Limited Data-Oriented Worker Intention Recognition Method in Worker-Robot Collaboration for Construction

Timely and accurate identification of workers' intentions in construction scenarios is crucial for seamless worker-robot collaboration. However, limited worker behavior due to varying behavioral styles and difficulties in collecting worker action data limit the practical application of existing methods that rely heavily on extensive worker action data. This paper addresses the dynamic nature of construction environments by proposing a few-shot worker intention recognition method. The proposed approach constructs worker intention query features using randomly sampled frame combinations and then applies metric learning to develop a few-shot worker intention recognition model. To validate the effectiveness of this method, a worker scaffolding installation action video dataset was used for the experiments on worker intent recognition. Given five categories with five worker action samples, the method achieved an accuracy of 71% in recognizing workers' intentions. The results demonstrate that the proposed method can effectively learn and detect novel worker actions with a minimal number of classified action videos, thereby improving model performance while reducing the number of required training videos. This approach not only reduces the labor needed for data labeling but also enhances the practicality of worker-robot collaboration in construction scenarios.