The defence cascade as an indicator of animal welfare in the lab and field

Lead Research Organisation: University of Bristol
Department Name: Clinical Veterinary Science

Abstract

The assumption that non-human animals can subjectively experience negative emotional states, and hence suffer, underlies many people's concerns about animal welfare. Whilst we cannot measure the subjective states of other animals directly, it is important that we develop accurate indirect measures. It is also important that these measures can be used in both lab and field, especially on farms. This is because assessment of welfare is becoming a significant part of on-farm quality assurance schemes which aim to provide reliable information for the consumer about how food is produced. These schemes tend to assess welfare by measuring the resources available to animals (e.g. trough space) - a very indirect measure - or physical damage to the animal which may only reveal relatively severe problems. We aim to develop a new measure of welfare that more closely reflects the emotional states which lie at the heart of animal welfare concerns. This measure will also be of value in lab studies, precluding the need to isolate animals for testing. When animals are disturbed by an alerting stimulus that may signal danger (e.g. loud noise), they show a suite of defensive responses including startle and orientation, freezing and evaluation of the situation, and a final response of fleeing or resuming ongoing behaviour. Theoretical predictions, and studies of humans and rodents, suggest that the components of this 'defence cascade' (DC) are modulated by the individual's emotional state. For example, individuals in a negative state are predicted to show a stronger and faster initial startle response, to be more likely to show a final fleeing response, to make this decision faster, and to be slower to make a final decision to stay put, than individuals in a positive state. We will test these predictions in an important farmed species, the domestic pig. The pig shows a clear DC response to sudden noises (e.g. door slam), and we will develop standardised methods for inducing this response. Parts of the DC have been studied in humans and rodents under 'gold standard' laboratory conditions using force plate technology. It would be impractical to use such equipment on farms, but video images of the defence cascade could easily be collected and analysed to quantify the response. To develop these video-based methods, we will study individual pigs' DC responses under controlled conditions where we can obtain video and force plate data simultaneously. We will use computational image analysis to derive numerical output from the video footage and correlate this with the conventional 'gold standard' measures to determine whether image analysis accurately measures the DC response. We will also develop novel image analysis measures of group DC responses (pigs are usually group housed) - e.g. how rapidly a response spreads across a group - and investigate whether manipulation of welfare / emotional state (e.g. by housing groups in different conditions) affects DC responses as predicted above. We will then trial the image analysis measures of DC responses that best reflect emotional state / welfare on farms. Because the image analysis approach we use does not need to identify individual animals, we anticipate that it will cope with the visual challenges of a 'real-life' farm environment. We will accompany farm assurance assessors to farms, measure DC responses, collect data on the conditions on farms and in pens, and evaluate the relationships between these different measures using statistical techniques. This will show us how our DC measures reflect independent assessments of welfare at farm and pen level. We thus hope to produce a validated, non-invasive, quick and practical method for measuring animal welfare that can be adapted to other species, can be used in the field as well as the lab (including as part of farm assurance audits), and gets closer to reflecting the important emotional component of welfare than any existing field-based measure.

Technical Summary

This project will develop a new method for assessing animal welfare in the lab and field. Good field-based measures are needed because welfare criteria are important components of on-farm quality assurance schemes. However, most field-based measures provide limited information on the affective states of animals which are central to people's concerns about welfare. The defensive responses shown to an alerting stimulus (e.g. startle / orientation, freezing / evaluation, fleeing or resuming ongoing behaviour) may reflect underlying affective states. Theory, and studies of humans and rodents, suggest that this 'defence cascade' (DC) is modulated by affect. Individuals in a negative state are predicted to: show a stronger and faster initial startle response; be more likely to show a final fleeing response; make this decision faster; be slower to make a final decision to stay put, than individuals in a positive state. We will test these predictions in an important farmed species, the pig, which shows a clear DC response. We will first develop a method for measuring DC responses that uses computational image analysis (IA) of videos of the behaviour. Readouts from computational IA will be correlated with simultaneously collected 'gold standard' measures of DC responses (force plate, behavioural observation) in singly tested pigs to identify the best correlated measures. These IA measures will then be used to quantify DC responses of individual pigs in groups (group housing is the norm on farms) and to develop novel emergent measures of group response. The effects of manipulations of affect / welfare on IA measures will be used to test the predictions listed above. IA measures that best reflect affective modulation of DC responses will then be trialled on farms. The relationship between DC responses, independent assessments of welfare, and other 'background' variables will be determined using multi-level modelling.

