Sir Airffic GoYawancin wannan ya fito ne daga takardun bincike, amma kalmomin an sauƙaƙe su. Tsaya tare da ni.
Matakai huɗu na fahimta
- Gano abu: nemo mutane, motoci da wayoyin wuta a hoto ka zana akwati a kewaye da su.
- Bin sawu: bin inda abin da aka gano zai motsa na gaba.
- Gano wurinka: lissafa nisa da inda ka motsa daga motsin hoto tare da IMU (VIO: visual-inertial odometry).
- Fahimta da umarni: fahimtar “duba bayan gidan mai jan rufi” da haɗa kalmomin da abin da kake gani (VLM: AI da ke sarrafa hoto da harshe tare).
Tashi yayin zana taswira: SLAM da VIO
A ƙarshen shekarun 2000, ɗakunan bincike na jami’o’i sun fara ƙiyasta wurin ƙaramin drone da kyamara ɗaya da IMU. Daga nan komai ya fara.
Hanyoyi kamar SVO, VINS-Mono da ORB-SLAM an wallafa su kuma suka zama tushe na kowa. Bincike a buɗe ya zama kayan haɗin samfura.
Drones na kasuwanci na yau suna iya tashi ƙarƙashin gada da cikin gini saboda wannan aiki da aka tara.
Daga sama, abubuwa suna ƙanƙanta kuma suna karkata
AI da aka horar da hotunan ƙasa tana wahala da kallo daga sama: mutane sun zama ɗigo, motoci kuma siffarsu ta canza.
An gina tarin bayanan sama kamar VisDrone da DOTA domin a horar da AI ta musamman ta sama.
Wanda ya mallaki bayanai shi ke tsara yadda AI za ta yi wayo. Wannan bai canza ba a zamanin foundation models.
Sir Airffic GoBambancin AI yakan dogara ne kan yawan bayanan sama da ka mallaka, fiye da dabarun lissafi. Ka tuna da wannan a babin dabaru.
Ranar da AI ta doke zakarun ’yan Adam
A 2023, drone na AI mai suna Swift, da aka gina a Jami’ar Zurich, ya doke zakarun duniya na tseren drone ’yan Adam. An wallafa aikin a mujallar Nature.
Swift ya ƙiyasta wurinsa da kyamara da IMU kaɗai, kuma ya koyi tashi ta reinforcement learning — yin atisaye sau da yawa a cikin kwaikwayo (simulation).
Wannan a kan hanya ce da aka tsayar. Kai wa ga wannan ƙwarewa a sararin sama na gaske, inda yanayi da cikas ke canzawa, har yanzu ƙalubale ne.
Tseren drone na matakin zakara ta amfani da deep reinforcement learning (UZH)
Shaida: [N124]
Igiyar gaba: gani, tunani da aiki da kalmomi
Daga kusan 2023, manyan samfuran AI da ke sarrafa hoto da harshe tare (foundation models) sun fara aiki a matsayin ƙwaƙwalwar robots.
Binciken drone kan bin umarnin da aka faɗa yayin fahimtar wurin (aerial vision-language navigation) yana ƙaruwa.
Manyan samfura suna da nauyi, kuma gudanar da su da sauri a cikin jirgi har yanzu yana da wuya. Shi ya sa dole a karanta chips — babi na gaba — tare.
Shaida: [N133]
Tambayoyin da ake yawan yi
Waɗanne ayyuka ne ke amfani da computer vision na drone?
Gano tsagewa a gadoji da wayoyin wuta, duba lafiyar amfanin gona, nemo mutane a ayyukan nema da ceto, tsaro da sa ido kan zirga-zirga, da duba wuraren kai kaya su ne amfanin da aka saba.
Mene ne visual-inertial odometry (VIO)?
Hanya ce da ke haɗa motsin hoto da ma’aunin gudu da juyawa na IMU domin lissafa nisan da drone ya motsa. Fasaha ce ta tushe don tashi ba tare da GPS ba.