Ukwakha izimoto ezizishayelayo ezingathatha izinqumo ezifuze abantu kube wumgomo wezinkampani ezifana neWaymo, GM Cruise, Uber nezinye. I-Intel Mobileye inikezela ngemodeli yezibalo ye-Responsibility-Sensitive Safety (RSS), le nkampani eyichaza ngokuthi "indlela evamile" ebonakala ngokuhlela umshayeli we-autopilot ukuthi aziphathe ngendlela "enhle", njengokunikeza ezinye izimoto ilungelo lokuhamba. . Ngakolunye uhlangothi, i-NVIDIA ithuthukisa i-Safety Force Field, ubuchwepheshe bokuthatha izinqumo obusekelwe ohlelweni obuqapha izenzo ezingaphephile zabasebenzisi bomgwaqo abazungezile ngokuhlaziya idatha evela kuzinzwa zezimoto ngesikhathi sangempela. Manje ithimba lososayensi abavela e-Massachusetts Institute of Technology (MIT) selijoyine lolu cwaningo futhi laphakamisa indlela entsha esekelwe ekusetshenzisweni kwamamephu afana ne-GPS kanye nedatha ebonakalayo etholwe kumakhamera afakwe emotweni ukuze umshayeli wemoto akwazi ukuzulazula endaweni engaziwa. imigwaqo efana nomuntu.indlela.
Abantu banekhono elimangalisayo ekushayeleni izimoto emigwaqweni abangakaze bayihambe phambilini. Simane siqhathanise esikubonayo esisizungezile nalokho esikubona emishinini yethu ye-GPS ukuze sibone ukuthi sikuphi nokuthi kudingeka siye kuphi. Izimoto ezizishayelayo, ngakolunye uhlangothi, zikuthola kunzima kakhulu ukuzulazula ezingxenyeni ezingaziwa zomgwaqo. Endaweni entsha ngayinye, umshayeli wendiza ozenzakalelayo udinga ukuhlaziya ngokucophelela umzila omusha, futhi ngokuvamile amasistimu okulawula okuzenzakalelayo ancike kumamephu e-3D ayinkimbinkimbi abahlinzeki abawalungiselela kusengaphambili.
Ephepheni elethulwe kuleli sonto eNgqungqutheleni Yamazwe Ngamazwe Ngamarobhothi Nokuzenzakalela, abacwaningi be-MIT bachaza uhlelo lokushayela oluzimele "olufundayo" futhi lukhumbula amaphethini okwenza izinqumo zomshayeli njengoba bezulazula emigwaqweni endaweni yedolobha elincane besebenzisa idatha kuphela. amakhamera kanye nemephu elula efana ne-GPS. Umshayeli oqeqeshiwe angakwazi ukushayela imoto engashayeli endaweni entsha ngokuphelele, elingisa ukushayela komuntu.
Njengomuntu, i-autopilot iphinde ithole noma yikuphi ukungafani phakathi kwemephu yayo nezici zomgwaqo. Lokhu kusiza isistimu inqume uma ngabe indawo yayo emgwaqeni, izinzwa, noma imephu ayilungile ukuze ikwazi ukulungisa indlela yemoto.
Ukuze aqale aqeqeshe isistimu, u-opharetha ongumuntu ushayele i-Toyota Prius ezenzakalelayo efakwe amakhamera amaningi kanye nesistimu yokuzulazula ye-GPS eyisisekelo ukuze kuqoqwe idatha emigwaqweni yasendaweni yasemadolobheni, okuhlanganisa izakhiwo zomgwaqo nezithiyo ezihlukahlukene. Isistimu ibe isishayela imoto ngempumelelo ngomzila ohlelwe ngaphambilini kwenye indawo eyihlathi ehloselwe ukuhlola izimoto ezizimele.
"Ngohlelo lwethu, awudingi ukuqeqeshwa kuyo yonke imigwaqo kusenesikhathi," kusho umlobi wocwaningo u-Alexander Amini, umfundi ophothule eMIT. "Ungalanda imephu entsha yemoto yakho ukuze ihambe emigwaqweni engakaze ibonwe ngaphambili."