Planned Impact

Our project will deliver tools and techniques for generating and monitoring defence cascade responses as validated indicators of animal affect and welfare under 'real-life' farm, and laboratory, conditions. It will have a number of benefits outside the academic sector. In the livestock industry, farm quality assurance schemes are increasingly keen to incorporate welfare measures based on the animal (rather than the resources provided) into their assessments. Our approach offers just such a measure, designed to reflect the animal's affective state - the core of people's concerns about welfare. Because of the image analysis procedures used, the approach should be readily generalisable beyond pigs (our study species) and hence applicable to a wide range of on-farm quality assurance schemes. It will benefit these schemes, the livestock industry, and the general public, by enhancing the scientific validity of on-farm welfare assessment, providing more accurate information for the public and decision makers, and contributing to improved quality of life for farm animals, and resulting product quality. We have agreement for collaboration within this project from the British Pig Executive (BPEX), the main pig industry levy board which is directly involved in livestock assurance schemes, allowing us to communicate findings directly to our industry contacts, as well as more generally through articles in the pig industry press. Our approach may also offer a new home-cage, group-based method for affect assessment in lab animals that alleviates the need for the stressful handling and isolation of animals that are tested in conventional Skinner box 'startle tests'. Any potential for further development and 'roll out' of this approach will be explored through our links with pharmaceutical companies (e.g. Pfizer, MSD). As well as providing a stronger scientific basis underpinning the quality assurance criteria upon which some members of the public base their food-purchasing decisions, the research will raise general public interest issues to do with 'animal mind' and 'animal emotion'. We will communicate relevant findings through press releases, articles in popular science magazines, public talks and exhibits (e.g. National Science and Engineering Week). We will engage with government and related NGO and charity bodies (e.g. Defra, HO, EU, NC3Rs, RSPCA), particularly those organizations involved in developing / policing legislation concerning farm animal welfare, through reports, meetings, expert advice on working groups, presentations to relevant bodies, and via contacts already formed through our work for, and advice to, these bodies. The methods developed in this research could be implemented by an integrated DC induction stimulus generator and image analysis software and / or hardware package. If the development of such a package appears feasible, we will liaise with the Bristol University Research and Enterprise Development Dept for advice on IP and commercialisation issues as necessary, and make an application for BBSRC 'follow-on funding'. The timescale of potential commercial benefits from our research is difficult to estimate until the project is in progress and the results become evident. Staff employed on this project already have considerable experience in communicating the impact of their work through interactions with the media, positions on advisory bodies and strategy / review groups, direct contact with the public at scientific exhibitions; personal contacts in governmental, NGO and industry groupings. Further training will be sought (e.g. BBSRC Media Training Course) to expand this skill set.
 
Description UPDATE Mar 2019: We have initiated collaboration with an EPSRC Research fellow team (Dr John Fennell and Dr Laszlo Tallas) and provided them with our extensive video dataset so that they can develop deep-learning approaches to detection of startle and freeze responses in pigs. This will complement our traditional computer vision approach. UPDATE March 2018: Delivery of the computer vision dataset was further delayed by problems with the algorithm that adjusts images to correct for lens distortion under field conditions. This has now been resolved and our computer science collaborators have delivered the dataset to us. We have carried out multilevel modelling bringing together the computer vision readouts and independent measures of on-farm conditions. These models reveal a number of interesting links between both startle and freeze components of the defence cascade and independent measures of on-farm conditions and measures of pig welfare. We are now in the process of writing one large paper spanning all the 'from lab to field' phases of the study, and which we hope to submit to Nature Methods in the first instance as planned. UPDATE: March 2017: The computer vision analysis of our data has proved to be more challenging than expected and necessitated hiring a PhD student (Oliver Moolan-Feroze) to carry out analysis and extraction of computer vision readouts under the supervision of Dr Sion Hannuna (PDRA now working on another project) and Dr Neill Campbell (Co-I). Oliver has run complex analyses and expects to deliver the finalised dataset of computer vision readouts within the next couple of weeks - much has been learnt about effective ways of analysing this particular dataset. Once these data are received from the computer science team, we will incorporate them into multilevel modelling analyses of treatment effects on conventionally recorded behaviour data. The delay has been extremely frustrating and has meant that published outputs have also been delayed because our plan to produce a paper detailing a 'from lab to field' story of method development in a high impact journal (see below) requires the full dataset to be available. We are still following this strategy. Its success will of course depend on how the final analysis pans out.

MARCH 2016: Because this multidisciplinary project is unusual in spanning from pure science to application, a strategic decision was made to withhold publication of data until the project was completed, and then to submit a substantial paper to a journal such as Nature Methods, summarising the progress from basic underpinning science in the lab to application on farm. This is still our plan (there have been delays due to a 9-month maternity leave by one RA and not the other, meaning that when the first RA returned to work, the second RA was only available for 50% of their time, hence slowing down the final computer vision analysis - see later for more details). In the meantime, we have presented the work at many behaviour / welfare and computer vision conferences, winning the best poster prize at the 2015 International Conference on Pig Welfare in Copenhagen. Overall, the project has been successful in achieving all objectives and developing a new and standardized way of measuring DC responses in the pig which can be used under field conditions. WE HERE PROVIDE A BRIEF 3-POINT SUMMARY OF KEY FINDINGS, FOLLOWED BY A LONGER REPORT:

(1) In the lab, we demonstrated that computer vision methods can be used to measure the defence cascade (DC: startle and freeze) responses of individual pigs to a sudden sound, showing strong correlation with human observer, force plate and motion capture metrics. The first in a series of sound stimuli showed the best correlations. The finding opens the way for using computer vision to assess DC responses in pigs on farms as an indicator of affective state and welfare (in humans and rodents the magnitude of startle responses is known to be a good indicator of an individual's affective state).

