Table 1.
Canine posture recognition algorithms using IMUs reported in literature.
Fig 1.
IMUs sensor placement on the dogs’ neck, back, and chest.
Table 2.
Behaviour Test Protocol for canine posture data monitoring and acquisition.
Table 3.
Categories of behaviour analysed.
Table 4.
Type, posture, and the number of observations in the IMU Posture dataset.
Fig 2.
Rolling window used for selecting data for feature extraction.
Hyper-parameters shown include re-sampling transition time (t_time) between postures, window size (w_size) used to calculate the statistical measures, and window offset (w_offset) to create separate observations.
Table 5.
Number of observations after feature extraction per dataset per posture.
Table 6.
Grid search hyper-parameter set for the classifiers.
Fig 3.
Direct estimator used in Classifier 1.
Fig 4.
Type, static and dynamic estimators used in Classifier 2.
Fig 5.
Anomaly and normal estimators used in Classifier 3.
Table 7.
Classifier 2 model’s hyper-parameter set selected by grid search, the time (in seconds) taken to train and validate the model using the development set, and f1-weighted performance on 10-fold validation sets (mean ± standard deviation) and test set.
Table 8.
Classifier 3 model’s hyper-parameter set selected by grid search, and f1-weighted performance on 10-fold validation sets (mean ± standard deviation) and test set.
Table 9.
Feature scores calculated by f-classification in Select K Best by IMU position and sensor.
Table 10.
Feature importance calculated by Random Forest classifier considering the 80 features previously selected by Select K Best.
Table 11.
Feature importance calculated by Random Forest classifier considering the 80 features selected by Select K Best by domain.
Table 12.
Classification metrics per posture achieved using the best models selected by grid search in Classifier 1 on the test set.
Table 13.
Classification metrics per posture achieved using the best models selected by grid search in Classifier 2 on the test set.
Table 14.
Classification metrics per posture achieved using the best models selected by grid search in Classifier 3 on the test set.
Fig 6.
Confusion matrix for Classifier 1 on the test set.
Fig 7.
Confusion matrix for Classifier 2 on the test set.
Fig 8.
Confusion matrix for Classifier 3 on the test set.
Table 15.
Classification metrics per posture achieved using the best models selected by grid search in Classifier 1 on the golden set.
Table 16.
Classification metrics per posture achieved using the best models selected by grid search in Classifier 2 on the golden set.
Fig 9.
Confusion matrix for Classifier 1 on the golden set.
Fig 10.
Confusion matrix for Classifier 2 on the golden set.
Fig 11.
Confusion matrix for Classifier 3 on the golden set.
Table 17.
Classification metrics per posture achieved using the best models selected by grid search in Classifier 3 on the golden set.
Table 18.
Classification metrics per posture achieved using the best models selected by grid search in Classifier 3 on the test and golden sets combined.
Fig 12.
Confusion matrix for Classifier 3 on the test and golden sets.
Table 19.
Comparison between Classifier 3 performance and previous studies reporting inter-subject classification performance metrics per posture.
The best geometric means between TPR and TNR (g-mean) are in bold.