"Umgomo wethu uwukwenza ukuzulazula okuzimele okukwazi ukumelana nokushayela ezindaweni ezintsha," kunezela umbhali-ndawonye uDaniela Rus, umqondisi weComputer Science and Artificial Intelligence Laboratory (CSAIL). βIsibonelo, uma siqeqesha imoto ezimele ukuthi ishayele endaweni yasemadolobheni njengemigwaqo yaseCambridge, uhlelo kumele futhi lukwazi ukushayela kahle ehlathini, noma lungakaze luyibone indawo enjalo phambilini.β
Amasistimu okuzulazula endabuko acubungula idatha yenzwa ngamamojula amaningi alungiselwe imisebenzi efana nokwenza kwasendaweni, ukumepha, ukutholwa kwezinto, ukuhlela ukunyakaza nokuqondisa. Sekuyiminyaka, iqembu likaDaniela lithuthukisa amasistimu okuzulazula asuka ekupheleni acubungula idatha yenzwa futhi alawule imoto ngaphandle kwesidingo sanoma yimaphi amamojula akhethekile. Kodwa-ke, kuze kube manje, lawa mamodeli asetshenziswe ngokuqinile ekuhambeni okuphephile emgwaqeni, ngaphandle kwenjongo yangempela. Emsebenzini omusha, abacwaningi balungise isistimu yabo yokuphela-kuya-ekupheleni ukunyakaza okuya endaweni okuyiwa kuyo endaweni engaziwa ngaphambilini. Ukuze benze lokhu, ososayensi baqeqeshe i-autopilot yabo ukuze ibikezele ukusatshalaliswa kwamathuba agcwele kuyo yonke imiyalo engase ibe khona yokulawula nganoma yisiphi isikhathi lapho ushayela.
Uhlelo lusebenzisa imodeli yokufunda yomshini ebizwa ngokuthi i-convolutional neural network (CNN), evamise ukusetshenziselwa ukubonwa kwesithombe. Ngesikhathi sokuqeqeshwa, uhlelo lubheka ukuziphatha kokushayela komuntu. I-CNN ihlobanisa amasondo esiteringi ngokujika komgwaqo, ewubuka ngamakhamera nakumephu yawo encane. Ngenxa yalokho, isistimu ifunda imiyalo yokuqondisa okungenzeka kakhulu yezimo ezihlukahlukene zokushayela, njengemigwaqo eqondile, izimpambanondlela ezine noma i-T-junctions, izimfoloko namajika.
"Ekuqaleni, ezimpambanweni zomgwaqo, kunezindlela eziningi ezahlukene imoto engajika ngayo," kusho uRus. βImodeli iqala ngokucabanga ngazo zonke lezi zikhombisi-ndlela, futhi njengoba i-CNN ithola imininingwane eminingi mayelana nokuthi abantu benzani ezimeni ezithile emgwaqeni, izobona ukuthi abanye abashayeli bajikela kwesokunxele abanye bajike kwesokudla, kodwa akekho oqonda ngqo. . Ukuhamba phambili akubhekwa njengento engenzeka, futhi imodeli iphetha ngokuthi kuma-T-junctions ingahamba kwesokunxele noma kwesokudla kuphela.
Ngenkathi ushayela, i-CNN iphinda ikhiphe izici zomgwaqo ezibukwayo kumakhamera, okuyivumela ukuthi ibikezele izinguquko ezingaba khona zemizila. Isibonelo, ikhomba uphawu olubomvu lokuma noma umugqa ophukile ohlangothini lomgwaqo njengezimpawu zempambana-mgwaqo ezayo. Ngaso sonke isikhathi, isebenzisa ukusatshalaliswa kwamathuba abikezelwe emiyalo yokulawula ukuze ikhethe umyalo ofanele kakhulu.
Kubalulekile ukuqaphela ukuthi, ngokusho kwabacwaningi, i-autopilot yabo isebenzisa amamephu okulula kakhulu ukuwagcina nokuwacubungula. Amasistimu okulawula okuzenzakalelayo ngokuvamile asebenzisa amamephu e-lidar, athatha cishe u-4000 GB wedatha ukugcina idolobha lase-San Francisco. Endaweni entsha ngayinye, imoto kufanele isebenzise futhi idale amamephu amasha, okudinga inani elikhulu lenkumbulo. Ngakolunye uhlangothi, imephu esetshenziswa i-Autopilot entsha ihlanganisa umhlaba wonke kuyilapho ithatha amagigabhayithi angu-40 kuphela edatha.
Ngesikhathi sokushayela ngokuzenzakalelayo, isistimu iphinda iqhathanise idatha yayo ebonakalayo nedatha yemephu futhi ihlabe umkhosi noma yikuphi ukungafani. Lokhu kusiza imoto ezimele ukuthi ibone kangcono ukuthi ikuphi emgwaqeni. Futhi lokhu kuqinisekisa ukuthi imoto ihlala endleleni ephephe kakhulu, ngisho noma ithola ulwazi lokokufaka olushayisanayo: uma, yithi, imoto ihamba emgwaqeni oqondile ongajiki, futhi i-GPS ikhombisa ukuthi imoto kufanele ijike kwesokudla, imoto ukwazi ukuya ngqo noma ukuma.
βEzweni langempela, izinzwa ziyehluleka,β kusho u-Amini. "Sifuna ukwenza isiqiniseko sokuthi i-autopilot yethu iyakwazi ukumelana nokwehluleka kwezinzwa okuhlukahlukene ngokwakha uhlelo olukwazi ukuthola noma yimaphi amasignali omsindo futhi lisahambe ngendlela efanele emgwaqeni."
Source: 3dnews.ru