CS229 Project Report Detect Leaders in Cow Group Movement using Pairwise Distances

Studies in animal social behavior show that group-living animals, such as cows, travel together in a collective pattern known as spontaneous group movement [2] Such movement exhibits the leader-follower phenomenon, that is, some individual are more likely than others, to initiate group movement that causes others to follow. We call these cows leaders of the group movement. An useful tool in cattle management is to automatically identify the group movement leaders by analyzing cow motion data. The goal of this project is thus to find out whether we can model the leader probability of individual cows based on the pattern of recent group movements. We define the leading cows of a group movement to be occupants of the front positions in the direction which the group is moving. Although it is easy to determine a cow’s relative position within the group given its accurate geographical location tracked using GPS, it is often infeasible to deploy GPS devices to cattle on a large scale due to cost, power and signal limitation. An alternate way of data collection proposed by students in ETH Zurich University1, is dispersing the environment with sparse landmarks and equipping the cows with inexpensive sensors that records contact events with other sensors and landmarks[1]. Previous work by Stanford students Daniel Chen and Johnathan Jiang showed that it is possible to estimate the distances between sensors from contact information. Knowing the cow-to-cow distance, we define the neighbour distance distribution of Cow i to be the distribution of distances between Cow i and other cows within the group. We can describe this distribution using a Gaussian model, p(di) ∼ N(μi, σ2 i ) where μi and σ2 i are empirical estimations of mean and variance. We expect leading cows to have very different μi and σi from non-leading cows. Figure 1 compares the neighbour distance of Cow 2 and Cow 11 at frame 700. In this case, we have μ2 = 8.09, μ11 = 4.80, σ 2 2 = 21.42 and σ 2 11 = 7.18. Base on this intuition, this project applies several supervised learning models to estimate a cow’s leader probability from its neighbour distance distribution.