Development and evaluation of tools and methodologies for estimating behaviour and predicting training outcome of working dogs

Background: The average training success rate in different dog industries is as low as 50% and the cost of training a guide dog is as high as 53,00 in Ireland. The key to reducing costs is in the assessment of trainee dogs for identifying likely to fail at an early stage. Objectives: This thesis aims to improve behavioural assessment methods by including machine learning methods to (1) predict future outcomes in trainee assistance dogs based on ratings and test batteries, and (2) estimate canine posture based on a recognition system specifically designed for working dogs. Methods: (1) Two standardised ratings were used, in particular, the Canine Behavioral Assessment and Research Questionnaire (C-BARQ) was completed by puppy raisers and the Monash Canine Personality Questionnaire - Revised (MCPQ-R) was answered by dog trainers. Rating data were independently analysed to investigate their relationship with training outcomes. The novel Assistance Dog Test Battery (ADTB) was designed to assess the suitability of trainee assistance dogs for assistance work during training. The test was conducted at 3 weeks - Data Collection 1 (DC1) - and 10 weeks - Data Collection 2 (DC2) - after the start of formal training to investigate the optimal timing to predict working outcomes. (2) Three Inertial Measurement Units (IMUs) were placed on the dogs in different positions (neck, back and chest) and five postures (walking, standing, sitting, lying down and body shake) were annotated. Advanced machine learning techniques were applied for the first time in this field to improve state-of-the-art posture prediction performance. Results: (1) The machine learning models achieved an area under the ROC of 0.84 and 0.85 when using the ratings C-BARQ and MCPQ-R to predict training outcome; and 0.74 and 0.84 when using the DC1 and DC2 of the ADTB to predict working outcomes, respectively. (2) The optimal canine posture classifier achieved an f1-weighted of 0.90. Conclusions: (1) These novel machine learning models provided the most effective early prediction of suitability for assistance work. The MCPQ-R and ADTB were demonstrated for the first time to be a reliable canine behavioural assessment method for estimating future outcomes in trainee dogs. (2) Comparison with previous work reveals a superior performance of the new canine posture estimation system for working

Grantor dc:publisher

University College Cork

Year dc:date.issued

2023

Author dc:creator

• Marcato, Marinara

Advisors dc:contributor.advisor

• Galvin, Paul
• O'Flynn, Brendan
• O'Mahony, Conor
• Tedesco, Salvatore

• Assitance dog
• Guide dog
• Training outcome
• Working outcome
• Canine posture recognition
• Machine learning

Statement dc:rights

• © 2023, Marinara Marcato.

Licence dc:rights.uri

• https://creativecommons.org/licenses/by-nc-nd/4.0/

Language dc:language.iso

en

Handle dc:identifier.uri

https://hdl.handle.net/10468/15906

OAI identifier oai:identifier

oai:cora.ucc.ie:10468/15906