Mapping Activities onto a Two-Dimensional Emotions Model for Dog Emotion Recognition Using Inertial Data

Understanding animal reactions is essential for the welfare of animals, but accurately interpreting dogs’ emotions, despite their bond with humans, is challenging and often yields subjective results from human observers. Emotions manifest through physiological changes, such as heart rate fluctuations, or behavioral patterns, such as dog movements. In the present study, we measured and analyzed the movements of a group of dogs during four localized activities in two dimensions of emotion: arousal and valence. These activities (frustration, toy, abandonment, petting) were performed in natural settings while wearing the PATITA capture device. Statistical and temporal features were derived from acceleration signals and used to train various classification models. An average F1-score of 0.92 ( \(\sigma =\) 0.05) was scored when classifying the four emotions with the ExtraTrees classifier. This work contributes to a more accurate and consistent understanding of canine emotional states using dog movements, which has potential applications in shelters, day-care centers, and even homes, where dogs often spend a lot of time alone.