Today, the technological paradigm represented by the 4th industrial
revolution has made it possible to effectively manage and analyze various big
data generated through people and Internet of Things (IoT) terminals. In
addition, the outputs were able ...
Today, the technological paradigm represented by the 4th industrial
revolution has made it possible to effectively manage and analyze various big
data generated through people and Internet of Things (IoT) terminals. In
addition, the outputs were able to converge into new values beyond all areas
of society. Among them, artificial intelligence (AI) successfully solves
problems that were difficult to solve with traditional computer algorithms, and
has established itself as the most core technology leading the 4th industrial
revolution.
In general, artificial intelligence learns how to represent the data by
appropriately transforming the feature space of the dataset. In order to
acquire an artificial intelligence model with adequate performance, data points
that can represent the entire data set well and are not biased must be
collected and selected and processed appropriately so that the machine can
understand them well. For this reason, preparing a training dataset is
considered so important that 80% of the entire project's time resources are
invested.
Data labeling is a kind of data preprocessing task that annotates data
so that machines can understand the data, which takes a long time and
requires a lot of human labor. The data labeling task does not require
high-level technology, but it is a very important starting point as a
preparatory task for learning artificial intelligence. However, most of the data
labeling task is currently carried out by people's labor, so it takes a lot of
time. In addition, consistency is poor due to differences in labeling skills
depending on the worker, which is emerging as an issue that negatively
affects learning process of the artificial intelligence model.
Therefore, this paper proposes a system that automatically labels pose
estimation and object tracking information focusing on action recognition to
increase the task efficiency of data labeling that requires a large number of
manpower and time in the artificial intelligence research process. The
proposed system labels object pose and position automatically using models;
YOLOv7-Pose and StrongSort.
The implemented automatic labeling system achieved a F1-score of 0.849,
and showed better performance on indoor images rather than outdoor. These
results show that, except for specific situations such as object occlusion, the
performance of the automatic labeling system was very similar to the results
of manual task by workers, and it can significantly reduce task time.