DEEP LEARNING FOR CLINICAL POSE ESTIMATION: A TOP DOWN APPROACH

Clinical pose estimation involves the localization of keypoints, joints or landmarks of clinical interests in an image. The results thereof are further processed for applications such as clinical gait analysis, patient rehabilitation, and assessment of human posture. This work is a subcategory within a larger framework - an end-to-end markerless motion capture system. The focus of the thesis is on the core part of the motion capture system, and it entails obtaining image signals from recordings of subjects walk cycles, proper annotations of clinically relevant anatomical landmarks and using a top-down pose estimation approach for the localization of the keypoints. Top-down pose estimation approach involves using two models. The first model is the person detection model, it is used to obtain the bounding box of the subjects in the images. The result is then passed on to the second model for key-points localization. For clinical and medical purposes, precision is very important and to achieve this, a dataset that captures the relevant details is important. The improvement this work has over existing approaches is the use of custom dataset that follows the IOR gait protocol for anatomical landmarks. IORgait is a gait analysis protocol and report for fundamental examination in clinical practice. As part of our optimization of the exisitng pose estimation framework, we proposed a replacement of the person detection model with the faster and more accurate YOLOv3 person detector. The improvement in prediction time of people's bounding-box by our replacement ensures that the system can be used in real-time. The evaluation metrics adopted in this work show promising results, although more work needs to be done in obtaining high quality datasets for a future deployment of our pose estimator in clinical applications.