Intelligent Perception System of Moving Robot for Dynamic Environment

Object detection poses a significant challenge in the realm of computer vision for mobile robot grasping tasks. The rapid advancements in deep learning have led to the emergence of numerous novel object detection algorithms, resulting in improved accuracy in large-scale environments. Mobile robot grasping tasks demand highly responsive target detection, as any delay could hinder the robot's ability to react appropriately, thus impacting task execution. Additionally, the constantly changing background environment caused by a fast-moving robot introduces dynamic interference. To address these issues of real-time processing and dynamic background interference in mobile robot grasping tasks, it is crucial to enhance target detection speed, comprehensively capture features, and interpret image context information. This paper introduces a new information fusion structure based on the Yolov8 model and Gold-yolo aggregation distribution mechanism. The sampling approach retains the traditional Faster Implementation of CSP Bottleneck with 2 convolutions module (C2f) from YOLOv8 while incorporating the variable convolution module from Data Communication Network (DCN2). These two sampling methods are interwoven, enhancing the accuracy of detecting various shape information of the same object as the robot moves.