Real-time image processing for productivity tracking

The production planning manager could not keep track of real-time data in the production line. They could not identify the issue until a few days later. The problem in the production line can make products produce less or more than demand. It makes the company waste time and money. The worker counted the shoe without any timestamp and noted it on the paper note, which is unmanageable and easy to lose or damage. This proposal proposes shoe counting with a timestamp to help work count products automatically. We use python programming language and OpenCV2, object training by YOLOv4-tiny, to make the system recognize the object “shoe.” Then, the detected shoes from camera streaming will count by using a 2-point intersection of each center point algorithm. There are 106 images selected among 40 shoe models to train the approach, and then images will separate into 80 training images, 16 validation images, and 10 testing images. The 10-fold algorithms are applied to a 10-times shuffle of training, validation, and testing images. This application evaluates counting and timestamp accuracy by comparing manual and system counting reports. After evaluation, our system achieves 99 percent of shoe counting, then 80 percent of timestamp extraction, and it can correctly detect and count another untrained shoe model. This approach reveals that our proposed method is suitable for counting objects in real time.