Rep⁃YOLOv8车辆行人检测分割算法  

Rep⁃YOLOv8 vehicle and pedestrian detection segmentation algorithm

作　　者：王译崧 华杭波 孔明 梁晓瑜 WANG Yisong;HUA Hangbo;KONG Ming;LIANG Xiaoyu(School of Metrology and Measurement Engineering,China Jiliang University,Hangzhou 310018,China)

机构地区：[1]中国计量大学计量测试工程学院,浙江杭州310018

出　　处：《现代电子技术》2024年第9期143-149,共7页Modern Electronics Technique

基　　金：国家市场监督管理总局技术保障专项(2022YJ21);浙江省市场监督管理局科技计划(全额自筹)项目(ZC2023057)。

摘　　要：车辆行人检测分割在自动驾驶、智能交通管理等场景广泛应用,但如何提高车辆行人识别精度以及处理分割不均匀等问题一直是项挑战。针对上述问题,文中提出一种YOLOv8的改进算法,该算法采用RepECA作为骨干网络,此骨干网络使用RepVGG模块代替原骨干网络的卷积层,并融合ECA注意力机制对图像进行特征提取,其中RepVGG模块在检测时转变多分支结构为单路径结构,不损失训练精度的同时提升执行效率,ECA注意力机制针对通道维度的注意力加权机制,通过学习通道之间的相关性,自适应地调整通道的权重,增加少量模型参数却带来大的性能提升;在C2f模块中,改进算法加入了eSE自注意力模块,避免因为通道数减少造成的通道信息损失,进一步提高模型精度。实验结果表明,使用Cityscapes数据集训练,Rep⁃YOLOv8算法在检测与分割任务的mAP@0.5指标分别达到85.4%和75.5%,与原YOLOv8相比分别提升了13.4%和16%,推理速度从65 f/s提升至83 f/s。Vehicle and pedestrian detection and segmentation are widely applied in scenarios such as autonomous driving and intelligent traffic management.However,improving the accuracy of vehicle and pedestrian recognition and addressing issues like uneven segmentation has been remaining a challenge.In view of this,an improved algorithm based on YOLOv8 is proposed.In the algorithm,RepECA is taken as the backbone network,which replaces the convolutional layers of the original backbone network with RepVGG modules,and integrates the efficient channel attention(ECA)mechanism for image feature extraction.The RepVGG module transforms the multi⁃branch structure into single⁃path structure during detection,enhancing execution efficiency without sacrificing training accuracy.In view of the attention weighting mechanism of channel⁃wise,ECA mechanism adaptively adjusts the weight of channels by learning the inter⁃channel correlations,which adds a few model parameters,but brings great performance improvement.In the C2f module,an eSE(effective squeeze⁃excitation)self⁃attention module is incorporated into the improved algorithm to avoid channel information loss caused by a reduction in the number of channels and further enhance the model accuracy.Experimental results,based on the training with the Cityscapes dataset,show that the Rep⁃YOLOv8 algorithm achieves mAP@0.5 of 85.4%and 75.5%for detection tasks and segmentation tasks,respectively,which represents 13.4%and 16%improvement in comparison with the original YOLOv8.In addition,its inference speed is increased from 65 f/s to 83 f/s.

关 键 词：YOLOv8 RepVGG ECA ESE 目标检测 语义分割 

分 类 号：TN911.73-34[电子电信—通信与信息系统]