Building Operation is recognized as a key factor influencing energy consumption. Effective methods to monitor occupant states and meet their comfort requirements are essential for enhancing building energy efficiency. Previous studies have encountered challenges in accurately monitoring and distinguishing between different activities. In this work, we present an innovative computer vision-based thermal environment regulation system that is optimized based on the Openpose framework to implement state recognition of occupants, which aims at giving demand-driven optimization feedback according to their corresponding characteristics. Experiments show that the proposed optimization algorithm exhibits high accuracy in all stages of personnel behavior detection, improving the accuracy by at least 16.67% over the unoptimized algorithm.