isiZulu
Indlela Yokuhlola Nokuvumela Ukusebenza Bekhamera Yokubona Eyifakiwe
Kwadalwa ngo 03.11
Amakhamera e-Embedded vision aseke abe umgogodla wezinhlelo ezibalulekile ezindaweni ezahlukahlukene—kusukela ezimotweni ezizihambelayo, ukulawulwa kwekhwalithi ezimbonini, kuya emadolobheni ahlakaniphile, nasezithombeni zezokwelapha. Ngokungafani namakhamera abantu abavamile ukuzisebenzisa, ukusebenza kwawo kuthinta ngqo ukuphepha, ukusebenza kahle, nokwethembeka kwezinhlelo. Nokho, izindlela zokuhlola ezindala zivame ukugxila kwezinto ezithile zehadiwe (isibonelo, isinqumo) futhi zishaye indiva ukusebenzisana phakathi kwehadiwe, isofthiwe, nezindawo zangempela. Ukuqinisekisa ngempela ukuba namandla, indlela ebanzi, ehambisana nezigameko iyadingeka. Kulo mhlahlandlela, sizochaza uhlaka olusha lokuhlola olunezendlalelo ezi-3 olubhekana nezimfuno zesimanje.umbono ofakelweyo izinselelo, ngezilinganiso ezingasebenzi, amathuluzi, nezinqubo ezinhle kakhulu zokuqinisekisa ukusebenza ngaphezu kwedathasethi.
1. Uhlaka Oluyinhloko: Ukudlula Izilinganiso Ezihlukene
Izivivinyo eziningi zamakhamera okubona afakwe ngaphakathi ziyayeka ekuhloleni izici eziyisisekelo njengesinqumo noma isivinini sezithombe, kodwa izinhlelo ezifakwe ngaphakathi zisebenza ezindaweni eziguquguqukayo, ezinomkhawulo wezinsiza. Isu lokuqinisekisa oluphelele kumele lihlanganise izendlalelo ezintathu ezincikene: ukwethembeka kwezingxenyekazi zekhompyutha, ukunemba kwezibalo, nokwazi ukuzivumelanisa nezimo zangempela. Lolu hlelo luqinisekisa ukuthi ikhamera yakho ayisebenzi kahle nje elabhorethri—iletha imiphumela ehambisanayo ezimweni ezingaphandle lapho izosebenza khona ngempela, kungaba phansi kwefektri enothuli, umgwaqo omkhulu osheshayo, noma idivayisi ye-IoT enamandla aphansi.
2. Ukuhlolwa Kwesendlalelo Sehadiwe: Ngaphezu Kokusombulula Nengxenye Yesikhathi
I-hardware iyisisekelo sokusebenza kombono ofakelweyo, kodwa ukuhlolwa kufanele kudlule izibonelo ezifakwe kudathasethi. Gxila ezilinganisweni ezithinta ngqo ukusetshenziswa kwezinto zangempela, ikakhulukazi ezinhlelweni ezifakelwe ezikhawulelwe izinsiza.
Okokuqala, ububanzi obuguquguqukayo nokusebenza ekukhanyeni okuphansi kuyinto okungadingekile ezimweni eziningi zezimboni nezimoto. Kunokuba ulinganise kuphela isinqumo esikhulu, hlola ukuthi ikhamera igcina imininingwane ezindaweni ezinokungafani okuphezulu (isibonelo, ilanga eliqondile nezithunzi) usebenzisa izilinganiso zobubanzi obuguquguqukayo njengezindinganiso ze-dB. Ezimweni zokukhanya okuphansi, linganisa isilinganiso sesiginali-kuya-umsindo (SNR) kumazinga ahlukene e-ISO—gqamisa i-SNR engaphezu kwe-30dB ukuze uthole izithombe ezisebenzisekayo ezindaweni ezimnyama. Amathuluzi afana ne-Keysight’s Image Quality Analyzer angakwazi ukwenza lezi zilinganiso ngokuzenzakalelayo, aqinisekise ukungaguquguquki kuyo yonke imizamo yokuhlola.
