This thesis develops and validates deep learning pipelines for poultry behavior analysis using smartphone-captured video to enhance precision livestock farming. An integrated framework is introduced, combining YOLOv11-Pose detection, ByteTrack multi-object tracking, and a spatiotemporal X3D video encoder fused with pose-derived metadata via a multilayer perceptron. The system achieved 87.08% accuracy across five welfare-indicating behaviors. This validated framework was then applied to analyze two genetic lines of broiler chickens (Ross 308 and African local ecotype) across different times of day. The analysis revealed clear genetic and temporal differences in activity, resting, and thermoregulatory behaviors. Collectively, this work demonstrates that artificial intelligence with affordable computer vision tools can provide scalable, objective, and noninvasive solutions for welfare monitoring and management in poultry production.