(2) In the lab, we upscaled the work on individual pigs to groups and used both environmental and pharmacological manipulations of affective state to see whether the manipulated states led to changes in DC responses. Correlations and multi-level modelling showed that diazepam treated pigs showed shorter group freeze durations that reboxetine treated pigs indicating that animals in a calmer state did indeed differ in their DC response. There were no observed differences in startle magnitude. Surprisingly, pigs in enriched environments showed a stronger group startle response than those in barren housing, though they also showed higher levels of salivary cortisol indicating a higher general state of arousal.

(3) We translated the method on to working farms and used our computer vision method to assess DC responses in groups of pigs kept under different conditions. We found that a stronger DC response was observed in pigs that were dirty, kept on slatted as opposed to solid flooring, had access to few manipulable objects, and were more avoidant of people. However, pigs with leg problems showed a less strong DC response, perhaps because of limitations to mobility. The DC response shows promise as a potential new and practical on-farm welfare indicator that can be quickly tested and measured from standard video recordings using computer vision methods.


A FULL REPORT ON KEY FINDINGS IS PROVIDED BELOW


This project brings together researchers in animal behaviour and welfare, computer vision, biomechanics, biostatistics and psychopharmacology to study the defence cascade (DC) response (startle, freeze, vigilance) which is shown by many mammalian species in response to sudden loud noises. The startle component of this response is reliably potentiated in humans and rodents experiencing negative affective states, and ameliorated in those experiencing positive states. The ultimate aim of the project is to develop a prototype practical approach to measuring this response in a commercially important farmed species, the domestic pig, which can be used as a new indicator of animal affect and welfare under field (farm) conditions. This would be an important contribution to the current increasing interest in and use of 'on-farm welfare' assessment and auditing which generally lacks animal-based measurements that really address the key aspect of an animal's welfare - its affective state. Instead there is a lot of reliance on measuring what resources are available to the animals (e.g. trough space, straw etc.) rather than exactly how these impact on the animal's welfare. The challenges of this project are to develop a method that can be used in the field - e.g. startle responses are usually measured using force plates (rodents) or electromyography (humans) - with a focus on (ultimately automated) computer vision analysis of video recordings; to investigate how DC responses vary according to manipulations of affective state in the pig; to test the approach under field conditions. We have addressed these challenges and have the following key findings:


Objective (i): to develop new methods for measuring the DC response in single pigs in the lab: In this part of the study we used a custom-built force plate to measure the DC response of pigs to a variety of standardised sudden loud noises. We also used motion-capture techniques (attaching markers to individual pigs), direct ethological observations, and computer vision analyses of videos (fast-capture: 50fps) of the responses. The aim was to see whether computer vision methods reliably correlate to the 'gold standards' of force plate, mo-cap, and ethological observation. 286 DC tests were carried out in total on 12 pigs from 9-19 weeks of age. Following initial analysis of videos, sparse feature tracking was selected as the most suitable method for computer vision analysis. Using sparse features is advantageous as only the easiest regions to track are considered which improves accuracy compared to dense feature tracking. Birchfield's implementation of the Kanade Lucas Tomasi (KLT) tracker was utilised. Accelerations of highly textured image regions were estimated using sparse feature tracking as indicators of the magnitude of the startle response (mean acceleration for the x KLT points displaying the largest accelerations). Data were extracted for a range of x values (from 10 to 150) in order to compare the measures - they were all highly correlated. Startle magnitude was defined as the maximum acceleration in a temporal window 0.7 seconds after the startle stimulus. For freeze, both speed and acceleration were extracted. Freeze duration for video footage and force plate data were defined as the total time that the speed of the y fastest KLT points and force plate readings respectively were below an empirically determined threshold for a continuous period of at least 0.4 seconds.