Ukusetshenziswa kwamandla kahle kuyisici esibalulekile sehadiwe esivame ukunganakwa. Amakhamera afakiwe avame ukusebenza ngamandla ebhethri noma izinto zikagesi zezimboni ezabiwe, ngakho-ke ukusetshenziswa kwamandla okuphezulu kunganciphisa ukuguquguquka kokufakwa. Sebenzisa i-precision power analyzer ukukala ukudonswa kwamandla ngesikhathi sokungasebenzi, ukuthwebula, nezigaba zokucubungula. Ngokwesibonelo, i-NVIDIA Jetson AGX Orin, inkundla ethandwayo yokubona efakiwe, iletha ukusetshenziswa kwamandla okuhlukahlukene (14.95W kuya ku-23.57W) kuye ngemodeli nomsebenzi, okugcizelela isidingo sokuhlola amandla kanye nokusebenza. Zibekele izinzwa ezisebenzisa amandla kahle njenge-Prophesee's GenX320, inzwa yokubona yomcimbi encane kunazo zonke futhi esebenza kahle kakhulu emhlabeni, esiza ukunciphisa ukudonswa kwamandla ohlelo lonke ngenkathi igcina ukusebenza.
Ekugcineni, ukuhlolwa kokumelana kwemishini nezimo zemvelo kubalulekile ekusebenzeni kwezimboni nasezindaweni zangaphandle. Hlola ukusebenza kwekhamera ngaphansi kwezinga lokushisa eliphezulu, umswakama, nokuhudula usebenzisa izikhala zemvelo—qinisekisa ukuthi ihlangabezana nezindinganiso ze-IP futhi igcine ikhwalithi yesithombe nokuxhumana ezimeni ezinzima. Le nyathelo ivimbela ukuwa kwezindleko emkhakheni okungase kuphuthume ukuhlolwa kwelabhu okujwayelekile.
3. Ukuhlolwa Kwe-Algoithm Layer: Ukunemba Kuhlangana Ne-Efficiency Yesikhathi Sangempela
Amakhamera e-Embedded vision axhomeke kwi-AI/ML algorithms ezisetshenziswa kumadivayisi emisebenzi efana nokutholwa kwezinto, ukuhlukaniswa kwemigqa, nokuhlola izimo—ukuhlola lezi algorithms kudinga ukuhlela ukunemba nokusebenza kwesikhathi sangempela, okukhombisa izilinganiso ezimbili ezivame ukuncintisana.
Qala ngezilinganiso zokunemba eziklanyelwe ukusetshenziswa kwakho. Ngokuthola izinto, sebenzisa i-mean Average Precision (mAP) ukukala ukuthi uhlelo lokusebenza luhlonza futhi lubeke izinto kahle kangakanani ezindaweni eziningi. Ngokuhlukanisa izithombe, gxila ku-top-1 ne-top-5 accuracy. Sebenzisa amasethi edatha ezibonelo njenge-COCO (Common Objects in Context) noma i-ImageNet, kodwa futhi dala amasethi edatha wangokwezifiso afana nokusetshenziswa kwakho - abasebenzisi bezimboni bangafaka izingxenye ezinokukhubazeka, kanti abasebenzisi bezimoto kufanele banikeze izinto ezibalulekile zabahamba ngezinyawo nezimoto. Imiphumela ye-NVIDIA Jetson's MLPerf inference ibonisa ukuthi izinhlelo ezilungiselelwe (nge-TensorRT) zingathuthukisa kakhulu ukunemba nesivinini; isibonelo, ukuhlukanisa izithombe okusekelwe ku-ResNet ku-Jetson AGX Orin kuletha izibonelo ezingama-6423.63 ngomzuzwana kumodi engaxhunyiwe ku-inthanethi, okubonisa umthelela wokulungiswa kwezinhlelo ekusebenzeni.
Ukusebenza kwesikhathi sangempela kulinganiswa nge-latency (isikhathi kusukela ekubambeni kuya ekuphumeni) kanye ne-frame rate (FPS). Izicelo ezibucayi ngesikhathi njengokushayela okuzenzakalelayo noma i-robotics, i-latency kufanele ibe ngaphansi kwe-100ms—ngisho nokubambezeleka okuncane kungaholela emaphutheni amabi kakhulu. Sebenzisa amathuluzi afana ne-OpenCV's video capture API noma i-Prophesee's Metavision SDK ukukala i-latency; amakhamera asekelwe kwisenzakalo se-Prophesee athola i-latency engaphansi kwe-150μs ku-1k lux, abeka ibha ephezulu yokusebenza kwesikhathi sangempela. I-frame rate kufanele ihambisane (hhayi nje ukusebenza okuphezulu)—hlola ngaphansi kwemisebenzi ehlukahlukene ukuqinisekisa ukuthi ikhamera ayilahli amafreyimu lapho icubungula izigcawu eziyinkimbinkimbi.
