Enhancing Captive Welfare Management with Deep Learning: Video-Based Detection of Gibbon Behaviors Using YOWOvG.

Abstract

Accurate monitoring of animal behavior is critical for assessing welfare and informing conservation strategies for vulnerable species like the eastern hoolock gibbon (Hoolock leuconedys). To overcome limitations of manual observation and single-frame analysis in captive settings, this study developed the first human-annotated spatiotemporal behavior dataset for this species and proposed YOWOvG, an improved deep learning model integrating the SE attention mechanism and GELAN for enhanced feature extraction. Trained on 69,919 labeled frames across four behaviors (Resting, Socializing, Climbing, Walking), YOWOvG achieved an 85.20% Frame-mAP in video-based recognition. This is a 6.3% improvement over the baseline result while maintaining computational efficiency. The model effectively captured temporal dynamics and spatial contexts, significantly improving recognition of climbing and walking despite data imbalances. The results demonstrate the potential of automated, noninvasive video monitoring to enhance welfare assessment in rescue centers by detecting subtle behavioral changes. Future work will expand behavioral categories, address stereotypic behaviors, and integrate audio cues for holistic monitoring. This approach provides a scalable framework for behavior-informed management of captive wildlife.