301 Determining Individual Dry Matter Intake of Charolais Bulls from Ingestion Time Recorded with Rgb-D Cameras

Abstract

Abstract Monitoring individual dry matter intake (DMI) may help decision-making of cattle producers to improve feed efficiency, health and welfare; however, classic methods to record DMI are expensive and time-consuming (Davison et al., 2021). Thus, since the last years new technologies, such as cameras and Deep Learning (DL) algorithms have been developed to capture individual animal performance in an easier and cheaper way. The aim of this work is to evaluate the capability of an RGB-D camera coupled to a DL algorithm to predict individual DMI of beef cattle. For this, 12 young Charolais bulls (581 ± 37) were recorded during 10 days after meal distribution though an Intel Realsense (TMD455) camera (12 mins per video at 5 frames per second) and their DMI was individually measured. The DL algorithm for video processing was the YOLO (V8). We first trained the model to characterize activity of the animals (biting, chewing, visiting, ruminating and others; Table 1) through manual labelling using 650 images randomly extracted form videos. The images were divided into three sets for training, validation, and testing, (60, 20, and 20%, respectively) and metrics used to evaluate the model were Precision (P), Recall (R) and the average precision (AP). Then, we estimated the ingestion time (IT) by summing predicted biting and chewing time. This predicted IT was compared with the real IT measured through a chronometer. Finally, a mixed effects model was used to predict individual DMI from IT. Results showed an average high precision for all activities predicted (P = 0.62, R = 0.83, AP = 0.81). In addition, activities encompassed on the IT also showed a high precision (AP > 0.84). However, the confusion matrix showed a slight confusion between chewing and visiting (r = 0.14) activities. After processing activity of the animal, we observed a high correlation (r = 0.82, P = 0.001) between predicted and measured IT. Finally, we obtained a significant relationship between DMI and predicted IT (R² = 0.45, P = 0.001, RMSE = ±0.18kg DMI). The present work shows how new technologies such as RGB-D cameras associated to a trained DL algorithm could facilitate data acquisition on cattle farms, which may help decision-making of producers in terms of efficiency, health and welfare of the animal.