Scholarlink

<p>In intelligent transportation surveillance scenarios, existing high-precision vehicle detection models commonly suffer from insufficient real-time performance due to their large parameter sizes and high computational complexity. Conversely, while lightweight models exhibit lower computational demands, they often fail to meet the accuracy requirements of practical applications. To overcome this fundamental trade-off between accuracy and efficiency, this paper proposes an optimized lightweight multi-scale vehicle detection model based on YOLOv8, termed SGS-YOLOv8. Its core innovation lies in the synergistic integration of multiple lightweight techniques into the YOLOv8 architecture, specifically tailored for vehicle detection tasks: (1) Within the backbone network, we innovatively integrate and optimize the SCDown module alongside an improved GCSPR module, significantly enhancing the parameter efficiency of multi-scale feature extraction. (2) In the neck structure, we creatively design a cascaded lightweight architecture composed of GSConv and VoV-GSCSP, effectively mitigating the performance degradation in feature fusion typically encountered during traditional neck lightweighting. (3) Within the detection head, we introduce the computationally efficient Multi-Head Self-Attention (MHSA) mechanism combined with feature screening and weighting strategies, enabling adaptive focus on critical target features and enhancing discriminative capability in complex scenes. Comprehensive experiments conducted on three benchmark datasets—BIT-Vehicle, UA-Detrac, and BDD100K—demonstrate that the proposed SGS-YOLOv8 model achieves significant reductions of 36.90% in parameters, 45.68% in computational load (GFLOPs), and 33.39% in model size compared to the baseline YOLOv8 model, while maintaining average precision with a marginal decrease of no more than 0.2%. These results effectively validate the superiority of the proposed synergistic lightweighting strategy in resolving the accuracy-efficiency trade-off dilemma, offering an efficient solution for real-time applications on edge computing platforms.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Lightweight multi-scale vehicle object detection algorithm based on optimized YOLOv8

  • 期刊论文

  • 发表日期：

    2025年10月9日

查看全文
  • Hanqing Duan,
  • Yixuan Zhang,
  • Songhao Zhu,
  • Zhiwei Liang

摘要

In intelligent transportation surveillance scenarios, existing high-precision vehicle detection models commonly suffer from insufficient real-time performance due to their large parameter sizes and high computational complexity. Conversely, while lightweight models exhibit lower computational demands, they often fail to meet the accuracy requirements of practical applications. To overcome this fundamental trade-off between accuracy and efficiency, this paper proposes an optimized lightweight multi-scale vehicle detection model based on YOLOv8, termed SGS-YOLOv8. Its core innovation lies in the synergistic integration of multiple lightweight techniques into the YOLOv8 architecture, specifically tailored for vehicle detection tasks: (1) Within the backbone network, we innovatively integrate and optimize the SCDown module alongside an improved GCSPR module, significantly enhancing the parameter efficiency of multi-scale feature extraction. (2) In the neck structure, we creatively design a cascaded lightweight architecture composed of GSConv and VoV-GSCSP, effectively mitigating the performance degradation in feature fusion typically encountered during traditional neck lightweighting. (3) Within the detection head, we introduce the computationally efficient Multi-Head Self-Attention (MHSA) mechanism combined with feature screening and weighting strategies, enabling adaptive focus on critical target features and enhancing discriminative capability in complex scenes. Comprehensive experiments conducted on three benchmark datasets—BIT-Vehicle, UA-Detrac, and BDD100K—demonstrate that the proposed SGS-YOLOv8 model achieves significant reductions of 36.90% in parameters, 45.68% in computational load (GFLOPs), and 33.39% in model size compared to the baseline YOLOv8 model, while maintaining average precision with a marginal decrease of no more than 0.2%. These results effectively validate the superiority of the proposed synergistic lightweighting strategy in resolving the accuracy-efficiency trade-off dilemma, offering an efficient solution for real-time applications on edge computing platforms.