YOLO-FS: a unified framework for object detection and semantic segmentation

Abstract

The integration of object detection and semantic segmentation leverages the advantages of object localization and pixel-level semantic understanding to provide enhanced environment awareness for robot navigation and autonomous driving systems. In this paper, we propose an innovative model that combines YOLOv5 for object detection with a Fast-SCNN-based semantic segmentation module to form a unified framework capable of performing both object detection and semantic segmentation tasks. The model is trained and tested on public and homemade dataset, and validated using camera data collected from self-driving vehicles and quadruped robots. The experimental results show that the model has a mAP50 of 48.3% an improvement of 1.6% over the original algorithm, and the segmentation mean intersection rate (MIoU) on the public dataset is 70.6% an improvement of 2.5% over the original algorithm . On the homemade dataset, the performance of the model improved significantly with vehicle detection accuracy of more than 90% and average intersection joint rate of 89.3%. These findings indicate that the model can effectively enhance perception in complex environments and provide key support for safer and more efficient autonomous navigation.