Counting cattle in UAV images using convolutional neural network

Abstract

Determining the number of cattle in countries with the most extensive livestock and large pastures is difficult requires a lot of time of the farm workforce and is stressful to the animals. Counting cattle in an agile way using tools that can automatically perform this task would be very useful for herd conferences and farm management. The proposed solution is to count cattle through images acquired by Unmanned Aerial Vehicles (UAVs). This allows faster acquisition of the number of cattle in a given area so management tasks can be more accurately done and, allowing better interventions towards technical improvements. Thus, models of architectures from Convolutional Neural Networks (CNN) to YOLOv4 and YOLOv5 models (X, L, M, and S) were used for comparison. In order to evaluate the efficiencies of these solutions for the bovine counting, 878 images were acquired through flights of 20, 40, 80, and 100 m high. YOLOv4 obtained a precision of 0.90, and the YOLOv5 architectures (X, L, M, and S) were 0.98, 0.96, 0.93, and 0.96, respectively. In conclusion, the use of CNN to identify and count cattle from UAV images is a viable solution.