Publications

2026

1. Automated detection of piling behaviour in UK laying hen flocks

   Jack O’Sullivan ,  Helen E. Gray ,  Genevieve Moat , and 1 more author

   Royal Society Open Science, May 2026

   Abs PDF

   Piling is an understudied group behaviour where chickens gather in dense clusters. It appears to be widespread in commercial laying hens, leads to mortalities (smothers) and impacts production. Automated detection of piling could support: (i) stockpersons to reduce the behaviour and (ii) scientific efforts to study the behaviour. Computer vision-based classification is increasingly used in animal behaviour research to accelerate annotation of datasets. Here we aimed to automatically detect piling using the well-established residual network (ResNet) architecture convolutional neural network (CNN) to classify video frames of piling and non-piling chickens in commercial free-range laying farms. Data from 10 flocks (8 for training, 2 for testing) were used. The model achieved high classification accuracy, correctly identifying 89.6% of individual frames and 90.9% of aggregated events, though misclassifications at both levels were driven primarily by false negatives (16.2% for frames and 16.8% for events). Cases of misclassification could arise from not accounting for the temporal aspect of piling, human annotation error, image quality, flock-specific characteristics or ambiguous cases. ResNet demonstrated good performance for classifying piling from non-piling from still frames in brown laying hens in commercial houses but would need further training data to be robust to different housing types and chicken breeds.

2025

1. Research note: The effect of passionflower supplementation on feather pecking in laying hens

   Elizabeth Brass ,  Jack O’Sullivan ,  and  Helen Gray

   Poultry Science, May 2025

   Abs HTML

   Feather pecking is a significant issue in non-caged poultry welfare that results in the removal or damage of the feather of a hen. The most common forms are classified into gentle feather pecking and severe feather pecking which, if undeterred, can develop into cannibalism. This case study explored one aspect of the prevention of feather pecking, investigating if the feed additive Gallicalm, containing Passionflower, reduced feather pecking behavior in a free-range flock. Video footage over 6 weeks was analysed for feather pecking incidence in 2-week phases; Pre-Treatment, Treatment and Post-Treatment. Standard commercial rations were fed in the Pre-Treatment and Post-Treatment phases, with the Treatment phase receiving the standard commercial ration plus 1 kg per ton of Gallicalm. Feather scores were completed using the AssureWel method at the end of each phase, with production data collected through an online flock management tool. A total of 373 minutes of footage from 18 days was analysed for pecking behavior. Supplementation resulted in reduced number of severe feather pecks in the Pre-Treatment phase to the Treatment phase. Gentle pecking failed to decrease significantly during Gallicalm Treatment but increased in the post-Treatment phase. Aggressive, stereotypical and beak pecking were rare in all experimental phases. Feather scores deteriorated between the Pre-Treatment and Treatment phase but plateaued between the Treatment and Post-Treatment phase. This case study provides the first evidence of passionflower-containing supplements reducing feather pecking in laying hens. Given the billions of laying hens kept globally and the extensive welfare and economic issues associated with feather pecking, we advocate for further study to build on our initial findings.

2. Research note: Impacts of piling behavior on temperature and carbon dioxide in laying hen sheds

   Helen E Gray ,  Jack O’Sullivan ,  and  Lucy Asher

   Poultry Science, May 2025

   Abs HTML

   Piling, a high density of chickens choosing to gather, is increasingly being recognized as a major problem behavior in the laying hen industry with both economic and welfare impacts. Groups of animals in close proximity generate heat, and observations of piling have noted instances of over 1200 hens in direct contact. Here, we investigate the impact of piling behavior on the temperature of the chicken shed. Since heat stress causes panting in chickens, piling also has potential to increase the CO2 concentration and as such, we also investigated the impact of piling on CO2. We used annotations of piling behavior from video footage of approximately 21 days for each of 12 flocks. The Birdbox system for flock management was used to obtain matched logged temperature (°C) and CO2 (ppm) from two sensor stations every minute, resulting in 17,396 datapoints. Bayesian multilevel modelling was used to estimate the effects of pile number and duration on temperature and CO2, including an effect to control for time of day. Since baseline daily fluctuations in temperature and CO2 could not be obtained, time of day effects were modelled in different ways, as autoregressive, random intercept, sinusoidal and polynomial terms. As autoregressive and non-autoregressive models could not be directly compared, we present the results of the autoregressive and best fit non-autoregressive models. We found no association between piling and temperature or CO2 for the autoregressive models but did find an association between pile number, pile duration and temperature in the random effects model. Higher temperature was associated with an interaction between increasing pile numbers and increasing pile duration. Since the effect size was very small and this result was not replicated in the autoregression model it should be interpreted with caution but does provide interesting rationale for future work investigating behavior-environment interactions.

