在絕大多數(shù)人眼里，高精度地圖是高級自動駕駛的核心使能技術之一。開發(fā)這類地圖是一項浩大的工程，需要在資金和人才方面進行巨大投資。
在傳統(tǒng)地圖中，盤山公路上的黃色標志表示前方有蛇形彎道，這很難忠實地反映詳細道路狀況，但卻可以向駕駛員傳遞一個簡單的信號：準備應對彎道。
幾個世紀以來，地圖最本質的作用從未改變：指導人類從 A 地到達 B 地。不難想象，有時，過于詳盡的地圖也會礙事。加州初創(chuàng)公司 Deepmaps 首席運營官羅偉（音）表示，“你不需要過多細節(jié)，細節(jié)過多反而可能讓人迷惑。”羅先生此前曾在谷歌地圖（Google Maps）擔任產品經理。
然而，在為用戶提供信息的同時，地圖公司也同時指望用戶能夠幫自己補充許多缺失的部分，并時刻對變化做出反應。畢竟，地圖的用戶是人類。地圖和語言一樣，是連接人類思維的橋梁符號。
自動駕駛時代的地圖繪制
然而，這種情況正在改變。作為地圖學中的最新領域，專為自動駕駛汽車繪制的地圖面對的則是一類完全不同的用戶：軟件程序。與人類駕駛員不同，導航系統(tǒng)需要的是細節(jié)，每一條曲線、每一個凸起的路沿、每一條車道，每個細節(jié)都至關重要，而且必須達到厘米級精度。除此之外，更具挑戰(zhàn)性的是，自動導航系統(tǒng)必須有能力判斷各種未知情況，并作出相應改變。比如，假如有棵樹倒在路上，導航系統(tǒng)則該如何繼續(xù)提供指引？對于人類駕駛員，這根本不是問題，不過是抱怨幾句，然后換條路而已。但大多數(shù)軟件在面臨類似的情況時都需要非常詳細的指導，否則根本不知道作何反應。
如今，新型地圖已經成為自動駕駛行業(yè)中的基礎技術之一，地圖行業(yè)也由此開始崛起。畢竟，車輛最本質的功能是將人或貨物運送至目的地，而地圖可以幫自動駕駛車輛確定自身位置以及如何到達目的地，與現(xiàn)實世界建立聯(lián)系。
新型地圖需要對城市中的每一條小街小道都進行精確的三維記錄，這本身沒有什么難度。但更重要的是，我們還需要人工智能（AI）技術，協(xié)助車輛判斷行駛路線上可能出現(xiàn)的各種情況，并作出適當?shù)姆磻?，而且通常需要在幾分之一秒內完成?/span>
如今，在自動駕駛汽車發(fā)展如火如荼之時，新一代地圖的繪制已然成為了一項龐大事業(yè)。其中的參與者不僅包括地圖行業(yè)巨頭谷歌 Waymo，還有很多初創(chuàng)公司，比如拿到風投的美國公司 DeepMind 和 Carmera，以及由戴姆勒（Daimler）、大眾（Volkswagen）及其他多家汽車制造商投資的歐洲領先地圖供應商 HereTechnologies。最終，在競爭中勝出的公司將有機會運行一個全球地理平臺，追蹤并指導地球上大部分車輛的行駛?？▋然仿〈髮W機器人系教授 John Dolan 表示，“這是一個非常熱門的研究領域。”
應對變化
對于針對自動駕駛汽車的地圖繪制而言，最大的挑戰(zhàn)在于如何應對變化。“事實上，（自動駕駛地圖）必須是 4D 的，”Deepmap 公司的羅偉Wei Luo 表示，“也就是傳統(tǒng)三維外加時間維度。”為了在地圖中融入時間維度，所有系統(tǒng)都必須通過某種方法來收集最新數(shù)據(jù)，而且必須保證這些數(shù)據(jù)的實效性和可靠性。Waymo 等公司選擇利用自己自動駕駛車隊上的傳感器，其他公司則傾向于采用“眾包”的思路，也就是利用其他車輛上裝載的激光雷達和各種傳感器。
一旦傳感器就位，并開始傳回報告流，數(shù)據(jù)收集部分的工作就很簡單了。“你可以從一張信息豐富的基礎地圖開始，”紐約初創(chuàng)公司 Carmera 創(chuàng)始人兼首席執(zhí)行官 Ro Gupta 表示，“這并不簡單，但從某種程度上來說已經不是問題了。”
事實上，真正構成巨大挑戰(zhàn)的正是大量數(shù)據(jù)本身。羅先生表示，每輛自動駕駛汽車每小時大約可以產生 1 PB 的導航數(shù)據(jù)，這非常龐大，相當于 2 的 50 次方字節(jié)。軟件必須對這些海量數(shù)據(jù)進行篩選，并從中找出有意義的片段，然后“決定”是否采取行動及采取何種行動。這將帶來非常龐大的認知工作，需要人工智能技術的深入?yún)⑴c。
在最初階段，單單識別變化就已經是一個挑戰(zhàn)了。隨著海量數(shù)據(jù)的不斷涌入，基礎地圖將持續(xù)確認各種信息匹配無誤。停車標志？沒問題。左轉車道？也沒問題。
然而，世界常會有新的變化，比如街角處的一棵松樹沒有了，出現(xiàn)了一片空地。系統(tǒng)可以發(fā)現(xiàn)這些變化，但這個變化是否比落葉或水洼的出現(xiàn)更重要？人類駕駛員想都不用想，就會立刻認出某片空地上停了一輛卡車。但軟件系統(tǒng)卻缺乏人類的這種經驗和直覺，因此必須通過更多線索才能進行判斷。觀察結果有更多的數(shù)據(jù)支持嗎？類似大樹這樣的目標曾有多少次消失不見呢？這種情況是否會造成任何事故或其他麻煩？會影響交通的通暢嗎？
在應對變化時，時間至關重要。一種符合邏輯的做法是通過對傳感器車輛進行編程，使其僅在檢測到與基礎地圖出現(xiàn)不符的情況時才進行報告，從而大幅減少數(shù)據(jù)通信量及相關延遲。如果 Broad Street 大街上的三條車道一切如故，那又何必再報告一些沒用的信息，給系統(tǒng)增加“噪聲”呢？Carmera 的 Gupta 表示，不過，問題在于我們可能會忽視掉一些未被察覺的變化。他說，“因而可能會丟失一些假陰性指標。”
