摘要
针对传统的行人车辆目标检测算法因参数量大和计算复杂度高而在现实应用中受限的问题,基于轻量化深度学习网络提出改进的YOLOv5s行人车辆目标检测算法.首先,选用ghost模块替换主干网络中部分卷积模块进行模型剪枝,同时向网络中引入注意力机制,使得网络在减少模型参数量和提升模型性能两方面实现更好的平衡;其次,采用边界框的宽高差值计算代替边界框回归损失函数中宽高比距离的计算,加速网络的收敛;最后,通过构建真实交通场景下的行人车辆目标检测数据集检验模型的准确性和实时性.实验结果表明,在保持原算法较高精度的同时,改进后YOLOv5s算法的参数量下降28%,模型大小降低27%,节省了硬件成本,拓宽了YOLOv5s算法的应用场景.
Traditional pedestrian and vehicle target detection algorithms often cause problems with a large number of parameters and high computational complexity,which restricts its practical applications.To solve this problem,this paper proposes a pedestrian and vehicle target detection algorithm by improving YOLOv5s,based on a lightweight deep learning network.Firstly,ghost module is used to replace some convolutional modules in the trunk network for.At the same time,attention mechanism is introduced into the network,which makes the network achieve a better balance in reducing the number of model parameters and improving the performance of the model.Secondly,the convergence of the network is accelerated by calculating the width and height difference of the bounding box instead of calculating the aspect ratio distance.Finally,the performance of the model is verified by constructing a pedestrian and vehicle target detection data set in a real traffic scene.Experimental results show that while maintaining the high accuracy of the original algorithm,the parameters of the improved YOLOv5s algorithm are reduced by 28%and the model size is reduced by 27%,which saves the hardware cost and broadens the application.
作者
蒋超
张豪
章恩泽
惠展
乐云亮
JIANG Chao;ZHANG Hao;ZHANG Enze;HUI Zhan;YUE Yunliang(School of Information Engineering,Yangzhou University,Yangzhou 225127,China)
出处
《扬州大学学报（自然科学版）》
CAS
北大核心
2022年第6期45-49,共5页
Journal of Yangzhou University：Natural Science Edition
基金
国家自然科学基金资助项目(62203381)
江苏省自然科学基金资助项目(BK20190878)
江苏省高等学校自然科学基金资助项目(18KJB120011,19KJB110026).
