军事知识服务平台
L3级自动驾驶工况下驾驶人非驾驶任务轻量级检测算法
Lightweight MobileViT-WT Detection Algorithm for Non-Driving Tasks under L3 Conditions
摘要 针对驾驶人非驾驶任务的检测和评估问题,在实车座舱驾驶模拟器中搭建L3级自动驾驶工况,对40名被试者开展了真实非驾驶任务采集试验,设计了移动设备操作、转身取物品等10种非驾驶任务真实试验,构建了相应的真实操作数据集.通过将小波变换卷积融入MobileViT架构,结合主干网络优化、嵌入位置调整及卷积核参数重构,将多尺度处理与增强局部-全局注意力机制进行融合建立了MobileViT-WT模型.测试结果表明,该模型在自建数据集上对非驾驶任务实现了93.85%的分类准确率,泛化性测试集准确率达82.90%,较原模型提高5.15百分点,参数量较原模型减少11%,模型处理图像的推理延迟仅为8.0 ms.研究结果可为L3级自动驾驶的人机接管与安全预警提供优化.
Abstract:
Under the L3 level automatic driving condition,the driver is in the state of leaving the driving link.Real time detection of the driver's non-driving task is an important mechanism to ensure the safe operation of the L3 level automatic driving system for the automatic driving system to understand the driver's state and evaluate the driver's emergency ability according to the driver's state.For the detection and evaluation of drivers'non-driving tasks,this study built L3 level automatic driving conditions in the real vehicle cockpit driving simulator,carried out real non-driving task collection tests on 40 subjects,carried out 10 kinds of real non-driving tasks experiments such as mobile device operation and turning around to pick up items,and constructed the corresponding real operation data set.By integrating wavelet transform convolution into MobileViT architecture,combined with backbone network optimization,embedding position adjustment and convolution kernel parameter reconstruction,the MobileViT-WT model is established by integrating multi-scale processing and enhanced local global attention mechanism.The test results show that the model achieves 93.85%classification accuracy for non-driving tasks on the self-built data set,and the accuracy of the generalized test set is 82.90%,which is 5.15%higher than the original model,and the number of parameters is 11%lower than the original model.The reasoning delay of the model processing image is only 8.0 ms.The research results can provide optimization for man-machine takeover and safety early warning of L3 automatic driving.
刊 名 汽车技术 ISTICPKU
年,卷(期): 2026, (10)
分类号: U461.91
机标分类号: TP391.41 TN929.5 U471.1
在线出版日期: 2026年11月2日
基金项目: 2023-YBGY-035:陕西省重点研发计划项目23-69K:陕西省交通运输厅交通科研项目52362050:国家自然科学基金,