Abstract: To address the limitations of autonomous driving data acquisition systems in flexibility, real-time performance, and coverage of long-tail scenarios, this paper proposes an autonomous driving flexible data acquisition system and shadow mode based on an integrated vehicle cloud architecture. Through modular middleware in the vehicle-cloud architecture, three data acquisition modes are designed: periodic collection, event-triggered collection, and historical retrieval. A nonintrusive shadow mode is implemented to capture edge scenarios, achieving "edge intelligence+ cloud evolution". The vehicle-side middleware is deployed on the autonomous driving domain controller chip, while the cloud-side middleware supports parallel processing of petabyte-scale data. Field tests demonstrate that the system's long-tail scenario collection efficiency improves by 4.3 times, cloud storage load decreases by 72.6%, and end-to-end latency for event-triggered collection remains below 50 ms. With limited capacity of 2G eMMC, the device can support approximately 30 days of local data caching. This research provides a new technical paradigm for closed-loop optimization of autonomous driving data, significantly enhancing data acquisition efficiency and quality while reducing R&D costs, thereby offering robust support for the commercialization of advanced autonomous driving technologies.

Key words: vehicle-to-everything; vehicle cloud computing; autonomous driving; unstructured data collection; shadow mode

摘要： 针对自动驾驶数据采集系统在灵活性、实时性及长尾场景覆盖的不足，文章提出基于 车云一体架构的自动驾驶灵活数据采集系统及影子模式。通过模块化车云一体架构中间件， 设计周期采集、事件触发、历史调取三种数据采集模式，并实现非侵入式影子模式以捕获边 缘场景，实现“边缘智能+云端进化”。车端中间件部署于自动驾驶域控制器芯片，云端中间 件支持 PB 级数据并发处理。实车测试表明，系统长尾场景采集效率提升 4.3 倍，云端存储负 载降低 72.6%，事件触发端到端延迟低于 50 ms，在 2G eMMC 的有限容量下，该设备可支持 约 30 天的本地数据缓存。该研究为自动驾驶数据闭环优化提供了新的技术范式，显著提升了 数据采集效率与质量，降低了研发成本，为高级别自动驾驶技术的商业化落地提供了有力支 撑。

关键词: 车联网；车云计算；自动驾驶；非结构化数据采集；影子模式