2021

1. The automatic classification of canine state

   Jack O’Sullivan

   Newcastle University , May 2021

   Abs Bib HTML PDF

   It is understood gait has the potential to be used as a window into neurodegenerative disorders, identify markers of subclinical pathology, inform diagnostic algorithms of disease progression and measure the efficacy of interventions. Dogs’ gaits are frequently assessed in a veterinary setting to detect signs of lameness. Despite this, a reliable, affordable and objective method to assess lameness in dogs is lacking. Most described canine lameness assessments are subjective, unvalidated and at high risk of bias. This means reliable, early detection of canine gait abnormalities is challenging, which may have detrimental implications for dogs’ welfare. In this paper, we draw from approaches and technologies used in human movement science and describe a system for objectively measuring temporal gait characteristics in dogs (step-time, swing-time, stance-time). Asymmetries and variabilities in these characteristics are of known clinical significance when assessing lameness but presently may only be assessed on coarse scales or under highly instrumented environments. The system consists an inertial measurement unit, containing a 3-axis accelerometer and gyroscope coupled with a standardized walking course. The measurement unit is attached to each leg of the dog under assessment before it is walked around the course. The data by the measurement unit is then processed to identify steps and subsequently, micro-gait characteristics. This method has been tested on a cohort of 19 healthy dogs of various breeds ranging in height from 34.2 cm to 84.9 cm. We report the system as capable of making precise step delineations with detections of initial and final contact times of foot-to-floor to a mean precision of 0.011 s and 0.048 s, respectively. Results are based on analysis of 12,678 foot falls and we report a sensitivity, positive predictive value and F-score of 0.81, 0.83 and 0.82 respectively. To investigate the effect of gait on system performance, the approach was tested in both walking and trotting with no significant performance deviation with 7249 steps reported for a walking gait and 4977 for a trotting gait. The number of steps reported for each leg were approximately equal and this consistency was true in both walking and trotting gaits. In the walking gait 1965, 1790, 1726 and 1768 steps were reported for the front left, front right, hind left and hind right legs respectively. 1361, 1250, 1176 and 1190 steps were reported for each of the four legs in the trotting gait. The proposed system is a pragmatic and precise solution for obtaining objective measurements of canine gait. With further development, it promises potential for a wide range of applications in both research and clinical practice.

   @phdthesis{o2021automatic,
     title = {The automatic classification of canine state},
     author = {O’Sullivan, Jack},
     year = {2021},
     school = {Newcastle University},
     url = {http://theses.ncl.ac.uk/jspui/handle/10443/5397},
     dimensions = {true}
   }

2018

1. A step in the right direction: an open-design pedometer algorithm for dogs

   Cassim Ladha ,  Zoe Belshaw ,  Jack O’Sullivan , and 1 more author

   BMC veterinary research, May 2018

   Abs Bib HTML PDF

   Background: Accelerometer-based technologies could be useful in providing objective measures of canine ambulation, but most are either not tailored to the idiosyncrasies of canine gait, or, use un-validated or closed source approaches. The aim of this paper was to validate algorithms which could be applied to accelerometer data for i) counting the number of steps and ii) distance travelled by a dog. To count steps, an approach based on partitioning acceleration was used. This was applied to accelerometer data from 13 dogs which were walked a set distance and filmed. Each footfall captured on video was annotated. In a second experiment, an approach based on signal features was used to estimate distance travelled. This was applied to accelerometer data from 10 dogs with osteoarthritis during normal walks with their owners where GPS (Global Positioning System) was also captured. Pearson’s correlations and Bland Altman statistics were used to compare i) the number of steps measured on video footage and predicted by the algorithm and ii) the distance travelled estimated by GPS and predicted by the algorithm.
   Results: Both step count and distance travelled could be estimated accurately by the algorithms presented in this paper: 4695 steps were annotated from the video and the pedometer was able to detect 91%. GPS logged a total of 20,184 m meters across all dogs; the mean difference between the predicted and GPS estimated walk length was 211 m and the mean similarity was 79%.
   Conclusions: The algorithms described show promise in detecting number of steps and distance travelled from an accelerometer. The approach for detecting steps might be advantageous to methods which estimate gross activity because these include energy output from stationary activities. The approach for estimating distance might be suited to replacing GPS in indoor environments or others with limited satellite signal. The algorithms also allow for temporal and spatial components of ambulation to be calculated. Temporal and spatial aspects of dog ambulation are clinical indicators which could be used for diagnosis or monitoring of certain diseases, or used to provide information in support of canine weight-loss programmes.

