゚❪㡺㠦㍲ YOLOv3 ₆ ₆ 㹾 㹾⨟ 㧎 㧎㔳 Ἒ Ἒ㑮 㞞
㧊䡲㰚, 㧊㦖㰖, ㏢䡚, 㧚㍶㡗, 㡗䢎,*
㑯ⳛ㡂㧦╖䞯ᾦ IT Ὃ䞯ὒ
㑯ⳛ㡂㧦╖䞯ᾦ ゛◆㧊䎆 䢲㣿 㡆ῂ㎒䎆
e-mail : {adorablehye96, iej9031, shpark, sunnyihm, yhpark}@sookmyung.ac.kr
*ᾦ㔶㩖㧦
YOLOv3-based Vehicle Detection and Counting Method through Video
Hye-Jin Lee, Eun-Ji Lee, So-Hyun Park, Sun-Young Ihm, Young-Ho Park*
Dept. of IT Engineering, Sookmyung Women¶V8QLYHUVLW\
%LJGDWD8VLQJ5HVHDUFK&HQWHU6RRNP\XQJ:RPHQ¶V8QLYHUVLW\
殚 檃檃
⽎ ⏒ⶎ㠦㍲⓪ YOLOv3 ₆㦮 㹾⨟ 㧎㔳 Ἒ㑮 㞞 㡆ῂ⯒ 㰚䟟䞲┺. 㧊⯒ 㥚䟊, ┾㧒 ┾Ἒ
㔳㦮 ῂ㫆⯒ Ṗ㰚 YOLOv3 ⯒ 䢲㣿䞮㡂 㡗㌗ ㏣ 㹾⨟ 㧎㔳 Ἒ㑮 䞮⓪ 㔺䠮㦚 㑮䟟䞲┺. 㥚㦮 㔺䠮㦚 䐋䟊 㡗㌗ ㏣ 㹾⨟㦮 㑮㢖 ₆㠦 㧎㔳⮶ 㩫☚㢖 㹾⨟ Ἒ㑮 䘟‶ Ṩ 㿪㿲㦚 䞮㡖┺. 㔺䠮 ἆὒ, 㡗㌗ ㏣㦮 㹾⨟㦮 㑮㢖 ₆㠦 ㌗ὖ㠜㧊 90%㧊㌗㦮 㹾⨟㠦 ╖䞲 ⏨㦖 㧎㔳⮶㦚 ⽊㡖㦒Ⳇ 1 㽞 ╏ 㟓 20 ⻞㝿 㧎㔳♮⓪ 㹾⨟㦮 㑮㠦 ╖䞲 Ἒ㑮㦮 䘟‶ Ṩ㦚 㿪㿲䞮㡂 䎣㓺䔎 䕢㧒 䡫䌲⪲ 㩖㧻 䞮㡂 㔲Ṛ㠦 ➆⯎ 㹾⨟ Ἒ㑮⯒ 䢫㧎䞮㡖┺. ⽎ ₆㑶㦖 䡚╖ ㌂䣢㦮 ᾦ䐋 㼊㯳㦚 䟊ἆ䞮₆ 㥚䞲 ᾦ 䐋 䦦⯚ 㡞䁷㠦 䢲㣿♶ 㑮 㧞┺.
1. 昢嵦
㾲⁒ IT ₆㑶㦮 ザ⯎ ╂ὒ 䞾℮ ₆㫊 ᾦ䐋 㔲㓺 䎲㠦 㻾┾ ₆㑶㦚 㩧⳿㔲䋺⓪ ITS(Intelligent Transport System) ₆㑶㦚 䢲㣿䞮㡂 㹾⨟ ᾦ䐋⨟㦒⪲ 㧎䞲 ᾦ䐋 㩫㼊, ㌂ἶ ㌳ ❇㦮 ṗ㫛 ᾦ䐋 ⶎ㩲⯒ 㼊Ἒ㩗 㦒⪲ ╖㻮䞮㡂 䣾㥾㩗㧎 ᾦ䐋 㤊㡗ὒ ᾦ䐋 㧊㣿 䘎㦮
⯒ ⏨㧊ἶ㧦䞮⓪ ┺㟧䞲 㡆ῂ✺㧊 ❇㧻䞮ἶ 㧞┺. ┺ 㟧䞲 㟒 㭧 㡗㌗㦚 䐋䞲 㹾⨟ 㧎㔳 Ἒ㑮⯒ 䐋䞲 ᾦ䐋 䦦⯚ 䕢㞛 㡆ῂ⯒ 㰚䟟䞲┺.
