摘要
随着辅助驾驶的快速发展,交通标志检测精度要求日益提高,但在嵌入式设备部署的实时性仍然具有挑战。于是提出一种基于GhostNet改进的YOLOv8s目标检测算法YOLOv8s-CGSA。首先,采用HardSwish激活函数的GhostConv替代网络Neck部分的全部Conv模块,降低模型参数量、提升表达能力的同时降低计算成本、提高推理速度。其次,采用改进的C2fGhost替代原C2f模块,进一步减少模型参数量、提高模型性能。最后,引入SA注意力机制,增强语义信息和位置信息的融合,提高模型特征融合和检测性能。实验结果表明,在中国道路交通标志检测数据集TT100K上,相对于YOLOv8s原始模型mAP@0.5提高了3.4%,模型参数减少了26%,模型减小了4.1 mb,在嵌入式设备Jetson Xavier NX上,FPS提高了39%,实现了在嵌入式设备上对交通标志实时且准确的检测。
With the rapid development of assisted driving,traffic sign detection accuracy requirements are increasing day by day,but real-time deployment in embedded devices is still a challenge.The paper proposes an improved YOLOv8s target detection algorithm YOLOv8s-CGSA based on GhostNet.Firstly,the paper uses GhostConv of HardSwish activation function to replace all Conv modules in the Neck part of the network,reducing the number of model parameters and improving expression ability while reducing computational costs and improving inference speed.Secondly,the paper uses the improved C2fGhost to replace the original C2f module to further reduce the number of model parameters and improve model performance.Finally,the SA attention mechanism is introduced to enhance the fusion of semantic information and location information,and improve the model feature fusion and detection performance.Experimental results show that in the Chinese road traffic sign detection data set TT100K,compared with the YOLOv8s original model mAP@0.5 is increased by 2.7%,model parameters are reduced by 26%,the model is reduced by 4.1 mb,and FPS is improved on the embedded device Jetson Xavier NX.A real-time and accurate detection of traffic signs on embedded devices is achieved.
作者
唐坤俊
宁媛
刘聂天和
TANG Kunjun;NING Yuan;LIUNIE Tianhe(College of Electrical Engineering,Guizhou University,Guiyang 550025,China;Guiyang Huaxi Power Supply Bureau of Guizhou Power Grid Co.,Ltd.,Guiyang 550025,China)
出处
《智能计算机与应用》
2025年第5期142-148,共7页
Intelligent Computer and Applications
基金
贵州省科技计划基金(黔科合ZK2022135)。
