自動駕駛在挑戰中進化的感知能力

{"type":"doc","content":[{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"滴滴感知大量使用機器學習和深度學習來解決問題,但要解決L4自動駕駛的感知問題,並非只是引入最先進的深度學習模型即可解決。本文歸納出感知能力逐步進化的三個階段。並分析了以下幾大難題給感知帶來的挑戰:深度學習模型本身存在的缺陷、多傳感器需要進行揚長避短的融合、低延遲要求和有限算力間的矛盾、難以準確表徵和處理不確定性。最終,對感知的未來發展提出展望。"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"1. "},{"type":"text","text":"感知準確是無人車算法的基礎"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"感知在自動駕駛中是一個比較直觀的模塊,即根據傳感器(如激光雷達、相機、毫米波雷達)信息對周圍的物體進行理解和抽象,輸出所有交通參與者的位置、速度、大小等屬性,下游模塊(預測、規劃)根據這些抽象信息進行運算和決策;感知如果遺漏了物體,對自車來說是“視而不見”的狀態,會造成嚴重的安全風險。因此,感知信息的準確度是無人車算法的重要基礎。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"2. "},{"type":"text","text":"進化的感知發展路徑"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"自動駕駛的最低要求是能識別到面前的物體,不撞上它。這個要求通過激光雷達就可以實現,因爲它可以準確地估計物體的 3D 位置,並刻畫其輪廓。但真實世界並非都是靜態障礙物,比如在跟車的場景下,前車的速度是我們考量是否需要減速或剎車的重要因素。同時,面對一個在向自車靠近的行人,和一個靜止的行人,自車對他的反應是截然不同的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"第一階段:基於規則的點雲分割算法和物體追蹤"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在第一階段,除了需要識別靜態障礙物,我們也需要識別常見交通參與者(車、行人、自行車)的類別、朝向和速度,以幫助自車做出決策。在深度學習出現之前,其實通過基於規則的點雲分割 \/ 分類算法¹,再加上物體追蹤,就可以做出一個基礎的版本。在這一階段,針對處理不好的問題需要專家設計規則和專門的算法進行處理,然而,許多情景我們難以設計規則處理。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/wechat\/images\/96\/969e8c4014e4d542a6160d7154247713.png","alt":null,"title":null,"style":null,"href":null,"fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"對於並行走的行人很難設計人工規則把兩者分開²"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"第二階段:大規模數據標記和深度學習"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"深度學習的出現和發展大幅度提高了感知的效果。面對規則難以處理的感知任務,我們可以運用大規模數據標記及訓練深度學習模型。我們不再依賴專家針對問題設計算法,而是從大量數據中萃取出經驗和知識。在這一階段,感知算法的設計更加數據驅動。感知通過收集更多的數據,設計更好的模型進行迭代。但深度學習準確率也有上限,且泛化性(在非典型樣本上的表現)、可解釋性都存在問題。因此在自動駕駛這個場景中,深度學習並不是感知唯一的組件。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"第三階段:可擴展性和自學習性的長尾數據處理系統"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"第三階段,需要做更細粒度的識別,以及解決更多長尾問題,如各種奇怪的大車、地上的塑料袋、行人更細粒度的意圖(如是否在打電話)等。這一階段要求系統有更強的可擴展性、自學習性。長尾問題絕對量佔比小,但並不容易解決。其難度可以用九九定律³來刻畫:剩餘 10% 的問題,還需要額外 90% 的時間才能解決。理想情況下,長尾問題應該有自動的流程進入到模型框架中自動進行學習,而不是簡單地靠堆人力來改善這些問題,甚至人過多會使進展變慢⁴。現在學術界在研究的 multi-task learning⁵, AutoML⁶等技術對這一階段的感知發展有極大的啓發。但因爲數據的價值邊際效用遞減,及下文會提到的深度學習的限制,目前業界也還在探索狀態,沒有特別成熟的思路能達到僅靠數據流就能使系統不斷進化的狀態。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/wechat\/images\/af\/afe1bce9ee69a26556b2f9f50f9521bb.png","alt":null,"title":null,"style":null,"href":null,"fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"長尾問題:識別到椎桶在卡車上,無需剎車"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"3. "},{"type":"text","text":"實現完全無人駕駛對感知存在多重挑戰"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"1."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"strong"}],"text":"深度學習模型存在缺陷"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"深度學習模型雖然效果顯著,但最先進的模型的效果也無法達到無人車感知的要求,且深度學習算法缺乏泛化性和可解釋性。許多研究已經證明了深度學習遠不如人類智能通用,如通過加人類無感知的噪聲,就可以誤導模型對結果的分類;對於罕見的數據(如一個穿着很奇怪衣服的人),深度學習也容易犯錯誤。簡單來說,深度學習模型只是以一種生硬的方式在“記憶”訓練數據⁷。而且其記憶能力有限,在模型學習達到飽和後,學習新的樣本可能造成已有能力產生退化⁸。如何結合深度學習模型和基於規則的白盒算法,同時保障感知的召回率和效果,是感知系統面臨的一大挑戰。綜合考量以上缺陷,我們不能僅依賴深度學習模型。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/wechat\/images\/15\/15f9657b4c1a86efbccd4947a35214e4.jpeg","alt":null,"title":null,"style":null,"href":null,"fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"通過添加人眼無法辨別的噪聲,深度學習模型就可以被誤導⁹"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/wechat\/images\/49\/490e9f4c308458df09a4dddf6df53944.png","alt":null,"title":null,"style":null,"href":null,"fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"感知自建的數據標註系統"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"2."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"strong"}],"text":"多傳感器需要進行揚長避短的融合"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"傳感器是感知能力的上限,不同的傳感器有不同的優缺點。