13:30
17:30

Kévin Reby thesis: Human pose analysis using deep learning

The convergence of the health and digital sectors has led to the development of digital therapeutics (DTx). This work was done in the context of the development of a DTx for pain management. Pain is an important issue facing individuals, families, healthcare providers and society as a whole. Pain is usually measured by patient self-report or clinician impressions, either through clinical interviews or through visual and numerical pain scales. Although useful, self-reported pain cannot be used with young children, patients with certain neurological or psychiatric impairments, or with many patients in post-operative care or in transient states of consciousness. Recent advances in computer vision and machine learning for automatic analysis and modelling of human behaviour could play a key role in overcoming some of the limitations in a clinical context. Therefore, the automatic and objective assessment of pain disorders from behavioural signals is of increasing interest to clinicians and computer scientists.

The development and success of Deep Learning-based methods in the field of computer vision has led to significant advances in the field of automatic pain recognition. Most studies have focused on facial pain detection, while little research has been conducted on human pose.

In this work, we propose to use a two-stream spatio-temporal convolutional network (ST-GCN) model and spatial and temporal attention mechanisms to evaluate a person's behaviour based on their posture. We tested our model on UI-PRMD, a reference dataset that provides skeletal data using motion capture systems. Our results show that our attention-based ST-GCN models outperform state-of-the-art methods for quality score prediction and binary classification.

Amphi E (Bât A29)