基于改进YOLOv4的公路路面信息识别算法研究

Research on Highway Pavement Information Recognition Algorithm Basedon Improved YOLOv4

公路路面信息的获取和识别是构建智慧公路系统的关键技术之一,针对现有的公路路面信息检测算法存在的耗时费力、识别精度差与端对端应用难以实现等问题,提出了一种改进的YOLOv4算法。为增强模型泛化能力,以IoU值度量边界框与先验框的距离,形成改进的k-means聚类算法对路面遗撒与病害数据集进行锚框聚类;为提升网络特征描述能力,在特征增强网络PANNet的最后3个分支上分别添加兼具通道与空间注意力机制的轻量化CBAM模块,保证了模块在现有网络架构中做到即插即用;为节约参数与计算力,对稀疏训练后的模型进行通道剪枝,针对道路检测任务对于小目标物体的高识别精度要求进行模型剪枝率的迭代优化,进一步实现公路路面信息识别算法的端对端应用。实验结果表明,改进的YOLOv4网络模型的mAP0.5较原网络模型提升了0.78%,mAP@0.75较原网络模型提升了1.06%,FPS达到了34.85帧/s,检测效果满足自动识别的性能要求;0.4的剪枝率得到的剪枝模型综合性能较好,在保证模型的mAP0.5达到98.3%的条件下,储存空间较原模型降低了47.6%,GFLOPs较原模型降低了34.0%,参数总数较原模型降低了51.4%,FPS较原模型提升了6.3%,计算复杂度和占用内存都显著降低。研究成果可应用于智慧公路的路网感知能力体系建设,实现对公路路面信息的高效精准采集。

The acquisition and recognition of highway pavement information is one of the key technologies for constructing intelligent highway systems.As for problems of time-consuming efforts,poor recognition accuracy,and difficulty in end-to-end application in existing highway pavement information detection algorithms,an improved YOLOv4 algorithm has been proposed.In order to enhance the model's generalization capability,the distance between bounding boxes and prior boxes is measured by IoU value to obtain an improved k-means clustering algorithm forms,which was applied to anchor box clustering of the road surface scatter and disease data;In order to improve the network feature description capability,lightweight CBAM modules with combination of channel and spatial attention mechanisms are added to the last three branches of the feature enhancement network of PANNet,ensuring plug-and-play of the modules in the existing network architecture;In order to save parameters and computational power,channel pruning is performed on the model after sparse training.In order to meet the high recognition accuracy requirements for small objects in road detection tasks,the iterative optimization of the model pruning rate is carried out to further realize end-to-end application of the highway pavement information recognition algorithm.Experimental results show that the mAP@0.5 and mAP@0.75 of the improved YOLOv4 network model increased by 0.78%and 1.06%respectively compared with the original model.Frames per second(FPS)reaches 34.85.The performance requirements of automatic recognition is satisfied;With a O.4 pruning rate,the pruned model showed good overall performance.When the mAP@0.5 of the model is 98.3%,compared with the original model,storage space,GFLOPs and the total number of parameters was reduced by 47.6%,34.0%and 51.4%respectively,and FPS was increased by 6.3%,which significantly reduce computational complexity and memory usage.The research results can be applied to the construction of the road network perception capability system of intelligent highways to achieve efficient and accurate collection of highway pavement information.