Ukuphuculwa kwe-Edge AI kuyisici esibalulekile sokuhlola ama-algorithm. Amakhamera afakwe ngaphakathi anamandla okucubungula alinganiselwe, ngakho hlola ukusebenza kwe-algorithm kumadivayisi afanele (isb., i-Jetson Orin NX, i-Raspberry Pi) kunokuba kube nje kwi-PC enamandla. Amathuluzi afana ne-TensorRT (kwamadivayisi e-NVIDIA) noma i-TensorFlow Lite (ukweseka amapulatifomu ahlukene) aphucula imodeli zokuhlola ezisebenzelayo, futhi ukuhlola ngalezi zithuluzi kuqinisekisa ukuthi i-algorithm yakho iyasebenza kahle ekukhiqizeni.
4. Ukuhlola Ukuguquguquka Kwezimo: Ukuqinisekiswa Okuphezulu
Ingxenye ehlakaniphile kakhulu yokuhlola ukubona okwakhiwe kwanamuhla ukuqinisekisa ukusebenza ezimeni zangempela—hhayi kuphela ez laboratory ezilawulwayo. Le ngxenye iqinisekisa ukuthi ikhamera isebenza njengoba kulindelekile ezindaweni ezizoyisebenzisa.
Ezinkambeni ezisebenzisa ikhamera eyodwa, hlola ezimweni ezihlukahlukene zokukhanya (ukukhanya okuphansi, ilanga eliqondile, ukukhanya kwangemuva) nezizinda (ezinobuningi, ezihlukahlukene, ezihambayo). Ngokwesibonelo, ikhamera yezimboni kufanele ikwazi ukuthola amaphutha ngokucacile noma ngabe indawo yefekthri ikhanyiswe kahle noma imnyama. Sebenzisa izifanisi zemvelo ukuze uphinde ubuyekeze lezi zimo, futhi ukale ukuthi ukucaca nesivinini sokuqhafaza siguquka kanjani—amakhamera aqinile azogcina ukusebenza ngaphakathi kwemingcele eyamukelekayo.
Ukuhlolwa kokubambisana kwamakamela amaningi kubalulekile ekusetshenzisweni okukhulu okufana namadolobha ahlakaniphile noma ukuzenzakalela kwesitolo. Qinisekisa ukuthi amakamela asebenza kanjani ndawonye ukulandelela izinto, ukuhlanganisa izithombe ezibanzi, noma ukwabelana ngedatha. Izilinganiso eziyinhloko zihlanganisa ukunemba kokulandelela okuhlosiwe (izinga lokulahlekelwa elingaphansi kuka-5% ngokuya ngamazinga emboni), ikhwalithi yokuhlanganisa izithombe ezibanzi (imiphetho engaphansi kwama-pixel angu-2), kanye nokubambezeleka kwempendulo yokubambisana (ngaphansi kuka-200ms). Sebenzisa izihluzi zenethiwekhi ezinembayo ukuze ubheke ukudluliswa kwedatha phakathi kwamakamela, uqinisekise ukubambezeleka okuncane futhi akukho ukulahleka kwedatha. Landela izindinganiso ezifana ne-GB/T 28181-2016 ezinhlelweni zokuqapha ividiyo noma i-ISO/IEC 29151:2017 yobumfihlo nokuphepha kwedatha kumalungiselelo amakamela amaningi.
Ukuhlolwa kwezimo ezingajwayelekile kuyisinyathelo esinye esigxile ezindabeni. Khomba izenzakalo ezingavamile kodwa ezibalulekile (isibonelo, into eqhamuka kungazelelwe esithombeni, ukuvinjelwa kwekhamera, ukuphazamiseka kwenethiwekhi) bese uqinisekisa ukuthi ikhamera isabela kanjani. Ngokwesibonelo, ikhamera yokuphepha kufanele yazise ngokushesha uma ilensi yayo ivalekile, futhi ikhamera yemoto ezihambelayo kufanele igcine ukutholwa kwezinto noma ngabe imvula noma inkungu kunciphisa ukubonakala. Lezi zivivinyo zihlukanisa amakhamera athembekile kulawa ahluleka ezimweni zangempela.
