Research

My current research explores the use of 3D human pose estimation for behavior analysis in autism diagnosis. By examining social situations and core ASD symptoms, such as restricted and repetitive behaviors, the 3D location of human joints can be analyzed to detect atypical patterns. This information serves as a strong indicator of ASD symptoms, enabling more accurate diagnoses. 

This research introduces an accurate and efficient Deep Learning model for Multi-Person Pose Estimation, optimized for Edge TPU devices. Focusing on rehabilitation centers and individuals with knee injuries, DeepRehab predicts 23 body keypoints to analyze exercises performed during the rehabilitation process. Findings from this study were published in the International Conference on Artificial Neural Networks.

This research in home rehabilitation scenarios focuses on a Composite AI approach, combining body pose estimation, dialogue systems, semantic segmentation, and rule-based systems. To achieve high-precision body pose estimation, three components were identified: 1) a 3D environment representation, 2) a navigation dialogue guiding patients to optimal poses, and 3) semantic and instance maps for verbal instructions. The findings were published in the Multimodal Technologies and Interaction journal by the Multidisciplinary Digital Publishing Institute. 

This thesis focuses on developing a Deep Learning-based eye-tracking system to control wheelchairs using only the eyes, for individuals with limited hand movement or paralysis. The proposed two-stage pipeline involves segmenting human eyes and classifying gaze direction. An Arduino prototype was used for testing.