Usage of video monitoring and machine learning to monitor dairy cow behaviour

To monitor dairy cow behaviour, often accelerometer-based systems are used. Despite the fact that these systems are quite useful to study the behaviour of dairy cows, the usage of accelerometer-based systems has also its limitations, mainly due to battery lifetime restrictions, which limit the time-resolution with which behavioural data can be collected. Moreover, accelerometers aren’t suited to analyse social interactions of dairy cows (mounting behaviour, dominant-submissive interactions…). Therefore, studying the social behaviour of dairy cows is often still a manual job, limiting the scale at which studies can be executed, both in terms of time and animal numbers. In our project, we want to cope with these current limitations by using video monitoring and machine learning to collect behavioural data of dairy cows. Using camera’s instead of accelerometers allows us to collect behavioural data with a high time resolution, and this both on the level of the individual cow as on the level of social interactions. To process the video recordings to a behavioural timeline for each animal within the herd, a modular machine learning pipeline will be used. In the first step, machine vision will be applied to convert the videoframes to a set of keypoint-coordinates for each cow (nose, ears, tail implant, hookbones). In the second step the behaviour of the animals will be determined by analyzing the intra- and inter-individual position and movement of these keypoints. Due to the fact that this keypoint-based approach limits the amount of requested data-storage capacity, we will be able to collect and store behavioural data during multiple months, offering us the opportunity to execute a profound study on the (social) behaviour of dairy cows.To monitor dairy cow behaviour, often accelerometer-based systems are used. Despite the fact that these systems are quite useful to study the behaviour of dairy cows, the usage of accelerometer-based systems has also its limitations, mainly due to battery lifetime restrictions, which limit the time-resolution with which behavioural data can be collected. Moreover, accelerometers aren’t suited to analyse social interactions of dairy cows (mounting behaviour, dominant-submissive interactions…). Therefore, studying the social behaviour of dairy cows is often still a manual job, limiting the scale at which studies can be executed, both in terms of time and animal numbers. In our project, we want to cope with these current limitations by using video monitoring and machine learning to collect behavioural data of dairy cows. Using camera’s instead of accelerometers allows us to collect behavioural data with a high time resolution, and this both on the level of the individual cow as on the level of social interactions. To process the video recordings to a behavioural timeline for each animal within the herd, a modular machine learning pipeline will be used. In the first step, machine vision will be applied to convert the videoframes to a set of keypoint-coordinates for each cow (nose, ears, tail implant, hookbones). In the second step the behaviour of the animals will be determined by analyzing the intra- and inter-individual position and movement of these keypoints. Due to the fact that this keypoint-based approach limits the amount of requested data-storage capacity, we will be able to collect and store behavioural data during multiple months, offering us the opportunity to execute a profound study on the (social) behaviour of dairy cows.C