Correlations and multi-level statistical models controlling for the hierarchical nature of the data (e.g. tests nested within pig) were used to investigate the relationship between different measures. All measures, including computer vision, were strongly positively correlated for both startle and freeze metrics. Multi-level models indicated that repeated testing influenced startle responses suggesting habituation to the startle stimulus across time and/or decreasing responsiveness in older animals. We also found that a simple subjective rating of the pigs' affective state (from calm to highly-aroused) when they entered the force plate test area predicted the magnitude of startle and the duration of freeze responses. Calmer pigs showed smaller startle magnitudes and shorter freezes (although highly aroused pigs did not show this relationship as they were attempting to escape from the test arena). This result is in line with the findings of human and rodent studies. For the freeze response, all automated measures had very high sensitivity, i.e. they rarely missed a freeze when the human observer scored one. However, they all showed some lack of specificity, by generating false positives (detecting a freeze when the human observer did not categorise one). This was likely because pigs can be still for a variety of reasons (e.g. resting, sleeping, freezing) and it was difficult for our automated measures to discriminate these. Overall, the findings indicated that the KLT computer vision analysis of videos was a reliable indicator of startle magnitude and freeze, and that the results of the first test in a sequence are likely to be most reliable as pigs tend to be alert (not resting) and have not habituated to the stimulus.


Objectives (ii) and (iii): to upscale DC measures from single subjects to groups, and to test the relationship between DC responses and manipulations of affective state / welfare: On farms, pigs are usually housed in groups so a group-level DC measure is required. In this study we measured DC responses in 12 groups of 4 pigs in their home pens rather than individually on a force plate, thus more closely simulating the conditions under which the measures would be taken on farm. We used the KLT sparse feature tracking computer vision methods developed in objective (i) and validated using the force plate, and we also carried out ethological recordings. We designed a standardised metal clapper board device to generate a rapid onset loud sound as the board hit a metal bar (mean: 102dB at 1m). To manipulate affective state / welfare, pigs were either housed with multiple enrichment objects which, on the basis of previous studies, are thought to generate better welfare, or in barren conditions. Pigs from both housing groups were also exposed to pharmacological manipulations of affective state that we piloted on a different set of animals. From previous work, including in other species, we predicted that azaperone (a commonly used neuroleptic in pigs: 'stresnil') and diazepam (DZP) would have anxiolytic effects and induce a relatively positive affective state, whilst pentylenetetrazole and a single dose of reboxetine (REBOX) would have anxiogenic effects inducing a negative state. In a study using two doses of each drug we found that DZP and REBOX most reliably showed the predicted effects. Additionally, providing these as oral doses removed the need for restraint and drug injection which itself had marked effects on the pigs' behaviour and state.

Correlations and multi-level models once more showed a strong positive relationship between observer 'gold standard' measures of DC responses and the computer vision methods. We analysed data at group level (N=288 events) using the KLT method, and also at individual pig level (N=1152 events) by developing a program which identifies pig head and tail regions from an overhead video and tracks individual movements on this basis. The correlations between observer analysis and computer vision measures of group DC responses were better than those based on the individual pig data. We also confirmed that the lower specificity of computer vision measures of the freeze response is alleviated when the first response in a session is analysed in comparison to when all responses across a session are analysed, probably because pigs are more likely to be active at the start of a test sequence due to new human presence, hence minimising the likelihood of false positive freeze detections. Overall, these findings corroborate those from objective (i) that computer vision can be reliably used to measure startle and freeze components of the DC response, and they also indicate that this can be done at group level.

Multi-level models investigated the effects of experimental treatments and other factors on DC responses. Surprisingly, pigs in enriched housing showed higher levels of group startle magnitudes than those in barren environments. This ran counter to our predictions, but it was noticeable that levels of salivary cortisol both prior to and following DC tests, and aggression in the home pen, were higher in pigs from the enriched environment. It is possible that our enrichment manipulation, involving more frequent and unpredictable interaction with humans supplying food, enrichments etc., inadvertently generated more stressed or aroused pigs. REBOX treatment resulted in prolonged group freeze responses relative to control and DZP treatments, in line with predictions, but drug manipulations did not lead to significant differences in measures of group startle magnitude. These findings all remained significant when entering the effects of other variables into models, including the findings that freeze magnitudes decreased in response to repeated DC stimuli in a test sequence, and with the percentage of pigs standing, and that startle magnitudes increased in lighter groups of pigs and with the percentage of pigs standing. Overall, the findings support our hypothesis that freeze duration increases in pigs in a putatively more negative state, but evidence is not clear for the hypothesis that startle magnitude also increases. They also indicate that standardising initial activity of pigs (e.g. getting them to stand) is likely to remove a potential confound and also minimise false positive freeze data, and that pig weight should be taken into account when using the test.


