Automatic labeling of the Training Dataset for Individual and Group Activities Detection
Sepehr Honarparvar, Mahnoush Mohammadi Jahromi, Sara Saeedi, Steve H. L. Liang · 2023
Real-time activity detection is one of the interesting fields of research due to its wide range of applications in human-computer interaction, health monitoring, security, workers' performance monitoring, safety, etc. Computer vision and deep learning is one of the most popular solutions to detect human activities. However, these solutions often suffer from inadequate training datasets in terms of quantity and quality. Additionally, the procedure of data labeling is time-consuming and expensive. This paper proposes an automated method for labeling data for automatic individual and group activity detection. The proposed algorithm takes advantage of combining simple object classes (such as hands, and persons) and spatial relationship rules to label the training dataset for detecting activities like covering coughs with an arm or hand, hugging, and hand-shaking. These activities were used for Covid risky activities monitoring at workplaces. To implement this algorithm, we used You Only Look Once (YOLO) and Mask Reginal Convolutional Neural Network (Mask-RCNN) object detection and we applied the transfer learning from the Common Object in Context (COCO) CNN model. The results showed 96% precision of data labeling for covering coughs with a hand, 93% for covering coughs with an arm, 89% for hugging, and 87% for handshaking activities. This is while, the average precision of activity detection using automatically annotated data, only decreased by 3% for Mask RCNN and 5% for YOLO. Also, the separation of the individual and group activity classes led to more detection precision. In conclusion, the proposed automatic data labeling algorithm preserved the quality of data annotation while outperforming the training-data annotation speed.