激光雷達能對物體輪廓進行較準確的刻畫,同時能準確地得到物體的 3D 位置信息,但缺乏相機所能得到的豐富色彩信息,同時對雨雪天氣較敏感;相機對 3D 位置的估計稍差;而毫米波雷達精度一般,但感知距離遠,且能直接得到物體縱向的速度。下圖更全面地反映了這些優缺點。感知系統需要針對不同的任務,揚長避短地使用多種傳感器信息。同時,多傳感器的融合也對標定的精度、可擴展性提出了較高要求。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/wechat\/images\/4f\/4f121a4e4b0b82bf4c2b1b9ed552751e.png","alt":null,"title":null,"style":null,"href":null,"fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"不同傳感器的優劣勢對比¹⁰"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"3."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"strong"}],"text":"低延遲要求和有限算力間的矛盾"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"無人車是一個實時計算系統,無法像Web後端系統一樣通過增加服務器來進行算力拓展。同時車載系統對能耗、散熱也有約束,這間接約束了感知能使用的算力。在有限算力下部署複雜模型,感知輸出延遲較大,會造成安全隱患及各種問題。我們需要通過模型壓縮、神經結構搜索、代碼優化的方式更巧妙地利用有限的算力資源,達成最佳的效果。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"4."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"strong"}],"text":"難以準確表徵和處理不確定性"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"感知的輸出是帶有不確定性的¹¹,一個近處物體,在無遮擋的情況下,我們對其估計較爲確定;而一個遠處物體,激光雷達打上的點少,我們對它的類別、位置的不確定性都較大。一般來說,我們需要輸出一個最置信的類別和位置信息,但此時該信息的不確定性是極大的,而感知內部或下游往往會直接忽略這種不確定性。如何更好地融合不確定性信息,需要感知內部和下游模塊從底層進行更好的思考。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/wechat\/images\/c9\/c981cf787d5f0503307a886a420208b4.png","alt":null,"title":null,"style":null,"href":null,"fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"不確定性:例如在遮擋嚴重的情況下,我們對物體的類別、位置、速度等信息變得不確定"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"4. "},{"type":"text","text":"充滿挑戰的進階之路"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在外界看來,無人車感知似乎是一個較容易的問題,使用先進的深度學習模型就可以解決得不錯。但無人車對安全性有極高的要求,同時路測出現的場景千奇百怪,這對感知的挑戰是巨大的。無人車感知是一個需要綜合算法、工程、數據的系統工程。算法方面,我們需要針對問題,組合、定製已有算法,同時緊跟學術界\/業界進展,引入新思路。工程方面,我們需要讓系統能夠吞吐更大量的數據;同時不斷完善系統,減少工程師在解決問題需要投入的時間。數據方面,我們需要利用路測數據積累對長尾數據的認知和評估,同時形成模型-路測-數據標註閉環,增加數據的量級和利用率。在這一方面,我們複用了滴滴已有的成熟基礎架構進行二次開發,因此可以較快建立感知的體系。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"一個更高層次的要求是自學習性。如果系統有更好的自學習性,僅需一些數據標註和自動學習,系統就可以適應一個新的環境。當前,我們的感知系統部署到一個環境變化的新城市,還需要投入一些人力進行重新開發和調整。這是一個需要努力的方向,完善的數據和算法架構是重要的基礎。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"目前爲止,無人車感知雖然已經取得了極大進展,但遺留的難題還很多。以上只是一些宏觀層面的思考,但細節決定成敗,每個難題的攻克需要二三素心人的潛心鑽研。期待無人車的未來進階之路上早日迎來技術的躍遷。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"References"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"[1] Montemerlo, M., Becker, J., Bhat, Suhrid., Dahlkamp, H and Dolgov D., Ettinger, Scott., & Haehnel Dirk. (2008). Junior: The Stanford Entry in the Urban Challenge. Journal of Field Robotics, 25(9), 569–597. Retrieved 11 December 2020."}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"from "},{"type":"link","attrs":{"href":"https:\/\/onlinelibrary.wiley.com\/doi\/abs\/10.1002\/rob.20258","title":"","type":null},"content":[{"type":"text","text":"https:\/\/onlinelibrary.wiley.com\/doi\/abs\/10.1002\/rob.20258"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"[2] Held, D., Guillory, D., Rebsamen, B., Thrun, S., & Savarese, S. (2016). A Probabilistic Framework for Real-time 3D Segmentation using Spatial, Temporal, and Semantic Cues. In Robotics: Science and Systems."}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"DOI: 10.15607\/RSS.2016.XII.024"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"[3] Ninety-ninety Rule. Retrieved 11 December 2020."}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"from "},{"type":"link","attrs":{"href":"https:\/\/en.wikipedia.org\/wiki\/Ninety-ninety_rule","title":"","type":null},"content":[{"type":"text","text":"https:\/\/en.wikipedia.org\/wiki\/Ninety-ninety_rule"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"[4] The Mythical Man-Month. Retrieved 11 December 2020."