㡺⓮⋶㦮 CCTV 㦮 㡃䞶㦖 ┾㑲 Ṧ㔲Ṗ 㞚┢ ⻪㬚 㡞, ⰺ㧻㦚 ⶎ䞮⓪ ἶṳ㦮 㑮, ἶṳ㦮 ☯㍶ 䕢㞛
❇ Ⱎ䅖䕛㠦 ┺㟧䞮Ợ 䢲㣿♮⓪ ㌂⪖Ṗ ⓮㠊⋮Ⳇ CCTV 㦮 㡃䞶㦖 ⋮⋶㧊 䄺㪎Ṗἶ 㧞┺. ┺㟧䞲 CCTV 㦮 㡃䞶 㭧 Ἒ㑮䞮⓪ 㡃䞶㦚 㹾⨟㠦 㩧⳿㔲䅲 㡗㌗ ㏣ 㹾⨟㦮 㑮⯒ 䕢㞛䞮㡂 ᾦ䐋 䦦⯚ 㡞䁷㦚 䞲
┺. 㧊⯒ 䐋䟊, 䡚╖ ㌂䣢㦮 ᾦ䐋 㩫㼊⯒ 䟊ἆ䞮㡂 䣾 㥾㩗㧎 ᾦ䐋 㤊㡗ὒ ᾦ䐋 㧊㣿 䘎㦮⯒ ⏨㧊ἶ㧦 䞲┺.
➆⧒㍲, ⽎ ⏒ⶎ㠦㍲⓪ CCTV 㡗㌗ ㏣ 㹾⨟ 㧎㔳
Ἒ㑮⯒ 䐋䞲 ᾦ䐋 䦦⯚ ㍳ 㡞䁷䞮⓪ 㞞㦚 㩲㞞䞮ἶ㧦 䞲┺. 㧊 ⻫㦖 YOLOv3 ₆㦒⪲ 㹾⨟
㧎㔳 Ἒ㑮⯒ 㻮Ⰲ䞮⓪ ⻫㧊┺. YOLOv3 㞢ἶⰂ㯮 㦖 2.2 㩞㠦㍲ ㍺ⳛ䞲┺.
⽎ ⏒ⶎ㦮 ῂ㎇㦖 ┺㦢ὒ ṯ┺. 2 㧻㠦㍲⓪ ╖䚲㩗 㧎 ṳ㼊 㧎㔳 㞢ἶⰂ㯮✺㠦 ╖䟊 ㍺ⳛ䞲┺. 3 㧻㠦㍲
⓪ 㩲㞞䞮⓪ ⻫㦚 ㍺ⳛ䞲┺. 4 㧻㠦㍲⓪ 㔺䠮㠦 ╖ 䞮㡂 ㍺ⳛ䞮⓪◆, Ⲓ㩖 Ṳ 䢮ἓ㦚 ㏢Ṳ䞮ἶ 䤚㠦 㔺䠮 ἆὒ ㍳㦚 㰚䟟䞲┺. Ⱎ㰖Ⱏ㦒⪲ 5 㧻㠦㍲
⓪ ἆ⪶ 䟻䤚 㡆ῂ㠦 ╖䟊 ㏢Ṳ䞲┺.
2. 分崮 櫶割
⽎ 㧻㠦㍲⓪ ╖䚲㩗㧎 ṳ㼊 㧎㔳 㞢ἶⰂ㯮㠦 ╖䞮 㡂 䋂Ợ 2 Ṗ㰖⪲ ⋮㠊 ㍺ⳛ䞲┺. 2.1 㩞㠦㍲⓪ 2 ┾ Ἒ 㔳 ῂ㫆㦮 ṳ㼊 㧎㔳 㞢ἶⰂ㯮㧎 Faster R-CNN 㞢ἶⰂ㯮㠦 ╖䞮㡂 ㍺ⳛ䞮ἶ, 2.2 㩞㠦㍲⓪ ┾㧒 ┾Ἒ
㔳 ῂ㫆㦮 ṳ㼊 㧎㔳 㞢ἶⰂ㯮㧎 SSD(Single shot Detector) 㞢ἶⰂ㯮㠦 ╖䞮㡂 ㍺ⳛ䞲┺.
