Recognition and localization method for high-speed robotic sorting in complex environments based on image interference assessment

Recognition and localization method for high-speed robotic sorting in complex environments based on image interference assessment

연구 분야: Artificial Intelligence

학회: International Journal of Intelligent Robotics and Applications

Machine vision has been widely applied in the field of high-speed sorting. However, how to address the decline in detection accuracy caused by external interference while maintaining real-time performance remains a key challenge. This work proposes a lightweight image interference evaluation network, combined with a field-of-view entry judgment module and an interval frame sampling mechanism, to effectively reduce the number of invalid inferences by the model and reduced computation time. Even under extreme conditions such as harsh lighting or object occlusion, this method can substantially mitigate the impact of external interference on target recognition and localization. The recognition accuracy of “you only look once” (YOLO) has been improved through transfer learning and anchor re-clustering. Meanwhile, a multi-point geometric localization method based on YOLO regression boxes is proposed, incorporating a penalty function to reduce the influence of distant targets, enabling precise localization of grasping points for dense objects. Finally, the reliability and effectiveness of above methods in real industrial scenarios are verified on the experimental platform.