Objective (iv): to translate the new methods to the 'real-life' field situation. Having developed a method for measuring DC responses that can be used under field conditions, the final phase of the study involved going on to farms to investigate how DC responses were related to independent measures of pig welfare on these farms. At this point, the RA in charge of the behavioural side of the project (bRA) took 9 month's maternity leave and this substantially affected the study. During the maternity leave period, the RA in charge of the computer vision part of the project (cRA) was able to move to 50% time due to the serendipitous availability of another 50% contract, and we were able to employ temporary technicians to collect some of the on-farm data. However, when bRA returned, cRA was consequently only able to devote 50% of his time to the project and this significantly delayed computer vision analysis of the last phase of the study. On returning to work, the bRA decreased her working hours to 60%FTE due to new child care commitments, thus extending the endpoint of her employment by a further 3 mo. Accordingly, BBSRC agreed a 12-month no-cost extension. Despite these issues, 21 farm visits were made and 423 DC tests carried out on 141 groups of pigs ranging from young weaners (N=39 groups) to grower/finishers (N=88 groups) to adult sows (N=14 groups), using the portable DC stimulus designed for objectives (ii) and (iii). On each farm, detailed 'conventional' measures of welfare were made for each group, including measures of resource availability and pig condition as used in the EU Welfare Quality® assessment protocol for pigs. Farm conditions provided new challenges to the computer vision analysis including poor lighting conditions, low ceilings necessitating use of fish-eye lenses, and different sized pens and camera angles. Extracted pixel trajectories thus needed to be transformed to a 'rectilinear projection' to recover straight lines and normalise the distances represented from the centre to the edge of the video frame. Perspective projection from camera angles not parallel to the ground plane also needed to be corrected to ensure that the motion of pigs was not over or under estimated depending on their position in the frame. Similar corrections were necessary to ensure that movements measured by the system were invariant to camera height. These were all achieved with significant work by cRA. Multi-level models were then used to relate computer vision DC measures to independent measures of welfare. These are currently being completed due to the issues mentioned above. Analysis of the 1st test in each sequence indicates that, controlling for pig age (older pigs show a lower startle magnitude), startle magnitude was higher in pigs that were housed on slatted as opposed to solid flooring, had access to fewer manipulable objects, were dirtier, and were more avoidant of people in a standardised test. Pigs with more leg problems (bursae, lameness) had lower startle magnitudes perhaps because a lack of mobility may interfere with startle responses. Overall, startle magnitude shows promise as a potential new on-farm welfare indicator.


Following the successful first objective of the task, and because this multidisciplinary project is unusual in spanning from pure science to application, a strategic decision was made to withhold publication of data until the project was completed, and then to submit a substantial paper to Nature Methods, summarising the progress from basic underpinning science in the lab to application on farm. This is still our plan. In the meantime, we have presented the work at many behaviour / welfare and computer vision conferences, winning the best poster prize at the 2015 International Conference on Pig Welfare in Copenhagen. Overall, the project has been successful in achieving all objectives and developing a new and standardized way of measuring DC responses in the pig which can be used under field conditions. Likely next steps will include further validation under field conditions, and automation of the computer vision methods to allow the test to be more made more widely available. We have received interest in using the approach for on-farm welfare assessment from researchers in the Netherlands, France and Denmark.
Exploitation Route We have a prototype method for using computer vision methods to measure defence cascade responses (startle, freeze) in pigs under real-life farming conditions as a measure of affective state and welfare. With further testing and development of the method, and automation of the computer vision analysis to provide an automated readout, this method could be used as a quick, practical, and validated measure of animal affect and welfare by on-farm welfare assurance schemes. Validated measures of animal affect are lacking in on-farm welfare assessment. The software approaches developed could also be used to measure other fast movements in other species (e.g. startle magnitude in humans) and to tackle the problem of translating perspective to a standardised plane of view (a major challenge when working in the real world with variable camera mounting positions and angles). UPDATE Mar 2019: our extensive database of video and accompanying metadata is available for use by other researchers and recently we have given access to a collaborating team who are interested in developing deep-learning approaches to detect startle and freeze behaviour.
Sectors Agriculture, Food and Drink,Digital/Communication/Information Technologies (including Software)

 
Description Winning the best poster prize at the International Conference on Pig Welfare (Copenhagen 2015) has resulted in some informal interest from industry representatives in our approach as a potential automated on-farm welfare assessment method. Actual impact and realisation of this possibility awaits the final data analyses, which have been delayed by the unexpectedly challenging nature of the computer vision task presented by the video sequences collected on-farm (see Key Findings). Subsequent development of methods for application will also be required.
First Year Of Impact 2015
Impact Types Societal

 
Description Member of 'Tesco & Animal Welfare' workshop group convened by Lord Krebs
Geographic Reach Europe 
Policy Influence Type Membership of a guideline committee
 
Description BBSRC Animal Welfare Research Network Grant
Amount £103,000 (GBP)
Organisation Biotechnology and Biological Sciences Research Council (BBSRC) 
Sector Public
Country United Kingdom
Start 01/2016 
End 12/2018
 
Description EPSRC Institutional Sponsorship
Amount £39,000 (GBP)
Organisation Engineering and Physical Sciences Research Council (EPSRC) 
Sector Academic/University
Country United Kingdom
Start 08/2015 
End 08/2016
 
