面向智能交通边缘设备的轻量级高精度车辆检测算法研究

Research on Lightweight High-Precision Vehicle Detection Algorithm for Intelligent Transportation Edge Equipment

摘要: 车辆检测是智能交通系统的关键技术，然而，将计算密集型的深度学习模型部署于资源受限的边缘设备面临严峻挑战，限制了实时性能的发挥。为解决此问题，本研究基于轻量级检测器YOLOv5s，提出了两种优化的轻量化车辆检测模型。第一种模型LVD-YOLO，以追求极致效率为目标，通过采用EfficientNetv2作为骨干网络、结合BiFPN与CBAM注意力机制增强特征融合、并引入SIoU损失函数优化边界框回归，旨在显著降低模型的参数量与计算复杂度。第二种模型EMO-YOLO，侧重于提升复杂交通场景下的检测精度，同时保持轻量化特性。该模型利用新颖的EMO网络作为骨干，通过引入SCConv重构颈部的C3模块以减少特征冗余，并设计了包含三重注意力机制的新检测头，以增强对小目标、遮挡目标等困难样本的检测能力。在公开数据集UA-DETRAC上的大量实验结果表明：与基准YOLOv5s相比，LVD-YOLO在模型参数量和FLOPs上实现了大幅削减，展现了优越的效率；EMO-YOLO则在保持显著轻量化的同时，在检测精度(特别是mAP@0.5)上超越了YOLOv5s及其他对比的轻量化方法，尤其在复杂场景下表现更佳。消融实验进一步验证了EMO-YOLO中各改进模块的有效性。本研究提出的LVD-YOLO和EMO-YOLO模型为智能交通系统在边缘设备上的高效、精准车辆检测提供了具有竞争力的解决方案。

Abstract: Vehicle detection is a crucial technology for Intelligent Transportation Systems. However, deploying computationally intensive deep learning models on resource-constrained edge devices poses significant challenges, limiting real-time performance. To address this issue, this study proposes two optimized lightweight vehicle detection models based on the lightweight detector YOLOv5s. The first model, LVD-YOLO, aims for ultimate efficiency by employing EfficientNetv2 as the backbone network, enhancing feature fusion with BiFPN combined with the CBAM attention mechanism, and optimizing bounding box regression using the SIoU loss function, intending to significantly reduce model parameters and computational complexity. The second model, EMO-YOLO, focuses on improving detection accuracy in complex traffic scenarios while maintaining lightweight characteristics. This model utilizes the novel EMO network as its backbone, reconstructs the C3 module in the neck using SCConv to reduce feature redundancy, and designs a new detection head incorporating a triple attention mechanism to enhance the detection capability for difficult samples such as small and occluded objects. Extensive experimental results on the public UA-DETRAC dataset demonstrate that: compared to the baseline YOLOv5s, LVD-YOLO achieves substantial reductions in model parameters and FLOPs, showcasing superior efficiency; EMO-YOLO, while remaining significantly lightweight, surpasses YOLOv5s and other compared lightweight methods in detection accuracy (especially mAP@0.5), performing particularly well in complex scenarios. Ablation studies further validate the effectiveness of each improved module within EMO-YOLO. The proposed LVD-YOLO and EMO-YOLO models in this study offer competitive solutions for efficient and accurate vehicle detection on edge devices within intelligent transportation systems.