基于半监督学习的自动驾驶场景下的目标检测  

Object detection in automatic driving scenarios based on Semi-supervised learning

作　　者：文海名 童孟军[1,2] WEN Haiming;TONG Mengjun(College of Mathematics and Computer Science,Zhejiang A&F University,Hangzhou 311300,Zhejiang,China;Zhejiang Key Laboratory of Forestry Intelligent Monitoring and Information Technology Research,Hangzhou 311300,Zhejiang,China)

机构地区：[1]浙江农林大学数学与计算机科学学院,浙江杭州311300 [2]浙江省林业智能监测与信息技术研究重点实验室,浙江杭州311300

出　　处：《微电子学与计算机》2023年第2期22-36,共15页Microelectronics & Computer

基　　金：国家自然科学基金项目(31570629);浙江省自然科学基金项目(LY16F020036)。

摘　　要：由于自动驾驶场景下拍摄的图像目标尺度变化剧烈和环境复杂多变,检测具有不小的难度;获取大量模型训练需要的标注数据图像存在困难,而获取大量未标注数据图像较容易.为了解决上述两个问题,提出一种基于半监督学习的自动驾驶场景下的目标检测模型TransDet.首先,在特征提取部分提出一个具有全局注意力的MSADark模块,以提取图像更多的全局信息以及捕获远程依赖关系;其次在特征融合部分提出一个位置注意力加权特征融合网络LAFFN,用于不同特征融合层捕获局部的位置和通道信息,增强多层次特征加权融合和网络特征表示能力,缓解目标尺度剧烈变化的影响;最后提出一种简单高效的半监督学习算法框架EODS,高效利用未标注数据的同时进一步提升了模型性能.实验结果表明:改进模型在保证实时性的情况下,mAP@50精度从55.1%提升到了61.6%,相比最新的YOLOv5模型精度增加了6.5%,在保证实时的检测速度的同时提升模型检测性能.特别是在仅使用少量未标注数据的情况下使用半监督学习算法EODS将mAP.50性能提升至65.4%,提升达到10.3%,表明了该模型在自动驾驶场景下的目标检测的有效性.Due to the dramatic changes in the scale of the images captured in the autonomous driving scene and the complex and changeable environment,the detection is not small,and it is difficult to obtain a large number of labeled data images required for model training,while it is easier to obtain a large number of unlabeled data images.In order to solve the above two problems,a semi-supervised learning-based object detection model TransDet in autonomous driving scenarios is proposed.Firstly,a MSADark module with global attention is proposed in the feature extraction part to extract more global information of the image and capture long-range dependencies;secondly,in the feature fusion part,a positional attention weighted feature fusion network LAFFN is proposed for different feature fusion layers.Capture local location and channel information,enhance the ability of multi-level feature weighted fusion and network feature representation,and alleviate the impact of drastic changes in target scale;finally,a simple and efficient semi-supervised learning algorithm framework EODS is proposed.Further improved model performance.The experimental results show that the accuracy of mAP@50 increases from 55.1%to 61.6%when the improved model guarantees real-time performance,which is 6.5%higher than the accuracy of the latest Y0L0v5 model.While ensuring the real-time detection speed,the model detection is improved,performance.Especially when using only a small amount of unlabeled data,the semi-supervised learning algorithm EODS improves the mAP.50 performance to 65.4%,and the improvement reaches 10.3%,which shows the effectiveness of the model in object detection in autonomous driving scenarios.

关 键 词：目标检测 自动驾驶 注意力机制 特征融合 半监督学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]