People: Alan Wilson, Jim Usherwood, Kyle Roskilly, Steve Amos

Project overview – new tools for a new area of research

flock of pigeons

Measuring the detailed movement and relative location of individual animals within groups has, up to now, not been possible in most situations. The CARDyAL project has been designed to open a new field of research in this area, and in so doing to develop tools and methods that can be used in many other applications.

The aim is to do this by:

  • designing new lightweight and minimally obtrusive devices (tags) to record detailed data on the location and activity of free ranging animals and birds
  • developing the new analytical methods necessary to make optimal use of the data collected
  • testing and validating these in a number of relevant real-world situations.

Animals move in groups for a number of different purposes such as to avoid predation, to facilitate hunting, to conserve energy during ranging and for navigational benefits. Being part of a group may have costs as well as benefits to the individual. The ability to measure the detailed movement dynamics of individuals within a group has profound implications for a number of different fields including wild animal conservation, food animal management, management of facilities and the design of buildings. In addition to this, an understanding of what happens in natural groupings will inform the development of new ways to improve flow and reduce energy consumption in human environments such as workplaces and public facilities.

Finally, the incorporation of empirical data from natural systems into the computer models used to design complex systems, computer graphics etc (see Particle Swarm Optimisation in the Technology and Analysis Page) will allow naturally evolved group relationships and behaviour to inform the design of artificial systems.

This work has implications for disease control, animal conservation, food animal management and a number of other fields.

Collaboration

CARDyAL is a major four-year project funded by the EPSRC (Engineering and Physical Sciences Research Council).

It is a collaborative project between two research groups: University College London Computer Science Department (Stephen Hailes, Professor of Wireless Systems, and John Shawe-Taylor, Director, Centre for Computational Statistics and Machine Learning) and the Structure & Motion Laboratory at the Royal Veterinary College (Alan Wilson, Professor of Locomotor Biology and Jim Usherwood, Wellcome Senior Research Fellow).

Interested in studying science at university?

The RVC offers the top rating animal and biological science courses in the UK. Graduates from our BSc Bioveterinary Sciences degree work in a broad range of fields across the world and we also offer a variety of taught MSc and research degrees.

Developing the technology and analysing the data

Lightweight multiple sensor tags for data collection

It is hard to track something that is constantly moving – especially when the need is to track individuals within a group of things that are moving in multiple dimensions, often quite close to each other – in a flock of birds for example.

A key aspect of the project concerns the development and customisation of data collection tags that are lightweight, low-cost and have low power consumption. These need to be attached non-invasively to different types of free ranging animals.

What the tags do

The data collection tags can contain a range of sensors:

  • GPS (Global Positioning Systems) and/or RF (radio frequency) tags to record location, speed and track up to five times per second.
  • Three axis accelerometers to measure acceleration, which provides stride timings for locomotion analysis and can be combined with other data to identify behaviour such as running, feeding, and sleeping. Acceleration data can also be integrated and combined with GPS data to provide more fine-grained location.
  • Three axis gyroscopes to measure changes in heading and orientation of the collar/animal.
  • Three axis magnetometers (electronic compasses) to measure the direction an animal is orientated and moving.
  • An altimeter to help measure changes in altitude accurately.

Combining the data

The data from these different sensors can be combined using a Kalman filter to deliver a more accurate and frequent (300 updates per second) position, speed and path than can be delivered by any one sensor in isolation.

Dynamic response of collars/tags

Importantly, the accelerometer data can be used to detect different behaviours and the tag software uses this along with information such as location, battery voltage and time of day to adjust data collection by other sensors. This will allow, for instance, detailed fine grain data collection during hunting and in areas of interest and less detailed, more general data (such as location) when resting.

Different tags for different purposes

In the design of these tags, a number of physical and technological factors have to be combined and customised depending on:

  • the species of animal
  • the environment (weather etc)
  • the duration of the measurement
  • the desired outcome in terms of data collection.

This customisation poses a number of challenges.

