Airffic InstituteAERIAL PHYSICAL AI
Imenyu ☰

Aerial Physical AI Institute

I-Vision AI: indlela idrone ezwisisa ngayo lokho ekubonako

Ukubona akwaneli bona undize. Ingqondo itlhoga ukuzwisisa lokho ekubonako.

Ipendulo efitjhani: I-Vision AI yamadrone ikhule ngamagadango ama-4: ① ukufumana izinto (detection), ② ukuzilandela (tracking), ③ ukwazi bona ukuphi (SLAM, VIO), ④ ukuzwisisa imiyalo ngamezwi (foundation models). Simi emnyango wegadango ④.

Kuvuselelwe: · 3 imizuzu yokufundaIndlela esenze ngayo irhubhululo →

Sir Airffic GoOkunengi lapha kuvela emaphepheni werhubhululo, kodwana amezwi alula. Hlala nami.

Amagadango ama-4 wokuzwisisa

  1. Ukufumana (detection): fumana abantu, iimotoro neentambo zombane esithombeni bese uzibiyela ngamabhoksi.
  2. Ukulandela (tracking): landela lapho into efunyenweko iya khona ngemva kwalokho.
  3. Ukuzifumana: bala bona uhambe kangangani begodu ukuphi ngokuhamba kwesithombe ne-IMU (VIO: visual-inertial odometry).
  4. Ukuzwisisa nomyalo: zwisisa "hlola ngemva kwendlu enophahla olubomvu" bese uhlanganisa amezwi nalokho okubonako (VLM: i-AI esebenza ngesithombe nelimi ndawonye).

Ukundiza ngesikhathi wenza imephu: i-SLAM ne-VIO

Ekupheleni kweminyaka yabo-2000, amalabhorethri wamayunivesithi athoma ukulinganisa indawo yedrone encani ngekhamera yinye ne-IMU. Koke kwathoma lapho.

Iindlela ezinjenge-SVO, VINS-Mono ne-ORB-SLAM zakhitjhwa zaba sisekelo esabelwanako. Irhubhululo elivulekileko laba ziingcenye zemikhiqizo.

Amadrone webhizinisi wanamhlanje akghona ukundiza ngaphasi kwamabhuroho nangaphakathi ngebanga lomsebenzi lo obuthelelweko.

Ubufakazi: [P01] [P02] [P08] [P09] [P21] [P15]

Ukusuka phezulu, izinto zibonakala zincani begodu zigwegwile

I-AI efundiswe ngezithombe zaphasi ayisebenzi kuhle nayibona kusuka phezulu: abantu baba ngamachaphazana, iimotoro zitjhentjha isimo.

Iinqoqo zedatha zemoyeni ezinjenge-VisDrone ne-DOTA zakhiwa bona i-AI yesibhakabhaka ikghone ukufundiswa.

Loyo ophethe idatha uqunta bona i-AI ihlakanipha kangangani. Lokho akukatjhentjhi nesikhathini sama-foundation model.

Sir Airffic GoUmehluko we-AI kanengi uvela ebunengini bedatha yemoyeni onayo, ukudlula ama-algorithm. Khumbula lokhu esahlukweni seqhinga.

Ubufakazi: [P14] [P23] [P24]

Ilanga i-AI ehlula ngalo iingwazi zabantu

Ngo-2023 idrone ye-AI i-Swift, eyakhiwe e-University of Zurich, yahlula iingwazi zephasi zabantu ekugijimiseni amadrone. Umsebenzi lo wakhitjhwa ku-Nature.

I-Swift yalinganisa indawo yayo ngekhamera ne-IMU kwaphela, begodu yafunda ukundiza nge-reinforcement learning — ukuzijayeza kanengi ku-simulation.

Lokhu bekwenzeka emzileni ongatjhentjhiko. Ukulingana nalelo khono esibhakabhakeni samambala, esinesimo sezulu esitjhentjhako nezithiyo, kuseyinselelo.

Ukugijimisa amadrone ezingeni leengwazi nge-deep reinforcement learning (UZH)

Ividiyo yedrone ye-AI i-Swift ekhitjhwe yi-Robotics and Perception Group ye-University of Zurich. YouTube ↗ · Akukho okulayitjhwa ku-YouTube ungakadlali ividiyo

Ubufakazi: [N124]

Igagasi elilandelako: bona, cabanga begodu wenze ngamezwi

Kusukela ngo-2023, amamodeli amakhulu we-AI asebenza ngesithombe nelimi ndawonye (ama-foundation model) athoma ukusebenza njengengqondo yamarobhothi.

Irhubhululo lamadrone alandela imiyalo ekhulunywako ngesikhathi azwisisa indawo (aerial vision-language navigation) liyakhula.

Amamodeli amakhulu anesisindo, begodu ukuwasebenzisa msinyana phakathi kwedrone akukalula. Kungalokho ama-chip — isahluko esilandelako — kufanele afundwe ndawonye nalokhu.

Ubufakazi: [N133]

Imibuzo ebuzwa kanengi

Ngimiphi imisebenzi esebenzisa ukubona kwekhompyutha kwamadrone?

Ukufumana imifantu emabhurohweni nakiintambo zombane, ukuhlola ipilo yezitjalo, ukufuna abantu ekuhlengeni, ukuvikela nokuqala ithrafiki, nokuhlola iindawo zokuletha iimpahla kujayelekile.

Iyini i-visual-inertial odometry (VIO)?

Yindlela ehlanganisa ukuhamba kwesithombe nezilinganiso ze-IMU zejubane nokujikeleza bona ibale ukuthi idrone ihambe kangangani. Bubuchwepheshe obusisekelo bokundiza ngaphandle kwe-GPS.