An Improved YOLOv8 Model for Obstacle Recognition and Positioning in Autonomous Driving

Based on YOLOv8 algorithm, this paper improves the loss function and adds a simple attention mechanism to study obstacle recognition and localization in autonomous driving. The improved YOLOv8 model achieved an average accuracy of 92.9% and an average recall of 88.7%, while the recognition speed is 62.7 frames per second. The SGM (Semi Global Block Matching) stereo matching algorithm is used to achieve three-dimensional spatial positioning of roadblocks. The relative error of achieved localization is within 1.1%, which can basically meet the visual system requirements for obstacle positioning in autonomous driving scenarios.