摘要
为解决已有不良驾驶行为检测算法中存在模型参数量过多、计算量较大、实时性差等不足,并且很难部署到边缘设备等问题,文章提出一种基于YOLOv5的不良驾驶行为检测改进算法YOLOv5-MBi。在该模型中将YOLOv5主干网络替换为轻量级的MobileNetV3网络模块,减少训练参数,同时在YOLOv5中的Neck层结合BiFPN特征融合网络来提高其性能,实现自上至下和自下至上的深浅层特征双向融合,最后对改进的模型以及原网络进行相关的测试,验证所修改方法的有效性以及实时性。实验结果表明,YOLOv5-MBi算法在State Farm数据集上准确率达到了93.3%,相比于原始的YOLOv5s算法,参数量相对原有网络降低了46.7%,每秒传输帧数比原有网络提高了53.3%。实验证明改进后的算法在能保证较高检测准确率的同时,模型参数量大幅下降,能更好地满足检测实时性,对交通安全方面具有重要的实际应用价值。
In order to solve the problems of excessive model parameters,high computational complexity,poor real-time performance,and difficulty in deploying to edge devices in existing bad driving behavior detection algorithms,the article proposes an improved algorithm for bad driving behavior detection based on YOLOv5,YOLOv5-MBi.In this model,the YOLOv5 backbone network is replaced with a lightweight MobileNetV3 network module to reduce training parameters.At the same time,the Neck layer in YOLOv5 is combined with the BiFPN feature fusion network to improve its performance,achieving bidirectional fusion of deep and shallow features from top to bottom and bottom to top.Finally,relevant tests are conducted on the improved model and the original network to verify the effectiveness and real-time performance of the modified method.The experimental results show that the YOLOv5-MBi algorithm has an accuracy of 93.3%on the State Farm dataset.Compared to the original YOLOv5s algorithm,the parameter count is reduced by 46.7%compared to the original network,and the number of frames transmitted per second is increased by 53.3%compared to the original network.Experiments have demonstrated that the improved algorithm can ensure high detection accuracy while significantly reducing the number of model parameters,which can better meet the real-time detection requirements and has important practical application value in traffic safety.
作者
李彬彬
丁纪峰
LI Binbin;DING Jifeng(School of Information and Communication Engineering,Dalian Minzu University,Dalian 116605,Liaoning,China)
出处
《智能计算机与应用》
2025年第5期216-220,F0003,共6页
Intelligent Computer and Applications
