Production Line Monitoring
The topic is available also in Hungarian.
The first step to optimize a manufacturing process is to be informed about the exact time needed for each step of the process, in other words, to monitor the process. The goal is to develop a minimally invasive solution, which does not affect the daily work. To accomplish this task, a camera based solution is proposed. The camera is connected to, e.g., a Raspberry PI, the PI is mounted over the manufacturing stand, so that it has a clear picture of the product being processed.
In the frames of this project, we focus on computer vision based monitoring. It means that we would like to identify, e.g., the timestamp when a certain part gets mounted on the product. The pictures of the parts to detect and the pictures of the product are available. The proposed solutions are template matching, feature matching and deep learning approaches. It means that knowing which parts are to be detected, a matching algorithm is run periodically. The fact that a part is mounted is detected by a template match over a certain threshold.
The goal of this project is to make the template matching algorithm more robust in order to reduce the false positive rate.
The task of the student is to develop and evaluate computer vision networks with regard to high quality template matching. The plan is to extract the lower layers of published computer vision networks (ResNet, AlexNet, VGG), to basically conduct transfer learning. However, the student can propose alternate techniques.