Spatial Prior-Enhanced Temporal Action Localization: A Novel Three-Stage Framework for Pig Behavior Recognition

• Three-stage cascade framework integrates object detection, feature extraction, and temporal localization to capture fine-grained pig behaviors in thermal videos. • Spatial prior integration via YOLOv10-detected excretion sites effectively resolves background noise interference for subtle actions, boosting defecation mAP by 12.6%. • ActionMamba network utilizing spatially enhanced features achieves 84.19% mean average precision, significantly outperforming conventional global feature baselines. • DETAD diagnostic analysis confirms that incorporating spatial priors significantly reduces localization errors and background false positives in complex thermal environments. Automatic animal behavior recognition is a crucial component of Precision Livestock Farming. To address the challenges where traditional temporal action localization models struggle to accurately identify behaviors such as excretion in thermal infrared videos of pigsties due to their short duration and subtle motion, this study proposes a three-stage cascaded architecture comprising "object detection, feature extraction, and temporal localization." This architecture is designed to achieve precise recognition and duration localization of key pig behaviors, particularly urination and defecation. This study constructed datasets for object detection and behavior recognition based on thermal infrared images and videos, comprising five detection targets and five behavior categories. YOLOv10 was employed to achieve high-precision detection of excretions and pig postures, thereby obtaining spatial priors for behaviors. Subsequently, a video feature extraction model based on a masked autoencoder mechanism was utilized for spatiotemporal dynamic representation. These spatially enhanced features were then input into the ActionMamba network to enable accurate behavior classification and precise localization of action start and end boundaries. Additionally, the DETAD tool was utilized to conduct analyses on false positives, false negatives, and sensitivity, facilitating a fine-grained diagnosis of model performance. The results demonstrate that the YOLOv10 object detection algorithm achieved an mAP@50 of 91.30% for recognizing feces, urine, and pig bodies on the validation set. The spatial aggregation results revealed that the predicted regions were highly consistent with the distribution of the pigs' natural excretion and resting areas. Regarding the ActionMamba-based temporal action localization algorithm, under three strategies—using features pre-trained on a public dataset, features trained on the in-house pig behavior dataset, and enhanced features integrating spatial priors from excretion detection—the average mAP@0.5 for behavior recognition reached 68.06%, 82.02%, and 84.19%, respectively. In terms of fine-grained behavior categories, integrating spatial priors increased the mAP@0.5 for urination behavior from 91.13% to 92.06%, and significantly improved the mAP@0.5 for defecation behavior from 65.82% to 78.42%. DETAD evaluation indicates that spatial priors effectively reduce localization errors and background false positives, thereby enhancing the model's boundary localization capability for short-duration and subtle actions. In summary, the three-stage cascaded architecture proposed in this study effectively improves the accuracy and stability of fine-grained pig behavior recognition in thermal infrared videos, providing reliable technical support for herd health monitoring and behavior analysis in precision livestock farming.