}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"from "},{"type":"link","attrs":{"href":"https:\/\/en.wikipedia.org\/wiki\/The_Mythical_ManMonth?oldformat=true#The_mythical_man-month","title":"","type":null},"content":[{"type":"text","text":"https:\/\/en.wikipedia.org\/wiki\/The_Mythical_ManMonth?oldformat=true#The_mythical_man-month"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"[5] Standley, T., Zamir, A., Chen, D., Guibas, L., Malik, J., & Savarese, S. (2020). Which Tasks Should Be Learned Together in Multi-task Learning?. Retrieved 11 December 2020."}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"from "},{"type":"link","attrs":{"href":"https:\/\/arxiv.org\/abs\/1905.07553","title":"","type":null},"content":[{"type":"text","text":"https:\/\/arxiv.org\/abs\/1905.07553"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"[6] He, X., Zhao, K., & Chu, X. (2020). AutoML: A Survey of the State-of-the-Art. Knowledge-Based Systems, 106622. Retrieved 11 December 2020."}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"from "},{"type":"link","attrs":{"href":"https:\/\/www.sciencedirect.com\/science\/article\/abs\/pii\/S0950705120307516","title":"","type":null},"content":[{"type":"text","text":"https:\/\/www.sciencedirect.com\/science\/article\/abs\/pii\/S0950705120307516"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"[7] Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., & Fergus, R. (2013). Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199. Retrieved 11 December 2020."}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"from "},{"type":"link","attrs":{"href":"https:\/\/arxiv.org\/abs\/1312.6199","title":"","type":null},"content":[{"type":"text","text":"https:\/\/arxiv.org\/abs\/1312.6199"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"[8] Goodfellow, I. J., Mirza, M., Xiao, D., Courville, A., & Bengio, Y. (2013). An empirical investigation of catastrophic forgetting in gradient-based neural networks. arXiv preprint arXiv:1312.6211. Retrieved 11 December 2020."}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"from "},{"type":"link","attrs":{"href":"https:\/\/arxiv.org\/abs\/1312.6211","title":"","type":null},"content":[{"type":"text","text":"https:\/\/arxiv.org\/abs\/1312.6211"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"[9] Goodfellow, I. J., Shlens, J., & Szegedy, C. (2014). Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572. Retrieved 11 December 2020, from "},{"type":"link","attrs":{"href":"https:\/\/arxiv.org\/abs\/1412.6572","title":"","type":null},"content":[{"type":"text","text":"https:\/\/arxiv.org\/abs\/1412.6572"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"[10] Fused Sensing vs. Sensor Fusion. Retrieved 11 December 2020."}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"from "},{"type":"link","attrs":{"href":"https:\/\/www.outsight.tech\/technology\/fused-sensing-vs.-sensor-fusion","title":"","type":null},"content":[{"type":"text","text":"https:\/\/www.outsight.tech\/technology\/fused-sensing-vs.-sensor-fusion"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"[11] Feng, D., Harakeh, A., Waslander, S., & Dietmayer, K. (2020). A Review and Comparative Study on Probabilistic Object Detection in Autonomous Driving. arXiv preprint arXiv:2011.10671. Retrieved 11 December 2020."}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"from "},{"type":"link","attrs":{"href":"https:\/\/arxiv.org\/abs\/2011.10671","title":"","type":null},"content":[{"type":"text","text":"https:\/\/arxiv.org\/abs\/2011.10671"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"本文作者"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Milton,滴滴自動駕駛高級專家算法工程師。"}]}]}
發表評論
所有評論
還沒有人評論,想成為第一個評論的人麼? 請在上方評論欄輸入並且點擊發布.
相關文章