Airffic InstituteAERIAL PHYSICAL AI
Imenyu ☰

Aerial Physical AI Institute

Vision AI: indlela idrone eyiqonda ngayo into eyibonayo

Ukubona akwanelanga ukuze ubhabhe. Ingqondo kufuneka iyiqonde into eyibonayo.

Impendulo emfutshane: I-Vision AI yeedrone iphuhle ngamanqanaba amane: ① ukufumana izinto (detection), ② ukuzilandela (tracking), ③ ukwazi apho ukhoyo (SLAM, VIO), ④ ukuqonda imiyalelo ngamagama (foundation models). Simi emnyango wenqanaba ④.

Kuhlaziywe: · Imizuzu eyi-3 yokufundaIndlela esiphande ngayo →

Sir Airffic GoUninzi lwale nto luvela kumaphepha ophando, kodwa amagama agcinwe elula. Hamba nam.

Amanqanaba amane okuqonda

  1. Ukufumana (detection): fumana abantu, iimoto neentambo zombane emfanekisweni uze uzobe iibhokisi ezizijikelezileyo.
  2. Ukulandela (tracking): landela apho into efunyenweyo iya khona emva koko.
  3. Ukuzifumana: bala ukuba uhambe kangakanani nokuba uye phi ngentshukumo yomfanekiso kunye ne-IMU (VIO: visual-inertial odometry).
  4. Ukuqonda nomyalelo: qonda “jonga emva kwendlu enophahla olubomvu” uze uqhagamshele amagama noko ukubonayo (VLM: i-AI ephatha imifanekiso nolwimi kunye).

Ukubhabha ngelixa usenza imaphu: SLAM ne-VIO

Ngasekupheleni kweminyaka yoo-2000, iilebhu zeeyunivesithi zaqala ukuqikelela indawo yedrone encinci ngekhamera enye ne-IMU. Yonke into yaqala apho.

Iindlela ezifana ne-SVO, VINS-Mono ne-ORB-SLAM zapapashwa zaba ziziseko ekwabelwana ngazo. Uphando oluvulekileyo lwaba ngamalungu eemveliso.

Iidrone zorhwebo zanamhlanje zingabhabha phantsi kweebhulorho nangaphakathi ngenxa yalo msebenzi uqokelelweyo.

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

Ukusuka phezulu, izinto zibonakala zincinci kwaye zigoso

I-AI eqeqeshwe ngeefoto ezithathwe emhlabeni isokola xa ijonga ukusuka phezulu: abantu baba ngamachaphaza, iimoto zitshintsha imilo.

Iiseti zedatha zasemoyeni ezifana neVisDrone neDOTA zakhiwa ukuze kuqeqeshwe i-AI eyenzelwe isibhakabhaka.

Lowo unedatha ugqiba ukuba i-AI iba krelekrele kangakanani. Oko akutshintshanga nakwixesha lee-foundation models.

Sir Airffic GoUmahluko kwi-AI uhlala uxhomekeke kubuninzi bedatha yasemoyeni onayo, ngaphezu kwee-algorithm. Khumbula oku kwisahluko seqhinga.

Ubungqina: [P14] [P23] [P24]

Usuku i-AI yoyisa ngalo iintshatsheli zabantu

Ngowama-2023 idrone ye-AI ebizwa ngokuba yiSwift, eyakhiwe kwiYunivesithi yaseZurich, yoyisa iintshatsheli zehlabathi zabantu kugqatso lweedrone. Lo msebenzi wapapashwa kwiNature.

ISwift yaqikelela indawo yayo ngekhamera ne-IMU kuphela, kwaye yafunda ukubhabha nge-reinforcement learning — iziqhelisa amaxesha amaninzi kwi-simulation.

Oku kwakusenzeka kwindlela engatshintshiyo. Ukufikelela kobo buchule kwisibhakabhaka sokwenyani, nemozulu nemiqobo etshintshayo, kusengumngeni ovulekileyo.

Ugqatso lweedrone kwinqanaba leentshatsheli nge-deep reinforcement learning (UZH)

Ividiyo yedrone ye-AI iSwift epapashwe liqela leRobotics and Perception leYunivesithi yaseZurich. YouTube ↗ · Akukho nto ilayishwa kwi-YouTube ude ucofe u-play

Ubungqina: [N124]

Iliza elilandelayo: bona, cinga kwaye wenze ngokwamagama

Ukususela malunga nowama-2023, iimodeli ezinkulu ze-AI eziphatha imifanekiso nolwimi kunye (foundation models) zaqala ukusebenza njengengqondo yeerobhothi.

Uphando lweedrone ezilandela imiyalelo ethethwayo ngelixa ziqonda imeko (aerial vision-language navigation) luyakhula.

Iimodeli ezinkulu zinzima, kwaye ukuziqhuba ngokukhawuleza kwidrone kusenzima. Yiyo loo nto iitshiphu — isahluko esilandelayo — kufuneka zifundwe kunye nale.

Ubungqina: [N133]

Imibuzo ebuzwa rhoqo

Yeyiphi imisebenzi esebenzisa computer vision yeedrone?

Ukufumana iintanda kwiibhulorho nakwiintambo zombane, ukuhlola impilo yezityalo, ukufumana abantu ekukhangeleni nasekuhlanguleni, ukhuseleko nokubeka iliso kwizithuthi, nokujonga iindawo zokuhambisa zezona zisetyenziswa kakhulu.

Yintoni i-visual-inertial odometry (VIO)?

Yindlela edibanisa intshukumo yomfanekiso nemilinganiselo ye-IMU yokunyuka kwesantya nokujikeleza ukubala ukuba idrone ihambe kangakanani. Bubuchwepheshe obusisiseko bokubhabha ngaphandle kweGPS.