Vehicle-pedestrian Instance Segmentation Algorithm Based on Improved YOLOv8n-seg

In the field of autonomous driving, the rapid and precise perception of the environment, along with effective segmentation of vehicles and pedestrians, is a critical area of research. The complexity of real-world scenes, often characterized by occlusions, can lead to suboptimal segmentation results....

Full description

Saved in:
Bibliographic Details
Published in:Engineering letters Vol. 33; no. 6; p. 1879
Main Authors: Fang, Siwen, Zhang, Xinhe, Su, Bochao, Zhu, Wenxuan
Format: Journal Article
Language:English
Published: Hong Kong International Association of Engineers 01.06.2025
Subjects:
Complexity
Instance segmentation
Modules
Pedestrians
Perception
Performance enhancement
Real time
ISSN:1816-093X, 1816-0948
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In the field of autonomous driving, the rapid and precise perception of the environment, along with effective segmentation of vehicles and pedestrians, is a critical area of research. The complexity of real-world scenes, often characterized by occlusions, can lead to suboptimal segmentation results. Moreover, the computational demands of existing models require substantial resources and time for processing image data. In autonomous driving systems, timely perception and decision-making are essential; computational delays can hinder vehicle responsiveness and increase the risk of driving errors. To enhance the performance of vehicle-pedestrian segmentation, this paper proposes a novel single-stage instance segmentation approach based on an improved YOLOv8n-seg model. This improvement involves redesigning the bottleneck module in the core C2f module of YOLOv8n-seg and replacing the feature fusion layer with a Bidirectional Feature Pyramid Network (BiFPN) structure. Evaluations conducted on the Cityscapes dataset demonstrate that our method achieves a 3.1% increase in the mAP50-95mask value and a 1% reduction in FLOPs compared to the original YOLOv8n-seg. Furthermore, experiments on the COCO subset show that our approach achieves a mAP50mask of 55.6%, mAP50-95mask of 34.2%, and significantly improves segmentation performance under various real-world conditions. Consequently, our approach not only enhances segmentation accuracy but also reduces computational complexity, effectively meeting the real-time requirements for vehicle segmentation in autonomous driving applications.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1816-093X
1816-0948