DEVELOPMENT OF COMPUTER VISION BASED SMART SYSTEM FOR BROILER CHICKENS HEAT STRESS RECOGNITION

• Main Author: Iqbal Anggoro Agung, Muhammad
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/83925
• Institution: Institut Teknologi Bandung

In 2022, approximately 80% of the Indonesian population’s demand for protein sources comes from poultry meat, which requires farmers to optimize poultry farming practices. Broiler chickens, a commonly raised poultry breed in Indonesia, requires optimal farming processes. To ensure this, the Feed Conversion Ratio (FCR) is used as one of the parameters to determine the effectiveness of the farming process. Livestock welfare, crucial for growth rates and FCR, can be monitored through computer vision technology. This research aims to develop an intelligent behavior recognition system for broiler chickens to assist operators in monitoring chicken welfare in commercial coops using a computer vision-based Precision Livestock Farming (PLF) method. This research involved observing CP 707 strain broiler chickens by recording their activity data using temperature, humidity sensors, and cameras from 06:00 to 20:00 local time. The collected data, comprising approximately 5,000 annotated images of objects, was compiled into a dataset and split into training, validation, and testing data with a 6:2:2 ratio. This study employed the Instance Segmentation method with the Mask R-CNN model, trained over 3000 iterations using transfer learning from the MS COCO dataset model to recognize chicken segments and three objects in the coop. The best-trained model was then used as a detector in the object tracking process to extract the Cluster Index, Unrest Index, and Average Displacement features. These features were then selected to be used to build a chicken comfort model based on the Temperature Humidity Index (THI). The selected features were then trained using the Support Vector Machine (SVM) machine learning method with various kernels and parameters. The research results indicated that the best model for object segmentation was the Mask R-CNN model with a ResNet-50-FPN backbone, achieving an mAP50:95 bounding box value of 75.9%, an mAP50:95 segment value of 75.6%, and an average inference time of 79.9 ms on the testing dataset. After this model was used to compare the features to the THI value, the Cluster Index feature showed the best correlation with the THI value, followed by the Unrest Index, but the Average Displacement’s correlation could not be determined. The Cluster Index and Unrest Index features were then selected and trained to distinguish chicken comfort conditions, with a THI below 25°C considered comfortable and above that considered uncomfortable. An SVM-based model with four types of kernels was used for training, resulting in the best SVM model with a Linear kernel and C value of 1, achieving an AP value of 92.26% across all data and a weighted F1 score of 91.11%. Using the Mask R-CNN instance segmentation model, object tracking algorithm, and chicken comfort recognition model, a prototype application was developed to help operators monitor broiler chicken conditions by visualizing tracking and object segmentation results, and succesfuly send alarm notifications for the operators via Telegram bot. Keywords: Machine Learning, Computer Vision, Mask R-CNN, Object Tracking, Broiler Behavior Recognition, Support Vector Machine (SVM). ?