Description Laterality in behavioural responses to novel and unexpected stimuli
Amount £2,000 (GBP)
Organisation Biotechnology and Biological Sciences Research Council (BBSRC) 
Sector Public
Country United Kingdom
Start 07/2013 
End 09/2013
 
Title Computer vision methods for measuring the defence cascade in freely moving groups of pigs 
Description The method we have developed allows a test of the defence cascade response of pigs (startle and freeze) to be carried out under field conditions, and for the response to be quantified using computer vision approaches. It is currently a prototype method but with potential for wider use by other researchers. Ultimately, the hope is that it could be used as a stand-alone test in on-farm welfare assurance scheme protocols. 
Type Of Material Technology assay or reagent 
Year Produced 2016 
Provided To Others? No  
Impact We have presented our work on this area at various meetings, and other research groups have been in contact with us concerning using the defence cascade measure in their own research. Currently they are doing this using standard ethological recording methods, but there is interest in using our computer vision approach in the future. 
 
Title High quality temporally synchronized multi-modal data set comprised of force plate, motion capture, Kinect and high speed multiple view video data 
Description These data are in an uncompressed format collected by highly skilled videographer and optimized for automated analysis. Large portions of this dataset have been exhaustively ground-truthed and offer a variety of welfare and computer vision research opportunities. 
Type Of Material Database/Collection of data 
Year Produced 2018 
Provided To Others? Yes  
Impact These data are the subject of ongoing analysis investigating the potential for computer vision recording of rapid defence cascade movements as indicators of affective state. MAR 2019: Our database is now being used by Dr John Fennell (EPSRC Research Fellow) to develop deep-learning approaches to detection of the defence cascade. 
 
Title Video database of defence cascade responses in pigs 
Description We have collected and stored a large number of video clips of rapid startle and freeze movements in pigs under both lab and field conditions. 
Type Of Material Database/Collection of data 
Provided To Others? No  
Impact The video database could be used by others to develop tools for quantifying rapid movements in animals. It also provides a range of sequences from farm situations which require perspective correction to ensure comparable computer vision quantification of movement. We have developed a method for perspective correction which has been implemented to allow our analysis, and which could be of use under other conditions and scenarios. 
 
Description Application of deep-learning to defence cascade dataset 
Organisation University of Bristol
Country United Kingdom 
Sector Academic/University 
PI Contribution We have teamed up with an EPSRC Research fellow team (Dr John Fennell and Dr Laszlo Tallas) at the University of Bristol and provided them with our extensive annotated video dataset so that they can develop deep-learning approaches to detection of startle and freeze responses in pigs. This will complement our traditional computer vision approach.
Collaborator Contribution Drs Fennel and Tallas will be developing deep-learning algorithms using our dataset.
Impact At present, we have no concrete outcomes, but Dr Fennell has already developed a 'pig detector' which can then be used to localise pixel analysis to the appropriate point/s in a video.
Start Year 2018
 
Description Combining computer vision techniques with behaviour, biosensing and machine learning 
Organisation Rothamsted Research
Department North Wyke Farm Platform
Country United Kingdom 
Sector Private 
PI Contribution A direct result of the new and ongoing collaboration between behaviour and welfare scientists on this grant is the development of new grant applications that involve not only our immediate team (Mendl, Campbell), but also sensor biologists, bioinformaticians and machine learning experts. An expression of interest has gone in from this wider team to the recent BBSRC machine Learning call (Feb 2017). Our contribution will be to provide knowledge and recording of behaviour as a potential early indicator of health and welfare problems under farm conditions that can be used to 'ground-truth' the other automated measures, and to develop machine vision methods for recording and detecting behaviour and behaviour change.
Collaborator Contribution Our partners' contributions are to provide expertise in sensor data collection (including mobile mass spectrograph technology for detecting airborne chemicals), 'big data' capture and interrogation methods, and machine learning of links between data changes and relevant health and welfare outcomes.
Impact We have just submitted an EOI to the recent BBSRC Machine Learning call (Feb 2017).
Start Year 2017
 
Description Combining computer vision techniques with behaviour, biosensing and machine learning 
Organisation University of Bristol
Department School of Veterinary Science
Country United Kingdom 
Sector Academic/University 
PI Contribution A direct result of the new and ongoing collaboration between behaviour and welfare scientists on this grant is the development of new grant applications that involve not only our immediate team (Mendl, Campbell), but also sensor biologists, bioinformaticians and machine learning experts. An expression of interest has gone in from this wider team to the recent BBSRC machine Learning call (Feb 2017). Our contribution will be to provide knowledge and recording of behaviour as a potential early indicator of health and welfare problems under farm conditions that can be used to 'ground-truth' the other automated measures, and to develop machine vision methods for recording and detecting behaviour and behaviour change.
Collaborator Contribution Our partners' contributions are to provide expertise in sensor data collection (including mobile mass spectrograph technology for detecting airborne chemicals), 'big data' capture and interrogation methods, and machine learning of links between data changes and relevant health and welfare outcomes.
Impact We have just submitted an EOI to the recent BBSRC Machine Learning call (Feb 2017).
Start Year 2017
 
