关键词: 秦岭金丝猴, 行为识别, 视觉-语言模型, 行为学语义, 描述符自适应匹配

Abstract

Aims: The Qinling golden snub-nosed monkey is a rare primate endemic to China. Accurate recognition of its behaviors is important for understanding population activity patterns, assessing habitat conditions, and supporting conservation management. However, fine-grained behavior recognition in wild infrared-triggered camera videos remains challenging because some behaviors show high visual similarity and key discriminative cues are easily weakened by camera viewpoint changes and body posture variations. 

Methods: To address these challenges, we proposed a behavior recognition method that integrates ethological semantics and descriptor-adaptive matching. First, based on the ActionCLIP vision-language framework, we constructed ethological descriptors for Qinling golden snub-nosed monkey behaviors, introducing prior knowledge such as action patterns, movement modes, and individual interactions into the textual semantic space. Second, we designed a bilateral semantic alignment mechanism, in which video features are jointly supervised by ActionCLIP category prompt texts and ethological descriptor prototypes. Finally, we introduced a dual-source descriptor-adaptive matching strategy based on optimal transport to dynamically model the correspondence between frame-level features and textual units. 

Results: Experiments were conducted on a Qinling golden snub-nosed monkey behavior dataset containing 11 behavior categories and 2,177 videos. The proposed method achieved a Top-1 Accuracy of 78.47% in the fully supervised setting, improving the baseline by 5.88 percentage points. Under the few-shot K=16 setting, the method achieved a Top-1 accuracy of 66.01%, improving the baseline by 7.25 percentage points. 

Conclusion: The results demonstrate that ethological semantics, bilateral semantic alignment, and dual-source descriptor-adaptive matching can effectively improve fine-grained behavior recognition of Qinling golden snub-nosed monkeys. The proposed method provides an effective approach for automatic behavior recognition of rare primates in wild environments.

Key words: Qinling golden snub-nosed monkey, behavior recognition, vision-language model, ethological semantics, descriptor-adaptive matching.