摘要
业务数据量不足、缺乏AI应用开发专业知识、终端计算能力弱等因素严重制约了我国轨道交通智能化产品的快速工程化落地。为此,文章构建了轨道交通AI应用开放平台,其打通了模型训练、边缘计算、云边协同等关键环节,提供一站式、全流程的AI应用开发解决方案。该平台在云端,从数据标注、算法设计、模型训练及应用生成等多个维度构建AI开发工具链;在边缘端,以模型推断优化为核心,实现高效推理;通过云边协同机制,实现数据回传和模型部署。由于无人驾驶矿用卡车采用与轨道交通机车相似的自动驾驶技术,因此文章以矿用卡车无人驾驶视觉感知应用为例进行验证。结果表明,采用AI应用开放平台进行应用开发,可以解决业务主体缺乏模型设计、集成与部署专业能力的问题和专业人才短缺的不足,开发部署时间可从常规的3~4个月缩短至1个月;针对石头、矿卡等目标的视觉检测模型测试集平均准确率达到0.988,实现了优秀的感知性能。
Insufficient data, lack of expertise in AI application development, and weak device computing capabilities have severely restricted the rapid engineering implementation of rail transit intelligent products. In order to solve these problems, this paper proposes an AI open platform for rail transit. It builds connection between model training, edge computing and cloud-edge collaboration, and provides full-process AI application development solutions. In the cloud, this platform builds an AI development tool chain including data annotation, algorithm design, model training and application generation. It also provides an efficient model inference framework at the edge. Data collection and model deployment are implemented through the cloud-edge collaboration mechanism. Since unmanned mining trucks use autonomous driving technology similar to rail transit, this paper takes the visual perception application of unmanned mining trucks as an example for verification. The result shows that using AI open platform for application development can effectively reduce application development and deployment time, from 3~4 months normal period to present 1 month.The mean average precision of the visual detection model for stones, mining truck and other targets reaches 0.988,achieving excellent perceptual performance.
作者
林军
刘悦
王泉东
游俊
丁驰
刘任
LIN Jun;LIU Yue;WANG Quandong;YOU Jun;DING Chi;LIU Ren(CRRC Zhuzhou Institute Co.,Ltd.,Zhuzhou,Hunan 412001,China)
出处
《控制与信息技术》
2022年第1期64-70,共7页
CONTROL AND INFORMATION TECHNOLOGY
基金
湖湘青年英才(2020RC3095)。
关键词
边缘计算
模型训练
云边协同
AI应用开放平台
视觉感知
edge computing
mode training
cloud-edge collaboration
AI application open platform
visual perception