5. Amathuluzi Abalulekile Nezindlela Ezinhle
Ukuze wenze leli thuluzi elinamasu amathathu kahle, sebenzisa inhlanganisela yamathuluzi ajwayelekile kanye nezinsiza eziphambili. Ukuze uhlola imishini: Ama-Keysight Image Quality Analyzers, Ama-Tektronix Power Analyzers, nezikhala zemvelo. Ukuze uhlola i-algorithm: i-MLPerf Inference (yokuhlola), i-OpenCV, i-TensorRT, kanye ne-Prophesee’s Metavision SDK. Ukuze uhlola izimo: izikhala zokuhlola ezenziwe ngokwezifiso, ama-robot aphathekayo angalungiswa (ukuze simule izinjongo ezihambayo), kanye nezisimulator zenethiwekhi (ukuze kuphinde kube nokuxhumana okuphansi).
Landela lezi zindlela ezinhle ukuze uqinisekise imiphumela ethembekile: 1) Lungisa izimo zokuhlola (ukukhanya, ibanga, izinga lokushisa) ukuze uqinisekise ukuphindaphinda. 2) Hlola kusenesikhathi futhi kaningi—hlanganisa ukuhlolwa kokusebenza phakathi komjikelezo wokuthuthukiswa, hhayi kuphela ekupheleni. 3) Sebenzisa inhlanganisela yokuhlola okuzenzakalelayo nokuhlola okwenziwa ngesandla: zenze imisebenzi ephindaphindayo (isb. ukukala isivinini seframe) futhi uqinisekise izimo ezinzima ngesandla. 4) Bhala konke—landela izilinganiso, izimo zokuhlola, nemiphumela ukuze uthole ukuthambekela nokuxazulula izinkinga.
6. Izinkinga Ezivamile Okufanele Uzigweme
Ngisho noma kunohlaka oluqinile, izinkinga ezivamile zingadala ukuhluleka kokuhlolwa. Gwema ukugxila kuphela ekusebenzeni kwelebhu—izimo zangempela lapho amakhamera ehluleka khona kakhulu. Ungayishayi indiva ukusebenza kahle kwamandla; ikhamera enokunemba okukhulu kodwa edla amandla amaningi ayizimpawu zamadivayisi asebenzisa ibhethri. Gwema ukuhlola ngokweqile kumadatha wokuhlola; amadatha wangokwezifiso abalulekile ukuqinisekiswa okukhethekile kwezimo zokusebenzisa. Ekugcineni, ungakhohlwa ukuhlola ukuhambisana—qinisekisa ukuthi ikhamera isebenza nezingxenye zakho zikagesayidi, isofthiwe, kanye nengqalasizinda yenethiwekhi, ikakhulukazi ezinhlelweni zamakhamera amaningi.
Isiphetho
Ukuhlola nokugunyaza ukusebenza kwekhamera ezibekwe ngaphakathi kudinga indlela ebanzi edlula izicaciso eziyisisekelo. Ngokwamukela uhlaka lwezendlalelo ezintathu—ukwethembeka kwezingxenyekazi zekhompyutha, ukusebenza kahle kwe-algorithm, nokuzivumelanisa nezimo—ungaqinisekisa ukuthi ikhamera yakho iletha ukusebenza okuzinzile, okuthembekile ezindaweni zangempela zangaphandle. Sebenzisa amathuluzi aphezulu njengezilinganiso ze-MLPerf, amakhithi okuvavanya asekelwe kumicimbi ye-Prophesee, kanye nezinhlelo zokuhlola amakhamera amaningi ukuze uhlale uphambili. Noma ngabe ufaka amakhamera okulawula ikhwalithi yezimboni, izimoto ezizihambelayo, noma amadolobha ahlakaniphile, lolu hlelo luzokusiza ukuthi ugweme ukwehluleka okubizayo futhi wakhe ukwethembana kubuchwepheshe bakho.
Ufuna ukuyisa ukuhlolwa kwakho kwe-embedded vision esigabeni esilandelayo? Qala ngokukhomba isimo sakho esibalulekile osisebenzisayo, ukwakha idatha yakho yokuhlola eyenziwe ngokwezifiso, nokubeka phambili izilinganiso ezibaluleke kakhulu kuhlelo lwakho lokusebenza—ukucaca, ukubambezeleka, ukonga amandla, noma ukusebenza ngokubambisana. Ngendlela efanele, ungavula amandla aphelele ubuchwepheshe be-embedded vision.
Uxhumane
Sicela uxhumane nathi uhambele
WhatsApp
WeChat