Back to home

Compare

Comparing: China's Driver-Assistance All Runs on BEV — And It Shapes Your Next Car & 中国辅助驾驶都在用'BEV' — 这项底层技术决定你下一辆车体验

AEN
BEVautonomous-drivingend-to-end·

China's Driver-Assistance All Runs on BEV — And It Shapes Your Next Car

Over the past three years, almost every new Chinese EV maker's driver-assistance stack — from R&D to deployment — has run on a perception architecture called BEV (Bird's Eye View). Tesla is the lone exception. Beneath this lies a two-path debate over how autonomous vehicles should "see" the world — a debate that matters to every consumer shopping for an intelligent driving-equipped car.

What This Is

BEV stands for Bird's Eye View. The simplest way to think about it: the car gets a top-down map of its surroundings.

Traditional cameras only capture what's in front. Even stitching multiple camera feeds together produces distortion. BEV's core idea: extract features from the six wide-angle cameras around the vehicle, then "project" them onto a 2D grid viewed from above. On this map, vehicles, pedestrians, lane lines, and obstacles all become coordinate points in top-down space — making distance, speed, and position precisely measurable.

BEVDet is one of the most representative open-source models in this space, released in 2022 by MIT's Song Han team. It breaks the pipeline into four stages: image feature extraction, view transformation, BEV feature encoding, and detection output. The code is built on the mmdet3d framework, runs out-of-the-box on GPUs, and has served as the starting point for AV teams both in China and abroad.

The Industry View

There are basically two camps in the mainstream.

The Chinese camp: XPeng, Li Auto, Huawei ADS, BYD, NIO, and most other players use BEV as the perception backbone. The reasoning: an explicit top-down view makes rule-based validation and safety redundancy easier, and it dovetails with the rollout of urban NOA (Navigation on Autopilot in city driving). BEV has since evolved into BEV + Occupancy networks, which can detect "never-before-seen obstacles" and handle long-tail edge cases.

The opposition comes from Tesla: Elon Musk has stuck to "vision-only end-to-end," arguing that an explicit BEV module is just a transitional "crutch" that end-to-end neural networks will eventually replace. After the 2024 FSD update went live, Tesla's stack no longer contains a standalone BEV module.

One caveat: the BEV path isn't free. It's compute-hungry — it needs high-end chips (NVIDIA Orin and above) to run smoothly, which limits adoption in vehicles priced under RMB 100,000. More critically, the industry is rapidly migrating toward end-to-end. Hiring is still booming for "BEV algorithm engineers," but this window may not stay open for long.

Impact on Regular People

For automotive IT: AV teams are seeing explosive demand for talent in BEV perception, Occupancy networks, and end-to-end planning. Starting salaries for these roles in 2024 were noticeably higher than for traditional CV positions — but watch out for "skill depreciation" as the tech stack shifts.

For individual careers: the window is open for engineers outside AV to pivot in, but the ceiling for those who only know BEV in isolation is becoming visible. Pairing it with end-to-end architecture and Transformer fundamentals is the safer bet.

For consumers: BEV has made it possible to roll out urban NOA at scale on vehicles in the RMB 200,000 price band. Next time we're car shopping, "does it use a BEV architecture?" may no longer be marketing fluff — it'll be a hard metric driving real-world differentiation.

Source: juejin.cn
BZH
BEV自动驾驶端到端·

中国辅助驾驶都在用'BEV' — 这项底层技术决定你下一辆车体验

过去三年,中国新势力车企的辅助驾驶系统从研发到落地,底层几乎都跑过一项叫 BEV(鸟瞰图)的感知架构 — 特斯拉是唯一的例外。这背后是自动驾驶"该用什么方式看世界"的两条路线之争,与每个打算买带智驾车型的消费者都相关。

这是什么

BEV 全称 Bird's Eye View(鸟瞰图),把它理解成"让车拥有一张俯视地图"就好。

传统摄像头只能拍到前方画面,多摄像头拼接也有畸变。BEV 方案的核心思路是:把车身周围 6 个广角摄像头的画面先抽取特征,再"投影"到一张从车顶往下看的二维网格上。在这张图里,车辆、行人、车道线、障碍物都变成了俯视坐标点 — 距离、速度、位置都能精确测量。

BEVDet 是这套技术里最有代表性的开源模型之一,2022 年由 MIT 韩松团队发布。它把上述流程拆成四步:图像特征提取、视角转换、BEV 特征编码、检测输出。代码层面基于 mmdet3d 框架有现成实现,可以直接跑在 GPU 上,国内外多个自动驾驶团队都以此作为研发起点。

行业怎么看

主流路线基本两派。

国内派:小鹏、理想、华为 ADS、比亚迪、蔚来等大多数玩家把 BEV 作为感知主干。理由是显式俯视图便于做规则校验和安全冗余,也方便后续落地城市 NOA(城区自动辅助驾驶)。BEV 之后又演化出 BEV + Occupancy(占据网络),可识别"没见过的障碍物",应对长尾场景。

反对声音来自特斯拉:马斯克坚持"纯视觉端到端",认为显式 BEV 模块只是过渡阶段的"拐杖",最终会被端到端神经网络完全替代。2024 年 FSD 新版本上线后,特斯拉方案里已看不到独立的 BEV 模块。

需要警惕的是:BEV 路线并非没有代价。它对算力要求高,需要大算力芯片(英伟达 Orin 及以上)才能流畅工作,10 万元以下车型落地受限。更关键的是,行业正在快速向端到端迁移,目前招聘市场还在大举招"BEV 算法工程师",但这个窗口期可能不会太长。

对普通人的影响

对汽车行业 IT:自动驾驶团队对 BEV 感知、Occupancy 占据网络、端到端规划的人才需求集中爆发,相关岗位起薪在 2024 年明显高于传统 CV 岗位,但需要警惕技术路线切换带来的"技能折旧"。

对个人职场:非自动驾驶行业的程序员转型窗口仍在,但只懂 BEV 单点技能的天花板正在显现 — 补足端到端架构、Transformer 基础会更稳妥。

对消费市场:BEV 让城市 NOA 在 20 万元价位车型上大规模落地成为可能。下次你买车时,"用的是不是 BEV 架构"可能不再是营销话术,而是影响实际体验差异化的硬指标。

Source: juejin.cn