基于YOLOv5s的轻量化行人检测算法

Pedestrian detection algorithm based on YOLOv5s lightweight

行人检测系统普遍安装在移动智能设备上,而这些设备对模型的轻量化要求较高,已有算法很难在精度和轻量化之间达到平衡。针对这一问题,提出一种改进的YOLOv5s轻量化行人检测模型。选用EIoU作为边界框损失函数,加速收敛并提高回归精度;结合CA(Coordinate Attention)注意力模块改进主干网络的C3模块,增强模型对行人目标的精确定位能力;引入一种新卷积层GSConv替换颈部网络的卷积层(Conv),以减轻模型的复杂度并保持准确性;引入改进的自注意力模块CoT,进一步提高网络模型的特征表达能力。使用INRIA数据集进行训练和测试,实验结果表明:改进后的模型mAP@0.5达到97%,相比于原始模型提高1.9%,mAP@0.5:0.95提高2.1%;而模型参数量降低10.5%,模型体积降低13%,计算量GFLOPS减少7%,能够在提高行人检测精度的同时使得模型更加轻量化。

Pedestrian detection systems are generally installed on mobile smart devices,which require the lightweight of the model.However,existing algorithms struggle to reach a balance between accuracy and lightweight.On this basis,an improved model of YOLOv5s lightweight pedestrian detection is proposed.The EIOU is selected as the bounding box loss function to accelerate convergence and improve the regression accuracy.The C3 module of the backbone network is improved with the CA(coordinate attention)attention module to enhance the model's precise positioning ability for pedestrian targets.A new convolutional layer GSConv is used to replace the convolutional layer(Conv)of the neck network to reduce model complexity and maintain accuracy.The improved self-attention module CoT is introduced to further improve the feature expression capability of the network model.The INRIA data set is used for training and testing,the experimental results show that the mAP@0.5 of the improved model can reach 97%,which is 1.9%higher than that of the original model,mAP@0.5:0.95 can be increased by 2.1%.The reduction in model parameters by 10.5%,model volume by 13%,and computational load GFLOPS by 7%can improve pedestrian detection accuracy while making the model more lightweight.