Evaluating Computer Vision Solutions for Automated Medicine Detection Applied to Grasping

• T. J. A. De Souza UFPE
• J. P. M. De Paula UFPE
• J. M. X. N. Teixeira UFPE
• A. Durand-Petiteville UFPE

Resumo

The process of creating a robust autonomous medication dispensing system relies on the system’s capability to collect data from the environment and accurately calculate the medication’s presence and position. Taking cylindrical ampoules as the test subjects, this paper evaluates two approaches to this problem based on the data acquired from an RGB-D camera. For the first approach, a dataset was collected and a YOLOv8 model was trained to identify medicine ampoules. As for the second one, the depth image was processed using the K-means algorithm. Both methods perform well in the experiments, identifying all 347 ampoules each. The experiments also revealed that while YOLOv8 is more accurate, the K-means method can represent a more flexible solution in unexpected scenarios.

Palavras-chave: Training, Computer vision, Accuracy, Costs, Image color analysis, Robot vision systems, Grasping, Cameras, End effectors, Biomedical imaging, Robotics, Computer Vision, Healthcare, Detection, YOLO, K-means