A Benchmark for Action Recognition of Large Animals

Abstract

The action recognition of large animals plays an important role in the intelligent and modern farming. People often use the actions as the key factors to achieve scientific feeding and improve the animal welfare, and then the quality and productivity of animals are greatly promoted. However, most present action recognition methods focus on the actions of human (as pedestrian, athletes) or man-made objects (as cars, bikes). This paper proposes a benchmark to recognize and evaluate the actions of a kind of large animals namely the cows. First, we construct a dataset including 60 videos to describe the popular actions existing in the daily life of cows, and manually denote the target regions of cows on every frame in the dataset. Second, six famous trackers are evaluated on this dataset to compute the trajectory of cows which is the basis of actions recognition. Third, we define the method to recognize the actions of cows via the trajectories and validate the proposed method on our dataset. Many experiments demonstrate that our method of action recognition performs favorable in detecting the actions of cows, and the proposed dataset basically satisfies the action evaluations for farmers. The work in this paper provides an automatic and scientific method for famers to design a scheme to promote the quality and productivity of cows.