A study on real-time semantic segmentation algorithms in autonomous driving environments based on YOLOv8
Abstract
This study proposes a real-time semantic segmentation algorithm based on an improved YOLOv8 network architecture, aiming to address the balance between inference time and pixel-level classification accuracy in edge perception computing for autonomous driving. By integrating a channel attention module into the backbone structure for feature weight redistribution and designing a lightweight decoupled mask generation branch in the decoding segment, this method achieves matrix dimensionality reduction for redundant spatial parameters. Experimental results show that, on the Cityscapes public benchmark, the algorithm achieves an average intersection-union ratio of 76.8%, a 2.4% improvement over the native framework. In NVIDIA Jetson Orin edge hardware deployment, the model inference frame rate reaches 80.5 FPS, with end-to-end latency controlled at 12.4 ms, quantitatively confirming its engineering feasibility in maintaining high-precision dynamic road scene parsing under limited computing resources.