Inferring Brown Bear Hair Snare Interactions by Automatically Detecting Bipedalism on Camera Trap Images Using Pose Estimation and a Multilayer Perceptron

Abstract

Hair snares, consisting of strands of barbed wire that collect tufts of fur, have long been used as a noninvasive sampling technique for DNA‐based monitoring of different species of bears. However, to prevent cross‐contamination by other animals or different bear individuals, frequent visits by technicians are required to collect samples and clean snares, a highly demanding process in terms of time and resources that hampers scalability. In the Catalan Pyrenees, only 24.6% of the visiting efforts come out positive, highlighting the need for automated, scalable solutions that can operate independently of cellular networks or human intervention. This study proposes a novel methodology to automatically infer hair snare interactions by detecting bipedalism in images acquired by camera traps. This is implemented by using a state‐of‐the‐art deep learning model for pose estimation, combined with a multilayer perceptron (MLP) to classify bipedalism or quadrupedy based on predicted keypoints. Due to the lack of annotated bear pose datasets, a custom dataset of 2373 images collected in the Catalan Pyrenees was manually annotated with 15 anatomical keypoints. Using YOLOv11, the trained pose estimation model achieved a keypoint inference precision of 93.2%, while the MLP reached an accuracy of 96.1% in distinguishing bipedalism from quadrupedy. Finally, to enable inference on the edge, several verification strategies are proposed using only the pose estimation output, including bear geometrical illustrations derived from the inferred keypoints. This innovative lightweight solution, designed for satellite‐based transmission, enables scalable deployment in low‐connectivity and remote areas, significantly reducing the need for frequent manual inspections.