A wild animal behavior recognition method based on multi-task learning
By employing a multi-task learning method involving progressive training and gradient coordination, a dual-channel spatiotemporal feature extraction network and a main-auxiliary dual-branch output network were constructed. This solved the real-time and stability issues in wildlife behavior recognition, enabling efficient monitoring of fine-grained behavior recognition.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INST OF SEMICONDUCTORS - CHINESE ACAD OF SCI
- Filing Date
- 2026-03-12
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies struggle to achieve real-time, stable, fine-grained behavioral understanding and cross-scene generalization in wildlife behavior recognition. This is especially true in bird monitoring, where targets are small-scale, frequently occluded, subject to large changes in lighting, and have strong background interference, leading to frequent missed and false detections in the model. Furthermore, gradient conflicts in multi-task learning cause training instability.
A progressively trained multi-task spatiotemporal action recognition model is adopted, which constructs a dual-channel spatiotemporal feature extraction network and a main-auxiliary dual-branch output network. The weights of the auxiliary task are adjusted by gradient cosine similarity detection to achieve joint learning of behavior category recognition and motion attribute prediction, thereby alleviating gradient conflict and improving recognition robustness.
It improves the accuracy and robustness of wildlife behavior identification, reduces processing time, is suitable for large-scale, long-term bird behavior monitoring, and significantly improves identification efficiency and stability.
Smart Images
Figure CN122135404A_ABST