Object detection technique applied to cattle production sites – Identifying individual Japanese black cows –

Abstract. A more efficient individual animal management system is necessary for cattle production in Japan. Individual management systems using wearable sensors for cattle at production sites have been introduced. However they entail many difficulties including high costs and stress of the animals. Recently, production sites have used network camera systems to monitor individual cow conditions. Some reports of earlier studies have described systems of analyzing animal behavior using object detection techniques. This report presents investigation of the possibility of identifying individual Japanese black cow faces using object detection technique. We examined the number of training images, the acquisition time of training images, the locations of cows during image capture, the bounding box range, and the distance between the camera and cow during image capture. Cow face images were taken using a digital video camera (FDR-X3000; Sony Corp.). Training data were labeled using annotation software (VoTT 1.72, VoTT 2.2.0; Microsoft Corp.). Then AI models were produced using the YOLOv4 object detection algorithm. Results demonstrated that the correct answer rate decreased to have different standing positions of cows between the training image and the test image. Accuracy improvement can be expected by including images of various standing positions and distances for each cow in the dataset.