Technological challenges

GPS versus RF for location

Conventionally, GPS has been used to provide information on location. However, it is not suitable for all applications. It has the disadvantages of being power hungry (need for heavier batteries), bulky, expensive and with location accuracy usually limited to a metre. Furthermore the orientation of the animal (tag antenna) to the satellite is important, as is the absence of cover such as buildings and trees. Additionally, high dynamics can be a problem and, of course, there has to be satellite coverage.

An important part of this research is to develop radio-based localisation technology for use in applications where GPS is not a good option. RF (Radio Frequency) tags will give position relative to other tags. These tags are cheaper and potentially use less power as well as updating much more rapidly. They too have limitations – the data tends to be noisy and may contain reflection artefacts. Read more about our development of RF tracking tags.

Remote programmability

By incorporating the capacity to change the software settings and indeed the actual tag software remotely it becomes possible to vary extensively the data that is collected and the tag response for new experiments. Importantly, this minimises human/animal interaction – especially with wild animals.

Power on the move

The addition of solar panels to provide some of the power which reduces the need for heavy batteries. This also extends the life of the tag before human interaction becomes necessary.

Mechanical challenges

Weight is usually the most important of these – especially when attaching tags to birds – clearly the lighter the better for any animal. RF tags have the potential to be lighter and less bulky than GPS tags.

A robust casing and waterproofing is important – the tags need to be able to withstand all the same activities as the animal, such as foraging in vegetation, grooming or fighting, and being out in all weathers. Heat and sun can cause just as many problems for the electronics as rain, especially where the tags are left attached to the animal for long periods of time.

Attachment to the animal – this can be by means of harness, collar or with the very lightweight tags a form of superglue. Collars and harnesses are specially designed and made for each species to ensure a good fit without hindering movement or chafing. Only collars are used on animals in the wild. All tagged animals are checked regularly by visual observation.

New analytical methods and mathematical models

This part of the project is aimed at the development of new ways of analysing the large amounts of high-resolution data made possible by the development of complex sensors detailed above. Raw data needs to be processed in a way that will generate useful results in terms of the questions about the behaviour, activity and relationships of members of a group. The development of these analytical methods is also aimed at validating computer models (see Particle Swarm Optimisation below) that can be informed by solutions developed in nature (such as energy conservation within a group).

Matching recorded data to animal behaviour

One of the strengths of the multi-disciplinary approach in this project is that biologists and computer scientists are working closely together. Thus sensor-generated data can be validated and combined with observational data collected in field studies. This is done in a sophisticated way by directing ground-based or airborne video cameras to the animal-mounted sensor tags allowing animals to be filmed during high-speed activities such as hunting. This is achieved by sending tag position, track and speed data over a radio link to a motorised camera mount that knows its own position and attitude. The camera mount compensates for the movement of its platform and points the camera at or just in front of the tag. In this way sensor data can be accurately time-correlated with video data, making it possible to classify behaviours from sensor data where direct observation has not taken place and to capture the behaviour of uncollared individuals such as prey.

Machine learning techniques

The generation of so much detailed and disparate data necessitates some form of automatic processing using specially written algorithms. Two types of data are collected: ‘labelled’, which is created by a researcher observing and recording the activities of the animal whilst the collar tag also collects data, and ‘unlabelled’ which is the data recorded by the tag alone. The UCL machine learning specialists can identify features in the labelled data by comparing it with the observer records. This is used to train software known as a Support Vector Machine (SVM) to recognise the same features in unlabelled data. This allows data from a much greater time period to be scanned for recognisable behaviours.

Additionally, because SVM does not consider temporal information, the results are then post processed with a Hidden Markov Model (HMM) to allow the introduction of time variables. By this method, supporting information from previous behavioural events identified in the data can be used to build up a picture of the animals’ behaviour over time. Hidden Markov Models are used to smooth out SVM interpretation of data so that the classification of state changes is less influenced by transient behaviour changes – for example, hunting would only be identified as such if more than six consecutive gallop strides were detected.

Using accelerometer-based activity data from commercial collars on six free ranging cheetahs, it has been found that only a relatively small number of behavioural observations are needed for robust classification. The combination of SVM and HMM has enabled a number of relevant questions to be addressed:

  • How frequently do cheetah feed?
  • For how long?
  • When are these animals active during the day?
  • Do seasonal variables affect feeding or activity?

