基于多任务学习的全景驾驶感知算法  

Panoramic driving perception algorithm based on multi-task learning

作　　者：吴伟林 刘春泉 余孝源 WU Wei-lin;LIU Chun-quan;YU Xiao-yuan(College of Electronic Information,Guangxi Minzu University,Nanning 530006,China;Guangxi Postdoctoral Innovation Practice Base,Wuzhou University,Wuzhou 543003,China;Key Laboratory of Atomic and Subatomic Structure and Quantum Control(Ministry of Education),School of Physics,South China Normal University,Guangzhou 510006,China)

机构地区：[1]广西民族大学电子信息学院,广西南宁530006 [2]梧州学院广西博士后创新实践基地,广西梧州543003 [3]华南师范大学物理学院原子亚原子结构与

出　　处：《计算机工程与设计》2025年第4期1127-1133,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(62241302);广西科技基地和人才基金项目(桂科AD23026199);广西机器视觉与智能控制重点实验室基金项目(2022B02);广西民族大学研究生创新计划基金项目(gxun-chxs2022095)。

摘　　要：针对全景驾驶感知算法YOLOP存在特征图池化操作自适应较差、下采样过程细节丢失和模型性能差的问题,提出一种基于多任务学习的全景驾驶感知算法,引入高效处理模块,提高对特征图池化操作自适应能力,采用不同加权系数的损失函数,提升网络的检测性能及鲁棒性。在BDD100K数据集的评估结果中,车道线检测准确率提高11.6%,可行驶区域检测的平均交并比(mIoU)提高2.1%,车辆检测的平均精确率均值的50%指标(mAP50)提高3.7%。在KITTI数据集的评估结果中,车辆检测mAP50指标提高3.4%。To address the issues of poor adaptability of feature map pooling operation,detail loss in the downsampling process,and inferior model performance in the YOLOP panoramic driving perception algorithm,a panoramic driving perception algorithm based on multi-task learning was proposed.Efficient processing modules were introduced to improve the adaptability of feature map pooling operations,and loss functions with different weighting coefficients were used to effectively enhance the detection performance and robustness of the network.In the evaluation results on the BDD100K dataset,the lane detection accuracy is improved by 11.6%,the mIoU for drivable area detection is increased by 2.1%,and the mAP50 for vehicle detection is improved by 3.7%.Furthermore,on the KITTI dataset,the mAP50 of vehicle detection is improved by 3.4%.

关 键 词：多任务学习网络 编码-解码器 车道线检测 可行驶区域检测 车辆检测 特征对齐 转置卷积 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]