated models of the motion are recovered. Emphasis has been placed on methods of calibration and reconstruction which are robust in the presence of noise. Currently, a Bayesian tracking scheme is under development. Research is also being carried out into the analysis of the 3D data - geometric algebra is being used for dealing with the rotational aspects of limb movement. The application of geometric algebra techniques to signal analysis and computer vision has also continued(K74,K75,K76,K77,K78).

Further results on efficient digital filter implementation using Multiplier Blocks have been published(K29,K30). Research into other DSP applications of minimal-complexity multipliers continues.

Further results from the EC Weigh in Motion project have been published(K139). This project involved optimal estimation of vehicle static weight using sparse noisy data from a small spatial array of pressure sensors.

A new method for estimation of resolved multiple sinusoidal tones has been published in a paper(K94), and its extension to high resolution case at the European Signal Processing conference(K93). Work continues on both fundamental techniques and their application to musical audio analysis.

A new edition of a telecommunications reference book has been published(K95). Results on adaptive rate channel coding have been reported(K79,K80,K81,K82,K83).

Speech, Vision and Robotics

Speech Processing

Research in speech processing covers the areas of recognition, synthesis, coding, language modelling, information retrieval and dialogue design. Work in the group is organised around externally funded research projects and individual PhD projects.

The group's association with the US NIST co-ordinated speech recognition evaluation programmes has continued and in the last year, CUED systems have continued to define state-of-the-art performance giving the lowest word error rates in two separate evaluations.

The EPSRC sponsored Multimedia Document Retrieval project is now well under way. This project is focused on the problem of transcribing video and audio material in order to index it and subsequently retrieve wanted articles by content. The development of the core technology required to do this has made good progress. The work in the transcription area uses the HTK Broadcast News Transcription System. Major progress has been made here on the problems of automatically segmenting a continuous audio stream containing different types of audio, speaker clustering and unsupervised speaker adaptation and class-based language modelling(K54,K104,K153). This system gave the lowest overall word error rate in the 1997 DARPA/NIST Hub4 Broadcast News transcription evaluation by a statistically significant margin(K152) in competition with systems from the top academic and commercial speech recognition research laboratories world-wide. On the content-based retrieval side (performed in collaboration with the Computer Laboratory), a basic retrieval engine was developed. The automatic transcription and retrieval components were integrated into a system that was included in the 1998 TREC-7 spoken document retrieval conference which required the retrieval of relevant broadcast news stories in response to natural language queries.

The EC sponsored SPRACH project has focused on the further development of the ABBOT recognition system. The connectionist framework employed allowed for work on the incorporation of acoustic confidences and the automatic learning of pronunciations. A system was produced for the French and Portuguese languages as well as continuing the development of the North American system for the Hub4 and TREC evaluations(K27,K28).

A Hewlett-Packard sponsored project exploited the speed and potential low memory requirements of the ABBOT system, concentrating on fast keyword spotting, low-memory large vocabulary decoding and compact language models(K119).

The GCHQ sponsored project on the Automatic Recognition of Conversational Telephone speech has continued with good progress being made. Significant improvements were made to the front-end signal processing, the development of robust speaker and channel normalisation procedures and the development of improved language modelling(K152). These factors taken together with the overall high performance of the HTK-based recognition engine led to a transcription system that gave the lowest overall error rate on the 1998 Hub5 Conversational Telephone transcription task.

Work on speech recognition in noise proceeded with sponsorship from Siemens. This included a study in which model compensation techniques were compared to environment adaptation techniques, the latter being shown to be superior.

A paper describing work on speech recognition using discriminative training techniques in speech(K149) resulting from an earlier EPSRC project was given the EURASIP Best Paper of the Year Award for 1997. The award is based on a comparison of all papers published in the leading journal Speech Communication.

Fundamental research on speech recognition has also included work on pronunciation modelling(K106); automatic accent modelling and adaptation(K58); multi-lingual language modelling and speaker normalisation. Other project and PhD work includes studies on the use of speech recognition for interactive language learning.

A low bit rate speech coder previously developed with Hewlett-Packard has been adopted by Microsoft for the portable computing environment. Further work on speech coding has resulted in a segmental based approach that has been shown to operate well at very low bit rates (350-650bps).