Social network analysis

By collecting locomotion and location data from many individuals within a group, it is possible to draw a picture of the social networks that exist between the individuals. For example, relative position of the individuals in a pack of hunting dogs during actual hunts can identify which dog or dogs usually take the lead. Information on social networks is highly informative about the extent to which the pack interacts and perhaps cooperates during hunting and locomotion.

Informing artificial systems – Particle Swarm Optimisation

In the field of computer science, models based on how social insects communicate and learn from each other are popular. They are used to solve problems involving large numbers of relatively simple autonomous units that can alter their behaviour as a result of interaction with other units. These approaches can be used in the design of complex systems as an alternative to approaches based on central planning.

These models are known as Particle Swarm Optimisation (PSO). The collection of detailed empirical data recording the behaviour of individuals within groups can be used to inform PSO models. The concept that group activity that has evolved in nature might be useful in the design of artificial complex systems is of considerable interest.

Animal behaviour studies

The research is structured in terms of a set of experimental scenarios, each of which has different scientific questions and technological challenges.

Study 1 – Hefted sheep

This study involved tracking free ranging hefted sheep in unfenced areas. The features of this research scenario are: outdoor, ground-based, low dynamics, large number and range of animals, uncluttered radio environment. Read more about the sheep study

Study 2 – Dynamic behaviour of packs of dogs

This study was devised to investigate hunting strategies in packs of African wild hunting dogs in an environment which is outdoor, ground-based, and with high dynamics. Read more about the African wild dog study

Study 3 – Flight patterns in migrating Ibises

This study was selected to investigate the flight characteristics of birds that fly in a V formation – ie outdoor, aerial, high dynamics, small number of individuals. Read more about the Ibis flight study

Study 4 – Aerodynamics of flight in flocking pigeons

This study investigates flight in a cluster flock, representing outdoor, aerial, high dynamics, large number of individuals. Read more about the pigeon study

Impact and Outcomes of the CARDyAL Project

Up to now, the internal dynamics of groups of animals has been difficult or impossible to study. We know almost nothing about the way that sheep heft, badgers socialise, packs of dogs run or pigeons fly. The development of sophisticated lightweight sensors which can be attached to animals and birds and adapted to collect many different types of highly dynamic data has implications for a number of fields.

In addition to the proof of concept studies which are part of this project, this technology with its transferable software and analytical base will have applications to other animal models, to the study of human group movement and the development of the computer models used to design complex systems.

The legacy of this project will enable many other investigations including:

  • the study of infectious disease transmission (e.g. foot and mouth, tuberculosis) between animals in a herd or flock through measurement of effective contact rates
  • domestic animal activity and behaviour (farm animal welfare)
  • wild animal activity and behaviour, providing new opportunities to inform conservation of endangered species
  • use of space by animals and humans (informing changes in architecture, management of facilities)
  • the development of synthetic optimisation techniques that are informed by empirical biological data (groups of aircraft, groups of unmanned aerial vehicles (UAVs)
  • the development of more realistic biologically informed models of group behaviour used in the computer graphics and animation industries

Key Publications

Portugal, S.J., Hubel, T.Y., Fritz, J., Heese, S., Trobe, D., Voelkl, B., Hailes, S., Wilson, A.M. & Usherwood, J.R. (2014). Upwash exploitation and downwash avoidance by flap phasing in ibis formation flight.
Nature 505, 399-402. doi:10.1038/nature12939

USHERWOOD, J.R., STAVROU, M., LOWE, J.C., ROSKILLY, K. AND WILSON, A.M. (2011). Flying in a flock comes at a cost in pigeons. Nature 474, 494-497. doi:10.1038/nature10164.

If you would like a copy of the paper and do not have institutional access please email Jusherwood@rvc.ac.uk.

Wilson, A.M. Lowe, J.C.Roskilly, K. Hudson,P.E. GolabekK.A.McNutt, J.W .Locomotion dynamics of hunting in wild cheetahs. Nature 498,185–189 (doi:10.1038/nature12295)

Top of page