| Popis: |
The task of human pose estimation (HPE) involves precisely locating body parts and constructing an accurate representation of the human body using unstructured input data such as images. HPE is a challenging problem, particularly when considering real-world applications such as aquatic safety, where data may be scarce. The domain of CCTV images of public swimming pools poses a severe challenge, as water can cause extreme distortions and occlusions that can significantly impact the visibility of each swimmer. This Master's thesis addresses this challenge. This study uses various methods to estimate the domain gap between the COCO keypoint dataset and the proprietary Lynxight 2D HPE dataset. Further, it develops a training regime using keypoint-specific initialization and aggressive data augmentation to bridge the domain gap between the two datasets. The findings demonstrate that combining multiple techniques can effectively bridge the gap and improve the accuracy of HPE in aquatic safety applications. In addition, the dynamics of multi-task learning with novel and domain-specific tasks are explored for improving HPE. The experiments show that multi-task learning can further enhance the performance of HPE. Overall, this study provides valuable insights into the challenges and opportunities of HPE in aquatic safety applications and demonstrates practical approaches for overcoming these challenges through domain adaptation and multi-task learning. In addition, these findings can have significant implications for the development of downstream tasks such as behavior recognition, drowning, and anomaly detection. |
