A Comparative Study of Environmental Perception and Object Recognition Methods for Robots

• Published in: 2024 3rd International Conference on Robotics, Artificial Intelligence and Intelligent Control (RAIIC) pp. 73 - 77
• Main Authors: Fan, Haoyu, Zhang, Guowei, Zhou, Youtao, Hu, Jinwen, Xiao, Yuntao, Zhang, Shimian
• Format: Conference Proceeding
• Language: English
• Published: IEEE 05.07.2024
• Subjects: Accuracy; Analytical models; component: Robotic Object Recognition; Grounding; Image segmentation; Object detection; Object recognition; Open-set object detection models; Real-time systems; SAM
• DOI: 10.1109/RAIIC61787.2024.10671088

To address the selection of generic image segmentation and open-set object detection models for robotic object recognition, this paper experimentally compares the generic image segmentation and open-set object detection models, specifically, compares and analyzes the performance of the SAM (Segment anything Model) image segmentation model as well as its derivatives Mobile SAM and HQ-SAM generic image segmentation model, and the Grounding DINO and YOLO-world open set object detection models. The applicability of these state-of-the-art object recognition methods in different robotic environment perception tasks is fully validated through detailed experimental evaluations. The results show that Mobile SAM and YOLO-world perform well in real-time demanding scenarios and can meet the needs of fast response and processing, while HQ-SAM and Grounding DINO show superior performance in segmentation or detection tasks requiring high accuracy and are more suitable for application scenarios requiring high segmentation or detection accuracy. These findings provide references and guidance for the practical applications of robotic object recognition, and can help researchers and engineers to choose the most suitable models and methods in scenarios with different requirements.