是否連“云”？
此外，新型地圖的更新還牽扯各種各樣的數(shù)據(jù)管理問題。例如，哪些地圖數(shù)據(jù)應交由車輛自己解讀，又有哪些應該上傳至基于云的人工智能系統(tǒng)進行判斷？
從一方面，云可以同時從多個來源接收信息，將其與歷史模式進行匹配，并提供更多的智能功能。然而，盡管超高速 5G 蜂窩網絡預計將在三年內得到普及，但數(shù)據(jù)的傳輸依然無法避免延遲問題。更重要的是，由于網絡連接很難得到 100% 的保證，因此自動駕駛汽車也必須配備車載系統(tǒng)，從而具備在不聯(lián)網的情況下對變化進行判斷，并做出適當反應的能力。
在早期階段，大多數(shù)地圖公司都會選擇將部分區(qū)域當作樣本，進行新型地圖開發(fā)。很自然，很多公司都把精力集中在正在進行自動駕駛測試或已經開始提供相關服務的區(qū)域。比如，Waymo 和 Deepmap 均在亞利桑那州和加利福尼亞州的部分地區(qū)投入了很大精力。Carmera 則已經與一些貨運公司簽訂了合作協(xié)議，目前正在與紐約、舊金山和佛羅里達州的老年村進行地圖建模，而這些地區(qū)都是其合作伙伴正在提供自動駕駛穿梭巴士服務的區(qū)域。Here Technologies 公司則是一個例外，這家公司憑借與多家主流歐洲汽車制造商的關系，可以通過這些制造商出售的數(shù)十萬輛汽車上的傳感器，收集歐洲和北美地區(qū)的匿名數(shù)據(jù)。
現(xiàn)階段的營利也很重要
對于一些獲得風投的創(chuàng)業(yè)公司而言，業(yè)務發(fā)展的時機也非常重要。盡管這些公司現(xiàn)在已經開始大量砸錢，但全自動駕駛汽車（也就是 SAE 4 級和 SAE 5 級自動駕駛汽車）的廣泛普及可能要到十幾年以后，甚至更久。因此，這些創(chuàng)業(yè)公司也在為他們的下一代地圖尋找過渡期的市場。Here Technologies 產品營銷經理 Mattew Preyss 提問到，“在過渡期中，我們該如何利用這些數(shù)據(jù)來幫助駕駛員？”
Preyss 表示，下一代地圖將為Waze、谷歌地圖及 TomTom 等當下主流導航服務提供有力補充，時時為駕駛員提供最新路況和路線修正信息。更重要的是，這些地圖還可以提供如增強現(xiàn)實或尋找車位等一系列全新服務，以音頻和畫面的形式為駕駛員提供詳細的路線信息。與以往一樣，只要同時牽涉人類駕駛員和地圖，我們就必須面臨一個永恒挑戰(zhàn)——如何讓地圖為駕駛員提供更多有用信息，但同時剔除可能分散駕駛員注意力的細節(jié)。
現(xiàn)階段，讓人類駕駛員繼續(xù)參與新型地圖繪制還有一個重要作用——地圖本身可以學習人類駕駛員是如何對數(shù)據(jù)做出反應的，進而將更多人工智能處理能力分配在行車路線的中需要車輛立即做出反應的重大變化上。在未來十年中，我們人類駕駛員也將“教導”的導航系統(tǒng)，使其真正做好取代我們的準備。
Most believe ultrahigh-definition mapping is crucial to make high-level automated driving possible. Developing these maps is a huge undertaking — one that’s enjoying a massive investment of money and talent.
A yellow sign on a mountain highway shows an S-shaped curve. This is a primitive map, and hardly a faithful representation of the road. Instead it delivers a simple signal to the driver: Get ready for turns.
Road cartography has evolved over centuries with a unifying purpose: to guide human beings from point A to point B. Complexity often gets in the way. “You don’t want too much detail,” says Wei Luo, formerly a product manager at Google Maps and now chief operating officer at Deepmaps, a Palo Alto, California-based startup. “That can confuse people.”
At the same time, though, the cartographer counts on the map’s user to fill in many of the missing pieces—and respond to changes. After all, the user is a fellow human being. Maps, like language, are symbols that bridge human minds.
