In this research, we developed a two-stage model that integrates first-person video and third-person skeletal information to perform task classification and task time estimation in a textile dyeing process. In the first stage, Visual Geometry Group 16 (VGG16) recognizes task-related objects, whereas in the second stage, Spatial Temporal Graph Convolutional Network (ST-GCN) classifies detailed task actions. Through the incorporation of additional training data for the latter stage, we have attained enhanced classification accuracy across all categories, with particularly noteworthy advancements observed in tasks characterized by ambiguous boundaries. Furthermore, by automatically detecting task start and end points from the predicted action label sequences, we estimated task durations and confirmed reductions in estimation error through noncontact sensing using cameras as optical sensors combined with advanced action recognition technologies. Moreover, by calculating the mean and standard deviation of task durations for each task, we were able to evaluate process bottlenecks and the stability of each task. The proposed method demonstrates the feasibility of achieving task analysis and time estimation through noncontact, camera-based measurement without disrupting on-site operations. This approach offers a promising framework for facilitating process optimization and task standardization in manufacturing environments.