Abstract:
"Tea factory packing area is a major stage in the whole tea manufacturing process. The packing area carries out tea blending and other packing activities. Factory supervisors and factory managers make major decisions on how, when and what aspects of the manufacturing process. Investing most of the times supervising areas such as the packing area takes away time, which instead could be used for other managerial decisions, from the factory supervisor or the factory manager.
The solution for the problem addressed is an Auto Supervisor system, which uses video footages to detect idle times and working times of the workers in the packing area. The solution is a step taken towards the introduction of industry 4.0 to tea factories. With the times detected of the poses, the factory supervisors and factory managers can invest the time spent on supervising on to more managerial decision making activities. The solution is an LSTM model which is trained on sequences of 30 keypoint data, which are detected by the YOLOv8 pose estimator feature. The sequential model consists of 3 LSTM layers and 3 Dropout layers one after the other. The LSTM layers take the temporal aspect of the data, which will consider a fixed sequence of frames to detect the activity. The purpose of the dropout layers were to prevent overfitting of the model.
With the custom dataset created with the specific poses, the model achieved an accuracy of 90%, identifying multiple correct poses from the manual testing of the model."