In Japan, the construction industry strongly be needed productivity improvement and increasing the number of new hires due to improvement of working environment. Site manager needs to grasp whether the daily progress is as planned and updates the schedule appropriately for improve site's productivity and safety. In image-based data acquisition approach in japan, there is a problem that learning is insufficient with only global public data, since construction worker in Japan has originality in image feature compare with other countries. In this study, we make original dataset for additional learning firstly. Then we proposed domain-specific algorithms specific to the Japan construction site, including a worker detection and tracking algorithm and a worker action recognition algorithm. As a result, our worker detection showed 87.9% accuracy in same-site evaluation and 77.5% accuracy in cross-site evaluation. Our worker action recognition showed 60.2% mean accuracy. Finally, the method of translation into activity element based on the output value of worker detection was indicated.