Neural Networks

Work on on-line learning and regularisation for nonstationary signal processing has made significant progress(K41,K42,K49,K52,K53).

A research project, called QAMC, exploring statistical techniques for predicting risk in pregnancy has recently been completed. This was funded by the European Community. It has resulted in improved techniques for classifier feature selection(K87) and the construction of a number of prototype risk prediction systems(K88,K89,K90).

Research continues on various aspects of 3D diagnostic ultrasound. New techniques have been developed for all stages of the process, including calibration of the acquisition system(K114), reconstruction of the 3D data(K122) and visualisation of the acquired data(K112,K113). Further research has shown how the quality of the 3D data can be improved by registration and spatial compounding(K120,K121).

Many clinical applications require accurate segmentation of structures within the 3D ultrasound data set. With this in mind, a new algorithm has been developed for statistical segmentation(K135). The segmented data can then be reconstructed to facilitate accurate volume measurement(K144,K145). The results of much of this research have been incorporated in the Stradx freehand 3D ultrasound system which is available on the web, free of charge, from http://svr-www.eng.cam.ac.uk/~rwp/stradx.

Work continues on the SOLUS-3D project(K17,K18), 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.

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(K78).

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. It has now been shown that viewer motion can also be recovered from the envelope of consecutive contour generators. These theories have been successfully implemented, thus allowing the registration of images taken from different viewpoints and the reconstruction of arbitrarily complex shapes(K23,K102,K128).

Progress continues in developing new techniques to match features(K5,K19,K136) and image curves over arbitrary viewpoints for extracting symmetry axes and recognition. Novel geometric invariants under affine and projective transformation groups were developed which do not suffer from the problems of noise sensitivity of differential invariants and the occlusion problems of global moment invariants. 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(K129,K130).

Work continues on the design of novel man-machine interfaces(K25). Pointing and face gestures can provide more natural ways of communicating with computers and machines(K24,K40). Automatic techniques for localising faces(K156) and a realtime visual tracking algorithm are being exploited in hands-off computer and robot interfaces for physically handicapped people as well as for an interface to 3D television displays. Collaboration with the Xerox Research Centre in Cambridge continues, progress has been made in the area of document image processing, including the automatic processing of handwritten annotations(K140).

Projects in rehabilitation engineering have led to the design of an interactive robotic visual inspection system which is undergoing extensive user-trials(K33).

Two new research projects were commenced in 1998. The first new project involves building photo-realistic models of architectural scenes from uncalibrated photographs. Preliminary results are extremely promising and a model of Jesus College has already been built. In the second, a prototype of a realtime robot-visual servoing system has been built which allows a robot to position itself relative to any target in an unstructured scene using visual information alone.

Laboratory for Communications Engineering

The Laboratory for Communications Engineering is a new research group in the Information Engineering Division. The group has a wide range of research interests under the generic banner of `Communications', and has close collaboration with several local companies including the Olivetti and Oracle Research Laboratory and Adaptive Broadband Limited. The group's interests range from the development of hardware to support novel networks through to the development of novel `middleware' software for LANs and WANs(K123,K125,K126).

Work in the area of Middleware support for distributed and mobile systems has continued. In particular the research area of mobile objects has reached maturity(K3). This work attempts to support the fact that mobile users require their applications to be mobile also. Rather than carrying around the equipment they need, the computing equipment deployed in users' current locations detects their presence and personalises itself to them, including stateful migration and reconfiguration of their applications environment.

Work in the area of distributed event-based systems has continued(K9). The revolutionary approach is based around registration of interest with active distributed services. When events of interest to a client occur, the client is notified through receipt of a parameterised event instance. The remote filtering mechanism prevents the network being flooded with unnecessary traffic. The asynchronous notification is more suitable for modern active applications (e.g. user mobility, transport systems and cooperative working) than current de facto approaches such as CORBA. Recent work has tackled the problem of making event services scalable on a world-wide scale and of making heterogeneous systems work together. A paper on using events for integrating real-world and computer-supported cooperative working was awarded best paper of conference (composed of multiple concurrent workshops) in the prestigious IEEE WETICE conference(K9).

