DCW-YOLO: Road Object Detection Algorithms for Autonomous Driving

Aiming at the problems of multiple parameters and poor detection accuracy of object detection network in automatic driving scenarios, an object detection algorithm based on improved YOLOv8 is proposed. First, a dynamic head framework is used to unify the object detection head and the attention mecha...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:IEEE access Ročník 13; s. 125676 - 125688
Hlavní autoři: Ren, Hongge, Jing, Fangke, Li, Song
Médium: Journal Article
Jazyk:angličtina
Vydáno: Piscataway IEEE 2025
The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
Témata:
Accuracy
Algorithms
Attention
Autonomous driving
Autonomous vehicles
Benchmark testing
Computational modeling
Deep learning
Feature extraction
Gradient methods
Heuristic algorithms
Location awareness
Object detection
Real-time systems
Roads
Task analysis
YOLO
YOLOv8
ISSN:2169-3536, 2169-3536
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Aiming at the problems of multiple parameters and poor detection accuracy of object detection network in automatic driving scenarios, an object detection algorithm based on improved YOLOv8 is proposed. First, a dynamic head framework is used to unify the object detection head and the attention mechanism, and the attention mechanism is used for scale-awareness, spatial-awareness, and task-awareness, respectively, which significantly improves the representation capability of the object detection head without increasing the computational overhead. Second, the Coordinate Attention mechanism is embedded in the SPPF layer, which embeds the target's location information into the channel attention to offer more precise localization for the model, suppress irrelevant aspects, and enable greater integration of local and global characteristics. Finally, the deleterious gradients generated by low-quality examples are reduced using the Wise-IoU v3 bounding box loss function in conjunction with a dynamic non-monotonic focusing mechanism utilizing an anchor box gradient gain assignment strategy. On the challenging public dataset KITTI, the accuracy is improved by 2.1% compared to the benchmark algorithm. In addition, the excellent performance on CCTSDB2021 and VOC highlights the generalization performance of the improved model.
Bibliografie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:2169-3536
2169-3536
DOI:10.1109/ACCESS.2024.3364681