   @article{ladha2018step,
     title = {A step in the right direction: an open-design pedometer algorithm for dogs},
     author = {Ladha, Cassim and Belshaw, Zoe and O’Sullivan, Jack and Asher, Lucy},
     journal = {BMC veterinary research},
     volume = {14},
     number = {1},
     pages = {1--10},
     year = {2018},
     publisher = {BioMed Central},
     doi = {10.1186/s12917-018-1422-3},
     url = {https://doi.org/10.1186/s12917-018-1422-3},
     dimensions = {true},
   }

2017

1. GaitKeeper: a system for measuring canine gait

   Cassim Ladha ,  Jack O’Sullivan ,  Zoe Belshaw , and 1 more author

   Sensors, May 2017

   Abs Bib HTML PDF

   It is understood gait has the potential to be used as a window into neurodegenerative disorders, identify markers of subclinical pathology, inform diagnostic algorithms of disease progression and measure the efficacy of interventions. Dogs’ gaits are frequently assessed in a veterinary setting to detect signs of lameness. Despite this, a reliable, affordable and objective method to assess lameness in dogs is lacking. Most described canine lameness assessments are subjective, unvalidated and at high risk of bias. This means reliable, early detection of canine gait abnormalities is challenging, which may have detrimental implications for dogs’ welfare. In this paper, we draw from approaches and technologies used in human movement science and describe a system for objectively measuring temporal gait characteristics in dogs (step-time, swing-time, stance-time). Asymmetries and variabilities in these characteristics are of known clinical significance when assessing lameness but presently may only be assessed on coarse scales or under highly instrumented environments. The system consists an inertial measurement unit, containing a 3-axis accelerometer and gyroscope coupled with a standardized walking course. The measurement unit is attached to each leg of the dog under assessment before it is walked around the course. The data by the measurement unit is then processed to identify steps and subsequently, micro-gait characteristics. This method has been tested on a cohort of 19 healthy dogs of various breeds ranging in height from 34.2 cm to 84.9 cm. We report the system as capable of making precise step delineations with detections of initial and final contact times of foot-to-floor to a mean precision of 0.011 s and 0.048 s, respectively. Results are based on analysis of 12,678 foot falls and we report a sensitivity, positive predictive value and F-score of 0.81, 0.83 and 0.82 respectively. To investigate the effect of gait on system performance, the approach was tested in both walking and trotting with no significant performance deviation with 7249 steps reported for a walking gait and 4977 for a trotting gait. The number of steps reported for each leg were approximately equal and this consistency was true in both walking and trotting gaits. In the walking gait 1965, 1790, 1726 and 1768 steps were reported for the front left, front right, hind left and hind right legs respectively. 1361, 1250, 1176 and 1190 steps were reported for each of the four legs in the trotting gait. The proposed system is a pragmatic and precise solution for obtaining objective measurements of canine gait. With further development, it promises potential for a wide range of applications in both research and clinical practice.

   @article{ladha2017gaitkeeper,
     title = {GaitKeeper: a system for measuring canine gait},
     author = {Ladha, Cassim and O’Sullivan, Jack and Belshaw, Zoe and Asher, Lucy},
     journal = {Sensors},
     volume = {17},
     number = {2},
     pages = {309},
     year = {2017},
     publisher = {MDPI},
     doi = {10.3390/s17020309},
     url = {https://doi.org/10.3390/s17020309},
     dimensions = {true},
   }

2. Identification of behaviours from accelerometer data in a wild social primate

   Gaelle Fehlmann ,  M Justin O’Riain ,  Phil W Hopkins , and 4 more authors

   Animal Biotelemetry, May 2017

   Abs Bib HTML PDF

   Background: The use of accelerometers in bio-logging devices has proved to be a powerful tool for the quantification of animal behaviour. While bio-logging techniques are being used on wide range of species, to date they have only been seldom used with non-human primates. This is likely due to three main factors: the long tradition of direct field observations, a difficulty of attaching bio-logging devices to wild primates and the challenge of deciphering acceleration signals in species’ with remarkable locomotor and behavioural diversity. Here, we overcome these aforementioned obstacles and provide methodology for identification of behaviours from accelerometer data of wild chacma baboons (Papio ursinus) in Cape Town, South Africa.
   Results: We apply machine learning techniques to process complex accelerometer data, collected by bespoke tracking collars to quantify a range of behaviours (focusing on locomotion and foraging behaviour). We successfully identify six broad state behaviours that represent 93.3% of the time budget of the baboons. Resting, walking, running and foraging were all identified with high recall and precision representing the first classification of multiple behavioural states from accelerometer data for a wild primate.
   Conclusion: Our ‘end to end’ process—from collar design and build to the collection and quantification of acceleration data—provides advantages over gathering data by traditional observation, not least because it affords data collection without the presence of an observer which may affect an animal’s behaviour. Furthermore, our methodology and findings open new possibilities for the fine-scale study of movement and foraging ecology in wild primates, and in particular our baboon study population which is in conflict with people.

   @article{fehlmann2017identification,
     title = {Identification of behaviours from accelerometer data in a wild social primate},
     author = {Fehlmann, Gaelle and O’Riain, M Justin and Hopkins, Phil W and O’Sullivan, Jack and Holton, Mark D and Shepard, Emily LC and King, Andrew J},
     journal = {Animal Biotelemetry},
     volume = {5},
     pages = {1--11},
     year = {2017},
     publisher = {BioMed Central},
     doi = {10.1186/s40317-017-0121-3},
     url = {https://doi.org/10.1186/s40317-017-0121-3},
     dimensions = {true},
   }