基于轻量化YOLOv5s的交通标志检测研究

Research on traffic sign detection based on lightweight YOLOv5s

针对交通标志检测存在模型精度较低、体积较大、不方便部署的问题,提出一种轻量化的YOLOv5s交通标志检测方法。首先使用注意力机制SENet改进C3模块,将CBAM模块融合进模型中,提升模型精度;然后删减一半卷积核实现模型的轻量化;最后,采用TensorRT技术实现模型在Jetson nano上的部署。实验结果表明:使用SEC3替代C3,模型精度由93.6%提高到了94.6%,通过轻量化方法,模型大小由14.2 M下降为3.98 M,而模型精度此时下降为92.0%,再将CBAM模块融合进轻量模型中,精度又提高到了92.9%。最终,运用TensorRT将模型部署在Jetson nano上并加速推理,FPS达到了24.1 f/s,满足了实时交通标志检测的需求。

Due to the problems of low model accuracy,large volume and inconvenient deployment in traffic sign detection task,a lightweight YOLOv5s traffic sign detection method was proposed.The attention mechanism SENet was used to improve the C3 module and the CBAM module was integrated into the model to improve the accuracy of the model.Half of the convolution kernel was cut to achieve the lightweight model.Finally,TensorRT technology was used to implement the deployment of the model on Jetson Nano.The experimental results showed that the model accuracy was improved from 93.6%to 94.6%by using SEC3 instead of C3.Through the lightweight method,the model size was reduced from 14.2 M to 3.98 M,and the model accuracy was reduced to 92.0%.Then the CBAM module was integrated into the lightweight model,and the accuracy was improved to 92.9%.Finally,TensorRT was used to deploy the model on Jetson Nano and accelerate the inference.The FPS reached 24.1 frames per second,which met the demand of real-time traffic sign detection.