Application of long-short term memories (LSTM) for pose recognition in quality control station

The time-motion study aims to find the standard time and improve production efficiency in manufacturing. This paper applies the Deep Learning model to recognize worker motion elements to have an automatic and effective system to improve productivity. Human pose and motion recognition are used to reduce production management costs. Each pose's training and testing video dataset is collected from the shoe manufacturing firm's camera at the specific Quality Control station. We propose applying the Long-Short Term Memories (LSTM) model, one of the particular kinds of RNN, with a kinematic base. Mediapipe Pose is used to express the body poses under the kinematic type, representing the human body's shape before feeding it into the LSTM model to train. The experiment conducted three work elements: pickup, inspection, and storage. After combining the LSTM model with Mediapipe Pose, we can detect the labels of workers' activities through video-based input and promising application to real-time cameras. The intersection rule is proposed in this study to improve the "flickering effect" to improve the accuracy up to 99.88%, 94.86%, and 100% for the inspection, pickup, and storage actions, respectively.