| 摘要: |
| 针对深度学习目标检测方法在实际交通场景应用中存在参数量、计算量大、难以部署的问题，提出了一种基于YOLOv8n的轻量化车辆检测算法RHL-YOLOv8。首先，在特征提取网络中使用重参数视觉模块（RepViTBlock，RVB）构建全新的C2f-RVB，在保留特征信息的同时消除与跳跃连接相关的计算和内存开销，降低网络参数量。其次，使用高效局部注意力机制（Efficient Local Attention，ELA）改进多尺度特征融合网络（Hybrid Structure Feature Pyramid Network，HSFPN）金字塔结构解决其过滤小尺度特征信息问题。最后，通过组归一化（Group Normalization，GN）卷积和Scale层设计全新的轻量级共享卷积检测头（Lightweight Shared Convolutional Detection Head，LSCD），在轻量化的同时弥补特征融合网络带来的精度损失。实验表明，所提算法相较于YOLOv8算法在UA-DETRAC数据集上mAP@0.5提升1.4%，参数量下降50.0%，计算量降低32.0%，帧率达到99.7；在KITTI数据上保持mAP@0.5稳定的同时，参数量和计算量分别降低43.3%、28.4%，均衡了检测精度与模型轻量化。 |
| 关键词: 车辆检测 深度学习 YOLOv8n |
| DOI：10.20079/j.issn.1001-893x.240905002 |
|
| 基金项目:重庆市科委科学技术研究项目（CSTB2022NSCQ-MSX1425）；重庆市教委科学技术研究项目（KJQN202101510）；重庆科技大学研究生创新项目（YKJCX2320403） |
|
| RHL-YOLOv8:a Lightweight Traffic Vehicle Detection Algorithm |
| PENG Jie琣,SU Yingying琣,DU Qian琣,LIU Can琣,LI Wenjie琤 |
| (a.School of Electronic and Electrical Engineering;b.School of Mathematical and Physical Sciences，Chongqing University of Science and Technology，Chongqing 401331，China) |
| Abstract: |
| To address the challenges of high parameter counts，substantial computational demands，and deployment difficulties associated with deep learning object detection methods in real-world traffic scenarios，a YOLOv8n-based lightweight vehicle detection algorithm named RHL-YOLOv8 is proposed.Firstly，a novel C2f-RVB module is introduced within the feature extraction network by leveraging the RepViTBlock module(RVB).This eliminates the computational and memory overhead associated with skip connections while preserving critical feature information，thereby reducing the overall network parameters.Secondly，the hybrid structure feature pyramid network(HSFPN) is enhanced by using an efficient local attention(ELA) mechanism to mitigate the loss of small-scale feature information.Finally，a new lightweight shared convolutional detection head(LSCD) incorporating group normalization(GN) convolution and scale layers is designed.This compensates for any accuracy loss due to feature fusion while maintaining a lightweight architecture.Experimental results demonstrate that the proposed algorithm achieves a 1.4% improvement in mAP@0.5 compared to the YOLOv8 algorithm on the UA-DETRAC dataset，reduces the number of parameters by 50.0%，decreases computational costs by 32.0%，and attains a frame rate of 997 frame per second.On the KITTI dataset，it maintains a stable mAP@0.5 while reducing the parameter count and computational cost by 43.3% and 28.4%，respectively，achieving a balanced trade-off between detection accuracy and model efficiency. |
| Key words: vehicle detection deep learning YOLOv8n |
《电讯技术》编辑部 