Abaphandi abavela kwiMicrosoft kunye neYunivesithi yaseCentral China indlela entsha yokusebenza ephezulu yokulandelela izinto ezininzi kwividiyo usebenzisa iteknoloji yokufunda ngomatshini - i-FairMOT (i-Fair Multi-Object Tracking). Ikhowudi kunye nokuphunyezwa kwendlela esekelwe kwiPytorch kunye neemodeli eziqeqeshiwe kwiGitHub.
Uninzi lweendlela ezikhoyo zokulandelela izinto zisebenzisa izigaba ezibini, nganye iphunyezwe yinethiwekhi ye-neural eyahlukileyo. Isigaba sokuqala siqhuba imodeli yokumisela indawo yezinto ezinomdla, kwaye isigaba sesibini sisebenzisa imodeli yokukhangela yombutho esetyenziselwa ukuchonga kwakhona izinto kunye nokunamathisela amahange kubo.
I-FairMOT isebenzisa ukuphunyezwa kwenqanaba elinye ngokusekelwe kuthungelwano lwe-neural convolutional (, Deformable Convolutional Network), ekuvumela ukuba ufezekise ukwanda okubonakalayo kwisantya sokulandelela into. I-FairMOT isebenza ngaphandle kweehange, isebenzisa indlela yokuchonga kwakhona ukumisela i-offsets yamaziko ezinto kwimephu yezinto ezichanekileyo. Ngokunxuseneyo, iprosesa iphunyeziwe evavanya iimpawu zomntu ngamnye zezinto ezinokuthi zisetyenziswe ukuqikelela ubuni bazo, kwaye imodyuli ephambili yenza ukuhlangana kwezi mpawu ukuxhaphaza izinto zezikali ezahlukeneyo.
Ukuqeqesha imodeli kwi-FairMOT, udibaniso lweedatha ezintandathu zoluntu zokukhangela abantu kunye nokukhangela zisetyenziswe (ETH, CityPerson, CalTech, MOT17, CUHK-SYSU). Imodeli yavavanywa kusetyenziswa iiseti zovavanyo lweevidiyo , , ΠΈ ebonelelwa yiprojekthi kunye nokugubungela iimeko ezahlukeneyo, ukunyakaza kwekhamera okanye ukujikeleza, ii-angles zokujonga ezahlukeneyo. Uvavanyo lwabonisa ukuba
FairMOT imodeli ekhuphisana ngokukhawuleza ΠΈ xa ivavanywa kwizakhelo ze-30 ngesekondi nganye kwimijelo yevidiyo, ebonisa ukusebenza okwaneleyo ukuhlalutya imijelo yevidiyo eqhelekileyo kubhabho.
umthombo: opennet.ru