Assessing machine learning classifiers for the detection of animals' behavior using depth-based tracking

There is growing interest in the automatic detection of animals' behaviors and body postures within the field of Animal Computer Interaction, and the benefits this could bring to animal welfare, enabling remote communication, welfare assessment, detection of behavioral patterns, interactive and adaptive systems, etc. Most of the works on animals' behavior recognition rely on wearable sensors to gather information about the animals' postures and movements, which are then processed using machine learning techniques. However, non-wearable mechanisms such as depth-based tracking could also make use of machine learning techniques and classifiers for the automatic detection of animals' behavior. These systems also offer the advantage of working in set-ups in which wearable devices would be difficult to use. This paper presents a depth-based tracking system for the automatic detection of animals' postures and body parts, as well as an exhaustive evaluation on the performance of several classification algorithms based on both a supervised and a knowledge-based approach. The evaluation of the depth-based tracking system and the different classifiers shows that the system proposed is promising for advancing the research on animals' behavior recognition within and outside the field of Animal Computer Interaction.