Description Development and widening of collaboration with Bristol computer vision researchers 
Organisation University of Bristol
Country United Kingdom 
Sector Academic/University 
PI Contribution We have broadened our collaboration with Bristol computer vision researchers to start work involving parallel undergraduate student projects on individual recognition and behaviour detection in domestic cattle. This extends our research focus from pigs to another species and introduces new goals and techniques, including automated recognition of individual animals.
Collaborator Contribution Our current computer science collaborators (Campbell, Hannuna) are focusing on movement detection, whilst a new team member (Burghardt) is investigating individual recognition possibilities.
Impact Joint student projects are now ongoing
Start Year 2014
 
Description Trialling of defence cascade measures by other research groups in consultation with us 
Organisation Federal Institute for Animal Health, Celle
Country Germany 
Sector Academic/University 
PI Contribution We have had discussions with researchers from all four of the above institutes regarding trialling defence cascade measures of animal affect and welfare at their institutes. These have all followed from talks that we have given on our work. Some data have been collected and we have received videos from the collaborators, and other studies are ongoing.
Collaborator Contribution Raising the profile of this approach and describing the methods that can be used to collect useful defence cascade data under lab and farm conditions.
Impact The main outcome has been sharing of data so far. We await the results of our collaborators' data collection to see whether the data can be analysed using the methods that we have developed.
Start Year 2015
 
Description Trialling of defence cascade measures by other research groups in consultation with us 
Organisation Institute for Agri-Food Research and Technology
Country Spain 
Sector Public 
PI Contribution We have had discussions with researchers from all four of the above institutes regarding trialling defence cascade measures of animal affect and welfare at their institutes. These have all followed from talks that we have given on our work. Some data have been collected and we have received videos from the collaborators, and other studies are ongoing.
Collaborator Contribution Raising the profile of this approach and describing the methods that can be used to collect useful defence cascade data under lab and farm conditions.
Impact The main outcome has been sharing of data so far. We await the results of our collaborators' data collection to see whether the data can be analysed using the methods that we have developed.
Start Year 2015
 
Description Trialling of defence cascade measures by other research groups in consultation with us 
Organisation University of Copenhagen
Department Faculty of Health and Medical Sciences
Country Denmark 
Sector Academic/University 
PI Contribution We have had discussions with researchers from all four of the above institutes regarding trialling defence cascade measures of animal affect and welfare at their institutes. These have all followed from talks that we have given on our work. Some data have been collected and we have received videos from the collaborators, and other studies are ongoing.
Collaborator Contribution Raising the profile of this approach and describing the methods that can be used to collect useful defence cascade data under lab and farm conditions.
Impact The main outcome has been sharing of data so far. We await the results of our collaborators' data collection to see whether the data can be analysed using the methods that we have developed.
Start Year 2015
 
Description Trialling of defence cascade measures by other research groups in consultation with us 
Organisation University of Wageningen
Department Wageningen Food & Biobased Research
Country Netherlands 
Sector Academic/University 
PI Contribution We have had discussions with researchers from all four of the above institutes regarding trialling defence cascade measures of animal affect and welfare at their institutes. These have all followed from talks that we have given on our work. Some data have been collected and we have received videos from the collaborators, and other studies are ongoing.
Collaborator Contribution Raising the profile of this approach and describing the methods that can be used to collect useful defence cascade data under lab and farm conditions.
Impact The main outcome has been sharing of data so far. We await the results of our collaborators' data collection to see whether the data can be analysed using the methods that we have developed.
Start Year 2015
 
Title Perspective correction to enable computer-vision quantification of animal movement 
Description This is a method (not yet available as a fully formed software package) that has allowed us to translate videos that vary in perspective / camera angles etc. to a standardised perspective such that computer vision analysis of rapid animal movements can proceed without being affected by angle of view, distorted pixel-to-dimension ratios etc. 
Type Of Technology Software 
Year Produced 2015 
Impact This has been essential to allow us to quantify pig movements from on-farm video recordings where many uncontrollable variables influence the field of view / perspective / camera angle etc. It allows us to control for all these issues and hence produce comparable measures of movement even under very different conditions. 
 