New-age cartography for autonomy
But this is changing. The newest field of cartography—creating maps for autonomous vehicles—is designed for a different user: a software program. Unlike a person, the navigation program demands specifics—every squiggle, every raised curb, every passing lane, all of them calibrated by the centimeter. At the same time, and far more challenging, automated navigation must adapt to immediate unknowns. How should it provide guidance to the destination if a fallen tree lies in its path? While a human driver might swear under her breath and improvise, most software programs will require detailed guidance.
An entire industry is rising up to create this new breed of map, a fundamental technology for the nascent autonomous industry. After all, the purpose of the vehicle is to reach a destination. The map tells whe it is and how to get there, the AV’s connection to the physical world.
Creating these maps requires precise three-dimensional recording of every street and byway—itself no mean feat. But it also requires muscular layers of artificial intelligence (AI) to interpret what it encounters along the way and then to respond appropriately. Often within a fraction of a second.
It’s a massive undertaking that feeds this growing field of research. Google’s Waymo, the industry’s AI behemoth, is developing maps for its autonomous fleets. It’s joined by a host of start-ups, including venture-funded DeepMind and Carmera in the U.S. and European-led Here Technologies, which is backed by Daimler, Volkswagen and other automakers. The winners in this market will be positioned to run the world's geo-platforms, tracking and guiding much of the movement on our planet. “It’s a very hot field for research,” says John Dolan, a robotics professor at Carnegie Mellon University.
Dealing with change
A central challenge for autonomy-centric mapping is adapting to change. “The system actually has to be 4D, says Deepmap’s Luo. “That’s 3D plus time.” To incorporate time into the map, each system must devise a method for harvesting reliable, up-to-the-minute data. Some, like Waymo, use the sensors on their own fleets of AVs. Others look to crowdsourced data or piggyback on the onboard LIDAR and other sensors.
