The rapid evolution of intelligent transportation systems (ITS) has necessitated the development of highly accurate and efficient real-time object detection models for tasks such as vehicle identification, traffic monitoring, and autonomous driving. These applications demand lightweight architectures that can operate under strict latency and resource constraints, especially in embedded or edge environments. Object detection, as a foundational task in computer vision, has witnessed significant breakthroughs driven by deep convolutional neural networks (CNNs), among which the You Only Look Once (YOLO) family of models stands out for balancing speed and accuracy.

The YOLO paradigm emerged with YOLOv1 [1], introducing a unified, single-stage architecture capable of real-time detection. This was further improved in YOLOv2 [2] and YOLOv3 [3] through enhancements in backbone networks and anchor-based predictions. YOLOv4 [4] leveraged additional tricks such as CSPDarknet53 and bag-of-freebies techniques to increase mAP without sacrificing speed. While YOLOv5 [5] did not emerge from a peer-reviewed outlet, it became widely adopted for its modularity and deployment flexibility. YOLOv6 [6] focused on industrial scenarios with anchor-free mechanisms, and YOLOv7 [7] introduced trainable bag-of-freebies and E-ELAN modules for efficient training.

The most recent versions, YOLOv8 [8], YOLOv9 [9], and YOLOv10 [10], integrate novel strategies such as native Python implementation, programmable gradient information, and end-to-end detection pipelines. These developments reflect the community’s shift toward optimizing both model architecture and training efficiency. Despite these advances, challenges remain in improving detection in cluttered, occluded, and dynamically varying environments typical in real-world ITS. Additionally, computational burdens imposed by dense feature extraction and large-scale convolutional layers often hinder deployment on resource-constrained platforms.

To address these limitations, recent research has explored advanced architectural modules aimed at reducing computational complexity while maintaining detection precision. Yadav et al. [11] incorporated Ghost convolution and GhostBottleneck layers into YOLOv5s to reduce redundancy in feature extraction. Similarly, Bao and Gao [12] introduced YED-YOLO for autonomous driving, optimizing the detection pipeline using lightweight convolutional mechanisms. Du et al. [13] proposed GhostC2f-enhanced YOLOv8 models for distracted driving detection, demonstrating real-time performance in resource-constrained environments. Lv et al. [14] further enhanced model efficiency by integrating GhostNetV2 with SE attention in GS-YOLO, while Tang et al. [15] extended the GhostNet architecture to support long-range attention with minimal overhead.

In addition to architectural modifications, optimization techniques have also shown promise. Liu et al. [16] designed a lightweight vehicle detection model suitable for real-time deployment in ITS. Ullah et al. [17] augmented YOLOv8 with GhostConv and attention modules for improved performance in dynamic traffic scenarios. The use of Exponential Moving Average (EMA) was explored by Li et al. [18] in EMA-YOLO to stabilize training and boost convergence, and similarly by Han et al. [19] in a maritime object detection context. Zhang et al. [20] leveraged multi-coordinate aggregation attention and shared convolution for fast and accurate road defect detection. Collectively, these studies demonstrate that incorporating Ghost-based convolutions, attention mechanisms, and EMA techniques can significantly improve detection performance, training stability, and computational efficiency in intelligent transportation systems.

In this context, we propose a novel EMA-GhostConv-enhanced YOLOv8 framework tailored for robust vehicle detection in ITS environments. By embedding GhostConv layers within the YOLOv8 backbone and integrating EMA-based parameter updates, our model achieves a better trade-off between speed and accuracy. Experimental evaluations on benchmark traffic datasets validate the efficacy of the proposed method, outperforming existing state-of-the-art lightweight detectors in both precision and inference time.

The key contributions of this paper are as follows:

To enhance the neck of YOLOv8n, we introduce a novel EMA feature layer for stabilizing multi-scale feature learning alongside a lightweight architecture built by replacing standard convolutions with unified GhostConv-C3Ghost blocks. This dual approach improves generalization against noise and significantly reduces computational redundancy while preserving expressive power for accurate detection in complex scenes.

The overall structure of the paper is illustrated in Fig. 1. The remainder of this paper is organized as follows: Section 2 reviews related work in vehicle detection and deep learning architectures. Section 3 presents the proposed methodology. Section 4 details the experimental setup and results. Finally, Section 5 concludes the paper and outlines potential future research directions.

