Carlos Eduardo Bezerra, Cauê Bittencourt, Renalvo Alves, Thiago Ribeiro, María Azorín-Ortuño, Gloria Nejar, Alba Barreras, Claudine Zemirline, Thomas Lewiner, Anahita Le Bourdiec, Gaëlle Pagny, Adriano Barbosa, Xu Yang, Krerley Oliveira, Thales Vieira
In image-based evaluation, the lightweight neural networks achieved 92.6% accuracy for side-view classification and 93.1% accuracy for overall posture classification while using approximately 38 times fewer trainable parameters than VGG19-based baselines. After image quality and viewpoint filtering, descriptor accuracy reached 88% for head pitch, 95% for back inclination, 90% for back curvature, and 95.47% for tail configuration. In end-to-end video experiments, frame retention after filtering depended strongly on recording conditions, particularly side-view visibility and occlusion. Nevertheless, retained frames still enabled high-accuracy estimation of head pitch, back inclination, and tail configuration.
INTRODUCTION: Body posture provides clinically relevant information in cats, but behavioral interpretation can be challenging because postural signs are often subtle, context-dependent, and subject to inter-observer variability. We developed and evaluated a non-invasive computer vision framework for extracting clinically interpretable feline posture descriptors from side-view RGB images acquired using low-cost cameras, with the aim of supporting behavioral assessment, welfare monitoring, and future clinical decision support.
METHODS: The proposed framework combines object detection, anatomical segmentation, lightweight convolutional neural networks for posture and side-view classification, pose estimation, and rule-based geometric analysis within a structured processing pipeline. From standing side-view frames, the system generates descriptors including overall posture, head pitch, back inclination, back curvature, and tail configuration.
RESULTS: In image-based evaluation, the lightweight neural networks achieved 92.6% accuracy for side-view classification and 93.1% accuracy for overall posture classification while using approximately 38 times fewer trainable parameters than VGG19-based baselines. After image quality and viewpoint filtering, descriptor accuracy reached 88% for head pitch, 95% for back inclination, 90% for back curvature, and 95.47% for tail configuration. In end-to-end video experiments, frame retention after filtering depended strongly on recording conditions, particularly side-view visibility and occlusion. Nevertheless, retained frames still enabled high-accuracy estimation of head pitch, back inclination, and tail configuration.
DISCUSSION: These findings demonstrate that automated posture analysis from standard RGB video can generate structured and clinically interpretable feline body-language descriptors relevant to behavioral medicine and animal welfare assessment. Further prospective validation against established pain, stress, and welfare assessment instruments is required before clinical or diagnostic application.