An automated video action recognition-based system for drinking time estimation of individual broilers

Abstract

Investigating broilers drinking behavior can provide vital information for farmers regarding broilers’ well-being, insights into farm management, and poultry resources to prepare optimal conditions in poultry farms. Therefore, the current study's objective was to assess individual broilers' drinking time using a three-step algorithm based on behavior recognition. First, the proposed algorithm detects the broilers using the U-Net architecture with ResNet50 as the encoder block. Second, the tracking algorithm based on Euclidean distance tracks the detected broilers. Third, if the algorithm recognizes that the tracked broiler satisfies the three-step pattern of drinking behavior, then a 3D convolutional network would estimate the drinking time of the tracked broiler. The proposed algorithm attained an accuracy level of 83.21 % and 88.32 % for the comprehensive estimation of drinking time and the assessment of individual broiler drinking time during each visit to the drinker, respectively. These values prove the efficiency of the proposed algorithm as a real-time and reliable tool in poultry farms.

Keywords

• Drinking time
• Broiler
• U-Net architecture
• MoviNet model
• Three-step drinking behavior pattern