Real-time object detection in intelligent transportation systems (ITS) presents persistent challenges, particularly in balancing computational efficiency with detection accuracy under complex conditions. A wide range of YOLO-based architectures has been adapted for ITS scenarios, with recent efforts shifting toward task-specific lightweight designs and embedded deployment. Liu et al. [16] proposed a compact detection model for real-time vehicle recognition on resource-limited hardware. Bakirci [21] explored aerial monitoring using YOLOv8 variants to classify vehicles across road networks. Zeng and Zhong [22] targeted road surface anomalies using a customized YOLOv8n variant, while Zhu et al. [23] enhanced infrastructure inspection through UAV-mounted YOLOv8 detectors. Du et al. [13] integrated GhostC2f modules within YOLOv8 to detect distracted driving with high speed and accuracy.

To reduce computational cost without sacrificing performance, Ghost convolution and its variants have become popular. Tang et al. [15] introduced GhostNetv2 to minimize redundant feature maps while incorporating long-range attention. Lv et al. [14] built on this by combining GhostNetV2 with SE attention in GS-YOLO, achieving superior performance in SAR ship detection. Yadav et al. [11] embedded GhostConv and GhostBottleneck layers in YOLOv5s to reduce latency while maintaining high precision. Ullah et al. [17] applied GhostConv and attention into YOLOv8 for dense ITS vehicle detection, while Elhenidy et al. [24] introduced ghost separable convolution in GY-YOLO for pedestrian detection.

The use of Exponential Moving Average (EMA) has shown potential in stabilizing training and improving model generalization. Li et al. [18] proposed EMA-YOLO, enhancing object detection consistency across dynamic scenes. Han et al. [19] also incorporated EMA with channel attention in Light-YOLOv7 to detect maritime targets efficiently. These contributions demonstrate that EMA is a powerful yet underutilized technique in lightweight detection.

Several studies have also fused attention modules to strengthen detection in low-contrast and occluded scenarios. Zhang et al. [20] developed a multi-coordinate aggregation attention block for road defect detection using shared convolution, and Wei et al. [25] introduced a multi-scale attention-enhanced YOLO model for remote sensing. Zhou et al. [26] combined layer-adaptive pruning and multidimensional feature fusion in Mp-YOLO for dense traffic object detection.

In addition, a range of YOLO-based models has been customized for ITS-specific applications. Thatikonda [27] built a lightweight helmet and license plate detector using YOLOv8. Bao and Gao [12] proposed YED-YOLO for autonomous driving, demonstrating a lightweight yet accurate solution. Yu et al. [28] optimized YOLOv8 for general vehicle detection across environments, and Pan et al. [29] introduced LVD-YOLO for low-resource ITS deployments.

Despite these valuable efforts, most existing models address only partial aspects of the challenges, such as computational overhead, unstable training, or limited generalization, without integrating all three effectively. As summarized in Table 1, prior approaches typically optimize for one or two factors, but few offer a unified solution. Our proposed EMA-GhostConv YOLOv8 model is designed to overcome these gaps by unifying Ghost convolution, EMA training stabilization, and lightweight attention into a single, coherent architecture. The result is a highly accurate, computationally efficient, and robust detector tailored for real-time ITS vehicle detection in complex environments.

This section outlines the proposed vehicle detection framework built upon the YOLOv8 object detection model. We first present an overview of the original YOLOv8 architecture, followed by the integration of our key enhancements: Exponential Moving Average (EMA) for stabilized learning, Ghost convolution for lightweight feature extraction, and a C3Ghost module to support multi-scale representation. The modified framework is trained and evaluated on two datasets, KITTI and a custom real-world traffic dataset, designed to test its performance under diverse road conditions. We also describe the metrics used to evaluate the detection performance.

This study presents the EMA-GhostConv YOLOv8, a refined vehicle detection framework that enhances feature stability and fusion efficiency by integrating an Exponential Moving Average (EMA) layer and GhostConv-based modules within the neck structure. Extensive experiments on both the KITTI benchmark and a challenging custom vehicle dataset demonstrate the proposed model’s consistent superiority over the YOLOv8n baseline and several state-of-the-art detectors, including YOLOv5, YOLOv7, Faster R-CNN, SSD, RetinaNet, DETR, and EfficientDet-D2. The EMA-GC YOLOv8 achieves the highest precision, recall, and mean Average Precision (mAP) values, as summarized in Table 6.

Ablation studies (Table 5) further validate the individual and synergistic benefits of the EMA and GhostConv components, confirming that these targeted architectural enhancements substantially improve detection robustness under challenging conditions such as partial occlusion, varying illumination, and dense traffic scenes. The proposed model achieves an optimal trade-off between accuracy and computational efficiency, making it highly suitable for deployment in edge-based intelligent transportation systems and real-time surveillance applications.

Future work will explore incorporating temporal feature modeling for video-based object detection, cross-domain adaptation techniques to enhance generalization across diverse environments, and lightweight quantization strategies to further optimize the model for embedded and resource-constrained platforms.