| Abstract [eng] |
This doctoral dissertation examines effective approaches for detecting and analysing human posture from occluded and limited video data, with particular attention to challenging real-world conditions such as occlusion, low image resolution, and constrained visibility, which commonly affect video-based monitoring environments. These challenges frequently arise in healthcare monitoring and intelligent digital systems and often reduce the reliability of conventional posture recognition methods. To address these issues, the dissertation developed optimized deep learning models based on hyperparameter-tuned transfer learning, data augmentation, and lightweight convolutional neural network architectures. Hybrid models that combine deep learning feature extraction with machine learning classifiers are also introduced to improve accuracy, generalization and computational efficiency. The proposed models are extensively evaluated on benchmark and real-world datasets, consistently demonstrating improved detection performance compared to existing conventional deep learning models. The results demonstrated that the proposed models achieved higher accuracy, robustness, and efficiency while remaining suitable for deployment in resource-constrained healthcare and smart monitoring systems. The dissertation contributes novel methodological insights and practical solutions for posture detection in digital health application. |