Once the sensors are in place and sending back streams of reports, the data-gathering part of job is straightforward. “You start with a very rich base map,” says Ro Gupta, founder and CEO of the New York start-up, Carmera. “That’s not trivial,” he says, “but it’s somewhat a solved problem.”
It’s the flood of data itself that creates immense challenges. Each AV, says Luo, generates about one petabyte per hour of navigational data. Software must sift through this avalanche of data to find the fragments that are meaningful and then “decide” whether take action. This is an enormous cognitive enterprise—and requires strong doses of AI.
The initial challenge is simply to spot a change. As the data pours in, the base map is certifying that everything is matching. Stop sign? Check. Left-turn lane? Check.
Then it encounters something new: A white space at a street corner where there used to be a pine tree. The system notes a change. But is it more significant than other changes, like falling leaves or fresh puddles? A human being might immediately recognize the white space as a parked truck, and not give it a second thought. The software, however, lacking human experience and intuition, must probe for clues. Is there more data to corroborate the observation? How many times have objects, like a tree, gone missing before? Is there any correlation in such cases to accidents or other troubles? Is traffic continuing unimpeded?
In responding to changes, time is of the essence. One logical approach would be to reduce data flows and associated latency by programming the sensor vehicles to report only when they detect changes from the base map. If the traffic is flowing on the usual three lanes on Broad Street, why add to system “noise” by reporting it? The trouble, though, says Carmera’s Gupta, is that unperceived changes will be missed. “You lose the false negatives,” he says.
Cloud or no cloud?
Updating this new variety of map raises all manner of issues regarding data management. How much of the geo-data, for example, should the vehicle itself interpret and what proportion should be uploaded to cloud-based AI systems?
On one hand, the cloud can harvest from multiple sources, match them with historical patterns, and provide expanded intelligence. But even with ultra-speedy 5G cellular networks expected to be widespread within three years, the back-and-forth of data transfer raises latency concerns. What’s more, since network connections are never guaranteed, autonomous vehicles must be equipped to interpret deviations from the base map for themselves and respond appropriately.
In these early days, most of the mapping companies are focusing on small samples of the earth’s roadways. Naturally, many concentrate on the areas where autonomous driving tests and services are underway. Waymo and Deepmap, for example, are busy in parts of Arizona and California. Carmera, which has agreements with companies that operate delivery fleets, is modeling New York City, San Francisco and retirement villages in Florida, where its partner, Voyage, is operating autonomous shuttle services. The exception is Here Technologies, which is harvesting anonymized data throughout much of Europe and North America from sensors on hundreds of thousands of vehicles manufactured by European automakers.
Monetization matters
One problem, particularly for the venture-backed startups, involves timing. While they’re making large investments now, the widespread use of fully-automated vehicles (SAE Level 4 and 5) may be a decade away, or perhaps longer. In the meantime, they’re searching for intermediate markets for their next-generation maps. “With this transition taking place, how can we use this data to help the driver [now]?” asks Matthew Preyss, a product marketing manager at Here Technologies.
Preyss suggests the new maps will enhance current navigation services, like Waze, Google Maps and TomTom, with more up-to-date road status and course corrections. But the maps could also feed new services, such as augmented reality and parking availability, providing detailed information on the route in both audio and video. The challenge, as always when it comes to maps and human beings, will be to provide helpful data while culling distracting detail.
However, keeping humans in the loop during this period of development also has an advantage: the maps themselves can learn from the drivers’ responses to the data—and focus the AI on significant changes along the route—the ones that demand a response. In this way, we human drivers, over the next decade, will be “educating” the navigation engines poised to replace us.
Author: Stephen Baker
Source: Autonomous Vehicle Engineering
Deepmap 數(shù)據(jù)點“視圖”下的世界。（圖片來源：Deepmap）
Here Technologies 的地圖含有對高精度自動導航至關重要的詳細道路信息。（圖片來源：Here Technologies）
地圖供應商 HereTechnologies 設想了一系列地圖數(shù)據(jù)的衍生應用。這些應用非常有用，比如可以用來尋找空閑停車位。（圖片來源：Here Technologies）
Here Technologies 公司高清地圖軟件下的城市街景。（圖片來源：Here Technogies）