Development of image-based chicken detection algorithms for monitoring farm activity indicators

Abstract

Guaranteeing consumers that broilers are produced in a way that respects animal welfare is the basis of the farming profession, but civil society is calling for greater transparency in farming practices. Farm welfare assessment methods are carried out punctually and require time and the presence of a trained observer. Image analysis, on the other hand, enables continuous, real-time measurements to be taken without disturbing the animals. The aim of this article is to evaluate the performance of a broiler detection algorithm based on image analysis and using artificial intelligence to quantify individual animal mobility. The majority of the images used to train the model, as well as the test database, reflect the commercial densities used in broiler farming. They are also representative of the different physiological stages of the broilers. The model achieves a detection rate of 80% across all ages. However, the sensitivity of the algorithms increases with the age, rising from 66% for 0-day-old broilers to 90% for 40-day-old broilers. As it stands, broiler detection is strongly linked to the number of pixels defining the animal. This detection model is the first step needed to carry out individual tracking over time and therefore over several successive images. The quality of this tracking is highly dependent on the system ability to detect the animals correctly, but also on their activity. Tracking performance is not discussed in this article. In view of the promising results, the data generated will be used to assess broiler welfare indicators and detect health problems early on in commercial broiler farms and could also be used in experimental farms.