Spatio-Temporal Human Activity Recognition on Manufacturing Floors
In industry, inspection of manufacturing floors is handled by humans. The drawback of using humans for this repetitive task is that collected data can be highly prone to bias due to small sample size and industrial engineers can miss relevant details during the cycle time analysis. Calculations often misinterpret the real factory situation. This research aims at finding an automatic computer-vision-based system for manual inspection. Some industries are trying to use wearable sensors, so that they can monitor worker activities. However, such an approach is not readily scalable and can be burdensome to the worker. On the contrary, computer-vision-based activity recognition systems can function without interrupting workflow. This research attempts to find a scalable vision-based AI solution that can perform activity recognition on gateway devices. The goal is to develop a neural network architecture with low computational complexity, taking advantage of state-of-the-art computer vision and deep learning tools.