As a development of the event work, a new service has been identified and constructed: the event store(K137,K138). This service acts as a temporally-aware repository for event instances. Event sequences can be replayed as they happened and sequences can be queried using temporal and content operators. Example applications are memory prosthesis ("show me the document Andy showed at the meeting sometime last week"), mobility support ("replay me all the conference events I missed whilst disconnected from the conference"), and visualisation ("replay me the evacuation drill for the oil rig in virtual reality").

In a slightly different area, collaborative research has investigated a systems architecture for the transparent parallelisation of declarative database queries(K1). Using the technique developed, languages such as SQL can be transformed to a parallel execution in real-time. This technique is based on sound theoretical principles and is suitable for efficient implementation on general-purpose parallel machines. A fully-functional implementation on a 128 node Fujitsu parallel machine has been produced. Initial performance figures are promising.

In addition, Dr. Bates has been asked to contribute to two books on distributed multimedia systems and the Internet(K6,K7).

Work on Radio ATM/Broadband communications has been done as a collaborative research project with Adaptive Broadband Ltd. (ABL), developing broadband radio ATM systems for wireless LAN's and wireless local loop (WLL). Architectures and protocols to allow seamless handover have been developed and implemented on a 10Mbps-prototype wireless ATM LAN(K103). Fast and efficient equalizers have been developed to combat multipath interference whilst keeping packet turnaround time low(K133). Extensive microwave radio propagation measurements have been conducted throughout the Cambridge residential Area in order to characterise path loss for WLL systems. Theoretical propagation models have been developed and compared with the field measurements.

Work on thin client systems has been an interest of Professor A. Hopper for many years. A `thin client' is a node on a network (or node in a system) which has little (or preferably no) computing power. Work recently has focused on extending the ORL Virtual Network Computer to provide support for asynchronous collaboration over the Internet(K80), and also on the development of protocols for the human control of devices using thin clients. A new project has been started which is looking at the application of networking to the home, and the software infrastructure to support the use of thin-client home appliances. It is also looking at the development of novel networks to work inside devices, such as fridges, and looking at how providing network capability to unusual home appliances can provide more functionality to the system. The Fridge Net is a passive multi-tag based system using bi-directional radio communication rather than the traditional approaches and allows sensors to be read/written on to the tag.

The LCE entered the 2nd International RoboCup Football competition (RoboCup'98) in Paris(K124). The aim is to develop an autonomous system of five small robots that can play football on a table tennis table. All parts of the system were developed by the group, including the highly specialised goal keeper. This year, 12 teams entered the competition and the LCE Team finished 4th overall, after being narrowly beaten in the semi-finals by the eventual champions. The work included the extension of Terrain Maps to provide Time encoding, to allow for good planning in the dynamic and hostile environment of the pitch.

Professor A. Hopper has a long-term interest in the development and deployment of novel location systems. Previous work has included the Active Badge System. Recent work by members of the LCE has looked at the development of a 3D Indoor Location System(K151). The system is able to locate an object in a room to within about 5 cm accuracy. This is achieved by using a sensor attached to the object being tracked. Laboratory wide installation of the system will be complete by the middle of 1999.

Several members of the LCE have been working in conjunction with ORL to investigate the use of low-power ad-hoc networking(K10,K56). Members of the LCE have been evaluating the use asynchronous processors in such networks as well as the use of novel routing algorithms in such networks to minimise power usage by the nodes(K118).

During the last part of 1998 a new project has been started to investigate the development of technologies to support alternative communication paths to, and around the human body. Traditional approaches are to use either radio communication or the capacitance properties of the body.

In this project we are looking at the development of carpet tile that can allow high-speed communication with a shoe (when it is placed on the carpet tile). The carpet tile therefore provides communication to the body. The project is also looking at how the communications channel can be routed around the body to devices attached to the body. It is currently working on the use of a number of specialised conductive cloths. This provides what we are calling "wired wireless".

Although the project is in its infancy, the initial results are very promising and it is hoped to have a working prototype system by the end of 1999.

Information Engineering References

[Table of Contents]

Last modified: October 1999