The Use of Triaxial Accelerometers and Machine Learning Algorithms for Behavioural Identification in Domestic Dogs (Canis familiaris): A Validation Study

Abstract

Assessing the behaviour and physical attributes of domesticated dogs is critical for predicting the suitability of animals for companionship or specific roles such as hunting, military or service. Common methods of behavioural assessment can be time consuming, labour-intensive, and subject to bias, making large-scale and rapid implementation challenging. Objective, practical and time effective behaviour measures may be facilitated by remote and automated devices such as accelerometers. This study, therefore, aimed to validate the ActiGraph^® accelerometer as a tool for behavioural classification. This study used a machine learning method that identified nine dog behaviours with an overall accuracy of 74% (range for each behaviour was 54 to 93%). In addition, overall body dynamic acceleration was found to be correlated with the amount of time spent exhibiting active behaviours (barking, locomotion, scratching, sniffing, and standing; R² = 0.91, p < 0.001). Machine learning was an effective method to build a model to classify behaviours such as barking, defecating, drinking, eating, locomotion, resting-asleep, resting-alert, sniffing, and standing with high overall accuracy whilst maintaining a large behavioural repertoire.

Keywords: algorithm; behaviour classification; overall activity; random forest.

Figure 1

(A) ActiGraph^® wGT3X-BT device (B) Ventral view of the ActiGraph^® wGT3X-BT accelerometer, which was consistently orientated, placed within a protective housing, and fitted ventrally to the collars of the dogs.

Figure 2

Diagram of the outdoor observation paddock showing its dimensions and features, including the positioning of the two surveillance cameras used to monitor the animals. Note that the image has not been drawn to scale.

Figure 3

The five modelling rounds showing the total number of observation(s) of each behaviour used for the training data set. The total data set consisted of 129,615 observations, excluding other and out of site categories. Of this data set, 90,741 (70%) and 38,874 observations (30%) were used to train and test the models, respectively. Abbreviations: Lateral (L.) Cells highlighted in orange have been removed from the subsequent models due to low accuracy/precision and/or sample size.

Figure 4

Raw (30 Hz) triaxial (x axis = blue line, y axis = orange line, z axis = grey line) acceleration profiles for each of the behaviours classified by Model 4. A total of 3 s to 5 s present per behaviour, although insufficient continuous acceleration data (Insuf. data) were available for defecation behaviour. These behaviours have been grouped according to the categories used for Model 5: Active, inactive, maintenance.

Figure 5

Graph showing the correlation between the time spent active (%) per hour (h) and the total ODBA per hour (h).