Efficient Object Detection with Color-Based Point Prompts for Densely Packed Scenarios in WRS FCSC 2024

This paper introduces a novel object detection method designed for densely packed environments, such as those encountered in the World Robot Summit Future Convenience Store Contest (FCSC) 2024.Our system leverages color-based point prompts in conjunction with Segment Anything (SAM) 2 to achieve precise object segmentation and grasp point estimation, specifically targeting scenarios where objects like rice ball cluster tightly within containers.Unlike traditional methods that depend on pre-defined grasping strategies susceptible to positional drift in mobile robots, our approach dynamically identifies and isolates objects without requiring extensive retraining inherent to CNN or Transformer models.We conducted comprehensive experiments comparing our method against SAM 2 and Grounding DINO using a dataset of 10 test images containing 202 rice balls.Additionally, we deployed the system on a Toyota Human Support Robot during the FCSC Stock Task to assess real-world performance metrics.Results demonstrate that our method achieves higher detection accuracy and operational efficiency, validating its potential for autonomous retail applications.