Filtering-Based Segmentation of Overlapping Food Items: A Case Study in Automated Chicken Breast Handling

This paper introduces a new method for separating objects with a high overlapping area. Standard segmentation methods often face challenges when addressing overlapping objects, particularly in scenarios where border contrast is low. For example, algorithms may mistakenly interpret a cluster of chicken breasts as a single entity, potentially leading to problems in packaging, counting, and size estimation. This study focuses on improving the accuracy of identifying the center, size, and orientation of each individual chicken breast within an image. The process begins with color detection to effectively distinguish the objects from the background, followed by OpenCV contour detection to identify their boundaries. To prevent the merging of distinct objects, even in cases of high overlap, we introduce a custom filling algorithm designed to maintain separation with improved precision. In the final stages of our process, we refine the shape of the objects using the watershed algorithm and apply classical geometric methods to determine the orientation of each breast. This enables the position observed by the camera to be accurately projected onto the robot frame for precise handling. It is anticipated that this approach could also be adapted for a range of other computer vision applications. This efficient, real-time approach shows promising results and has potential applications in other industrial computer vision tasks involving overlapping, deformable objects.