Postural behavior is an integral response of animals to their environmental stimuli. Huddling, nearly touching one another and spreading apart among group-housed animals are the qualitative stereotypes of postural patterns corresponding, respectively, to cold, comfortable and warm/hot sensation. It is these postural patterns that have been routinely used by animal caretakers to assess thermal comfort state of the animals and to adjust the environment or management accordingly. This human observation and manual adjustment approach, however, has two inherent pitfalls. First, it is impossible for the caretakers to attend the animals around the clock despite the circadian thermal needs of the animals. Secondly, the interpretation of optimal animal comfort behavior may vary considerably from one caretaker to another.

The goal of this project is to develop a machine vision system that automatically assesses thermal comfort state of swine and make according environmental adjustments. As the first step toward this goal, a study was conducted to examine the feasibility of classifying thermal comfort state of pigs by neural network (NN) analysis of their behavioral images. Specifically, pigs from 2 to 4 weeks of age were exposed in groups of 10 to cold, comfortable, and warm environments. Postural behaviors of the pigs were recorded and processed into binary images. Fourier coefficients (FC), moments (M), perimeter (P), area (A) and combination of M, P and A were evaluated as the feature representations of the processed behavioral images. Using these features as inputs, a 3-layer NN was established and then used to classify each of the postural images into the cold, comfortable or warm category. The combination of M, P and A as inputs to the NN model produced the best classification rate.

The results suggest that this approach has a good potential as a non-invasive assessment of the animal’s thermal comfort and ultimately may provide a practical management tool for an enhanced animal well-being. Subsequent studies have been conducted to develop algorithms for automatic image segmentation of the pigs in a commercial pen setting and automatic selection of the eligible behavioral images (via motion detection). Work is also underway to quantify relationships between physiological responses of the pigs (i.e., thermograph) and their behavioral displays. This information is essential for objectively building the NN, particularly fuzzy-logic NN, and thus improving the performance of the NN.