Real-time monitoring of metabolic rate (MET) and clothing insulation (CLO) is essential to ensure effective occupant-centric control (OCC) for thermal comfort. This study aims to develop multi-task model using semi-supervised learning to enhance occupant activities and clothes detection performance by utilizing both labeled and unlabeled data. The convolutional neural network-based model and training approach with pseudo labels to update all parameters comprehensively were proposed. The developed model is validated by conducting comparative analysis with state-of-the-art models and applying it in a real-world environment. The results demonstrate that the developed model, employing semi-supervised learning and the dual-phase training method (DPTM), achieves superior performance in activity and clothes detection outperforming previous studies with a 15.8 % higher mean Average Precision (mAP) for activity detection and a 25 % improvement for clothes detection. The findings highlight the potential of this multi-task model using semi-supervised learning to automate data collection improving the accuracy of estimating occupant thermal comfort. This approach can dynamically optimize indoor environments tailored to individual needs within the OCC framework, enhancing thermal comfort and energy efficiency through precise monitoring of occupant information.