Model selection for 24/7 pig position and posture detection by 2D camera imaging and deep learning

• Published in: Computers and electronics in agriculture 2021-08, Vol.187, p.106213, Article 106213
• Main Authors: Riekert, Martin, Opderbeck, Svenja, Wild, Andrea, Gallmann, Eva
• Format: Article
• Language: English
• Subjects: 2D vision cameras; Annotations; Datasets; Deep learning; Image degradation; Image quality; Industrial applications; Infrared imagery; Model selection; Night; Night near-infrared recordings; Pig lying behavior; Two dimensional models
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2021.106213

•Deep learning system for continuous 24/7 pig position and posture detection.•Selected deep learning model based on over 150 different configurations.•Dataset consists of 13,047 human-made annotations of 18 pens and 21 cameras.•Validation and test images covered 24/7 video recordings for 6 months from 10 pens.•Detected position and posture of 84% mAP@50 for day and 58% mAP@50 night recordings. Continuous monitoring of pig posture is important for better understanding animal behavior. Previous studies focused on day recordings and did not investigate how deep learning models could be applied during longer periods including night recordings under near-infrared light from several pens. Therefore, the objective of this research was to study how a suitable deep learning model for continuous 24/7 pig posture detection could be achieved. We selected a deep learning model from over 150 different model configurations covering experiments concerning 3 detection heads, 4 base networks, 5 transfer datasets and 12 data augmentations. For this purpose, we test and validate our models using 4690 annotations of randomly drawn images from 24/7 video recordings covering 2 fattening periods from 10 pens. Our results indicate that pig position and posture was detected on the test set with 84% mAP@0.50 (49% mAP@[0.50:0.05:0.95]) for day recordings and for night recordings 58% mAP@0.50 (29% mAP@[0.50:0.05:0.95]) was achieved. The main reason for lower mAP during night recordings was degraded near-infrared image quality. Our work reports important findings concerning the applicability of deep learning models on night near-infrared recordings for posture detection. The dataset is publicly available for further research and industrial applications.