Title: Hie-Graph-YOLOv9: a hierarchical YOLOv9 model with graph-based SE attention mechanism for vehicle detection in complex background

Authors: T. Selvamuthukumar; K. Vijayalakshmi; P. Dhanalakshmi; R. Abinaya

Addresses: Department of Computer Science and Engineering, KL University, Vadeswarram, Guntur, Andhra Pradesh, India; Annamalai University, Chidambaram, Tamil Nadu, India ' Department of Computer Science and Engineering, KL University, Vadeswarram, Guntur, Andhra Pradesh, India; Annamalai University, Chidambaram, Tamil Nadu, India ' Department of Computer Science and Engineering, KL University, Vadeswarram, Guntur, Andhra Pradesh, India; Annamalai University, Chidambaram, Tamil Nadu, India ' Department of Computer Science and Engineering, KL University, Vadeswarram, Guntur, Andhra Pradesh, India; Annamalai University, Chidambaram, Tamil Nadu, India

Abstract: Advanced vehicle detection algorithms are key to Intelligent Transportation Systems (ITS), enabling real-time traffic analysis and congestion and security management. Existing models like YOLOv9 face challenges in feature selection and learning, especially in dynamic or cluttered environments. To address these limitations, this research proposes Hie-Graph-YOLOv9 which is an extended version of YOLOv9 based on improving the feature selecting, feature learning and loss function by incorporating Hiera Transformers, Graph-based GAN-SE attention mechanism and Geometric-based Weighted Smooth L1 loss function. Hiera Transformers, integrated into the backbone network across four stages, refine multi-scale feature learning, ensuring robust representation of fine-grained and global patterns. The Graph-based GAN-SE, embedded in the bottleneck module, emphasises critical regions of feature maps, enhancing detection accuracy. Additionally, a Geometric-based Weighted Smooth L1 loss function is employed for bounding box regression, improving convergence speed and training stability. Experimental evaluations demonstrate the superiority of Hie-Graph-YOLOv9, achieving an AP (0.5) of 79.5%, improvement of faster convergence by 120 Epochs and an increased inference speed of 41.95 FPS, outperforming state-of-the-art models. This work offers a significant step forward in vehicle detection under complex real-world conditions.

Keywords: object detection; YOLO; vehicle; Hiera; graph; squeeze and excitation.

DOI: 10.1504/IJCAT.2026.153101

International Journal of Computer Applications in Technology, 2026 Vol.78 No.3, pp.167 - 185

Received: 13 Sep 2024
Accepted: 27 Mar 2025
Published online: 22 Apr 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article