Design of Obstacle Recognition and Localization System for Intelligent AGV

With the development of industrial automation, Automated Guided Vehicles (AGVs) have become increasingly widely used in logistics transportation. Their safe operation relies on efficient obstacle recognition and localization capabilities. Aiming at the problems of insufficient precision of single sensors and difficulties in multi-source information fusion in traditional AGV environment perception systems, this paper designs an obstacle recognition and localization system based on binocular vision and deep learning. The system uses the YOLOv8 network to perform target segmentation and recognition on images captured by the ZED2 binocular camera, identifying four common obstacles: box, person, fire extinguisher, and trash can. Three-dimensional position estimation of obstacles is achieved through feature point matching and stereovision principles, and a lidar is used to verify the localization results. To improve the model’s generalization ability under different imaging conditions, a hybrid dataset comprising mobile phone images and binocular camera images under natural daylight and artificial night lighting is constructed for training. Experimental results show that the system maintains a recognition accuracy of over $99 \%$ in complex lighting environments, with obstacle localization errors controlled within $\pm 5 \mathrm{~cm}$, verifying the system’s robustness and practicality.