Description Advisor to Humble Bee films on a programme exploring pig behaviour and biology 
Form Of Engagement Activity A broadcast e.g. TV/radio/film/podcast (other than news/press)
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Public/other audiences
Results and Impact I advised, including attending on location and assisting during post-production, on the filming of our pig cognition work for a TV documentary on pig behaviour and biology made by Humble Bee films and to be distributed internationally including by the Smithsonian Channel (USA) and Terra Mater Factual Studios (Austria). The film is now in post-production and so has yet to be released.
Year(s) Of Engagement Activity 2018
 
Description Appearance on BBC Radio 4's 'Would you eat an alien' documentary series (9-30 Dec 2015) 
Form Of Engagement Activity A press release, press conference or response to a media enquiry/interview
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Public/other audiences
Results and Impact This was a BBC Radio 4 series of four 30-min programmes covering the ethical and scientific issues concerning the welfare of animals managed by man, how to scientifically assess their welfare, and how this bears upon decisions to use animals for food.
Year(s) Of Engagement Activity 2015
URL http://www.bbc.co.uk/programmes/b06r82mz
 
Description Interview for Belgian science magazine 'eos' 
Form Of Engagement Activity A magazine, newsletter or online publication
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Public/other audiences
Results and Impact Interview with Belgian science magazine 'eos' regarding the cognitive and affective lives of domestic pigs.
Year(s) Of Engagement Activity 2017
URL http://eostrace.be/artikelen/de-slimste-big-ter-wereld
 
Description Interview for German TV production company 'nonfiction society' on animal emotion and the ethics of animal treatment 
Form Of Engagement Activity A broadcast e.g. TV/radio/film/podcast (other than news/press)
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Public/other audiences
Results and Impact German TV crew came to Bristol to interview me about animal emotions, how one might go about studying them and assessing them scientifically, and what the implications of understanding animal affective states may have for our treatment of animals in many contexts (e.g. laboratories, farming etc.)
Year(s) Of Engagement Activity 2019
URL https://www.nonfictionsociety.de/
 
Description Lecture and workshops at Millfield School 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach Regional
Primary Audience Schools
Results and Impact Mike Mendl spoke on his work on animal emotion in the opening talk of the 2015 Millfield School Biology and Psychology Symposium, attended by c.350 sixth form students from 10 schools in the region. He was also involved in two workshop events after the talk.
Year(s) Of Engagement Activity 2016
URL https://millfieldschool.com/senior/news-article/Millfield-hosts-biology-and-psychology-symposium
 
Description Poster display on 'New measures of animal emotion and welfare' at Bristol Neuroscience Festival 
Form Of Engagement Activity Participation in an activity, workshop or similar
Part Of Official Scheme? Yes
Geographic Reach Regional
Primary Audience Public/other audiences
Results and Impact We presented a poster display and laptop demonstration of our work on new measures of animal emotion and welfare at the 2-day Bristol University Bristol Neuroscience Festival on 11-12 October 2013. Our poster and laptop demonstration was manned for both days of this 2-day festival. It is estimated that the festival was attended by c.3000 members of the public, including school children. A poster and laptop demonstration of some of our work was produced.

General discussion of how one can measure animal affect and welfare in a scientific way.
Year(s) Of Engagement Activity 2013
URL http://www.bristol.ac.uk/neuroscience/events/diary/2013/101365.html
 
Description Talk at Public Engagement event - Bristol Neuroscience Festival 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? Yes
Geographic Reach Regional
Primary Audience Public/other audiences
Results and Impact Lecture on 'Optimistic animals': a new measure of animal emotion' was one of a series of public lectures over the 2-day festival, and was given on 12th October. It is estimated that the festival was attended by c.3000 members of the public, including school children. the lecture was attended by c.290 members of the public.

Questions received afterwards concerning the assessment of emotions in non-human animals. Interest in our exhibit display at the Festival.
Year(s) Of Engagement Activity 2013
URL http://www.bristol.ac.uk/neuroscience/events/diary/2013/101365.html
 
Description Workshop on Emotion and Value-Based Decision-Making in Humans and Animals organised at the Gatsby Computational Neuroscience Unit, UCL, London 
Form Of Engagement Activity Participation in an activity, workshop or similar
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Professional Practitioners
Results and Impact We (Paul, Mendl, Dayan) organised an international workshop on Emotion and Value-Based Decision-Making in Humans and Animals held at the Gatsby Computational Neuroscience Unit, UCL, London, UK. The workshop was attended by leading international experts in affect and decision-making and also by local postgraduate students from our group. It provided a state-of-the-art snapshot of work in the area, including on animal affect and decision-making
Year(s) Of Engagement Activity 2016
 
Description Workshop on Measuring Animal Emotion held at the Assembly Rooms, Edinburgh 
Form Of Engagement Activity Participation in an activity, workshop or similar
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Professional Practitioners
Results and Impact We ran a workshop on Measuring Animal Emotions at the Assembly Rooms in central Edinburgh. It was attended by over 300 delegates including academics, undergraduate and postgraduate students, representatives from charitable organisations, funders, industry and government. A plenary talk was followed by 10 short talks on the state of the art with respect to 10 different approaches to measuring animal emotion and welfare. The workshop ended with a panel discussion which engaged the audience.
Year(s) Of Engagement Activity 2016
URL http://awrn.co.uk/2016/08/02/986/