An Enhanced Multi-Target Detection Model for Autonomous Driving Based on Deep Learning Optimization
IEEE Access · 2026 · 14 · 34726-34739
DOI: 10.1109/access.2026.3669801 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Yongjie Zhang und 3 weiteren Autor:innen. Die Arbeit nennt 30 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National University Innovation and Entrepreneurship Training Project: Research on Behavioral Decision Making of Autonomous Driving Vehicles under Complex Urban…, 2025 Guangxi University Young and Middle-Aged Teachers basic Scientific Research Ability improvement Project “Research on multi-Object Tracking Fusion Model of….
Abstract (Originalsprache)
To tackle the core challenges of low real-time performance, frequent misdetections and missed detections in complex scenes, and poor small object recognition of small objects in autonomous driving multi-object detection, this paper proposes an enhanced object detection solution with the following key innovations: 1) Replacing traditional convolutions in the backbone network with the T-CMUNeXt module to overcome computational inefficiency and shallow feature extraction, thereby providing high-quality initial feature maps; 2) Embedding the SA (Shuffle Attention) mechanism at the backbone’s output and the neck’s input layers. This enhances cross-channel interactions by dynamically adjusting weights through channel rearrangement, emphasizing targets while suppressing background noise. 3) A P2 small object detection layer has been added to the neck to compensate for the deficiency of low-level target information. This strengthens semantic and spatial fusion, improving the accuracy of small object capture in complex scenes. 4) The dual attention mechanisms of SE and SimAM are deployed to jointly optimise features in the channel and spatial dimensions, making the solution more robust in complex scenes. 5) EfficiCIoU_Loss is used to adapt to objects of different sizes, improving the accuracy of small object recognition. These improvements enable full-chain optimisation, from feature extraction and fusion enhancement to loss supervision. Experiments show that the model has achieved an average precision improvement of about 35% over the baseline model for small object detection. The detection speed is 40 FPS on an NVIDIA Jetson AGX Orin. In complex scenes, the average false positive and false negative rates decrease by around 15.3 and 12.7 percentage points respectively. These results significantly enhance adaptability in complex environments and the practicality of engineering.
Förderung
- National University Innovation and Entrepreneurship Training Project: Research on Behavioral Decision Making of Autonomous Driving Vehicles under Complex Urban…
- 2025 Guangxi University Young and Middle-Aged Teachers basic Scientific Research Ability improvement Project “Research on multi-Object Tracking Fusion Model of…
Zitieren
APA
Zhang, Y., Feng, X., Hu, X., & Zou, F. (2026). An Enhanced Multi-Target Detection Model for Autonomous Driving Based on Deep Learning Optimization. IEEE Access, 14, 34726-34739. https://doi.org/10.1109/access.2026.3669801
Vancouver
Zhang Y, Feng X, Hu X, Zou F. An Enhanced Multi-Target Detection Model for Autonomous Driving Based on Deep Learning Optimization. IEEE Access. 2026;14:34726-34739. doi:10.1109/access.2026.3669801
BibTeX
@article{Zhang2026,
author = {Zhang, Yongjie and Feng, Xiaoyang and Hu, Xiaoyue and Zou, Fanglin},
title = {An Enhanced Multi-Target Detection Model for Autonomous Driving Based on Deep Learning Optimization},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {34726--34739},
doi = {10.1109/access.2026.3669801}
}
EUVIMED