2.1 Faster R-CNN㦮 㞚䋺䎣㻮
Faster R-CNN 㦖 ⍺䔎㤢䋂㠦㍲ ㌳㎇䞲 䞒㼦 ⱋ 㥚 㠦 n * n 䋂₆㦮 㥞☚㤆⯒ ㌳㎇䞮㡂 㔂⧒㧊❿䞮Ⳇ Region proposals 㦚 ㌳㎇䞮⓪ FCN(Fully Convolutional Network)䡫䌲㦮 RPN(Region Proposal Network)㦚 ☚㧛 䞮㡂 ₆㫊㦮 Fast R-CNN 㦚 Ṳ㍶䞲 㞢ἶⰂ㯮㧊Ⳇ[1, 5],
㌗╖㩗㦒⪲ ㏣☚Ṗ ⓦⰆ CPU ⪲ Ἒ㌆䞮⓪ ╖㔶 ザ⯎
RPN(GPU)㦚 ㌂㣿䞮㡂 ㎇⓻㦚 Ṳ㍶䞮㡖┺. Faster R- CNN 㦮 㞚䋺䎣㻮⓪ ⁎Ⱂ 1 㦚 䐋䟊 䢫㧎 Ṗ⓻䞮┺.
RPN ⓪ 㧊⹎㰖⯒ 㧛⩻ 㞚 ㌂ṗ䡫 䡫䌲㦮 㡺ぢ㩳䔎 䝚⪲䙂㧮ὒ ṳὖ㩗㧎 㩦㑮⯒ 㿲⩻䟊㭒⓪ 㡃䞶㦚 䞲┺.
2019년 추계학술발표대회 논문집 제26권 제2호 (2019. 11)
- 935 -
⁎Ⱂ 2[6]⯒ ⽊Ⳋ ☯㧒䞲 䋂₆㦮 㔂⧒㧊❿ 㥞☚㤆⯒
㧊☯㔲䋺Ⳇ 㥞☚㤆㦮 㥚䂮⯒ 㭧㕂㦒⪲ ㌂㩚㠦 㩫㦮♲
┺㟧䞲 ゚㥾ὒ 䋂₆⪲ 㧊⬾㠊㰚 㟋䄺 㓺✺㦚 㩗㣿 䞮㡂 䔏㰫㦚 㿪㿲䞲┺[6]. ⁎⩂⸖⪲ 䞒⧒⹎✲ 䔏㰫 Ἒ 䂋 ῂ㫆㻮⩒ 㧊⹎㰖 䋂₆⯒ 㫆㩫䞶 䞚㣪Ṗ 㠜㦒Ⳇ 㡂
⩂ ′⳾㦮 㔂⧒㧊❿ 㥞☚㤆㻮⩒ 䞚䎆 䋂₆⯒ ⼖ἓ䞶 䞚㣪☚ 㠜㦒⸖⪲ Ἒ㌆ 䣾㥾㧊 ⏨┺⓪ 㧻㩦㦚 Ṭἶ 㧞
┺. 䞮㰖Ⱒ, 2 ┾Ἒ 㔳 ῂ㫆⓪ Ὃ䐋㩗㦒⪲ ⽋㧷䞲 䕢 㧊䝚⧒㧎㦚 Ṗ㰖Ⳇ ㏣☚Ṗ ⓦ⩺ 㔺㔲Ṛ ṳ㼊 Ṧ㰖Ṗ 䧮✺┺⓪ ┾㩦㦚 Ṭἶ 㧞┺.
(⁎Ⱂ 1) Faster R-CNN 㦮 㞚䋺䎣㻮
(⁎Ⱂ 2) RPN 㦮 㞚䋺䎣㻮
2.2 SSD(Single Shot Detector)㦮 㞚䋺䎣㻮
SSD ⓪ VGG16[4] ⍺䔎㤢䋂⯒ ₆⓻ 㿪㿲₆⪲ ㌂㣿 䞮⓪ ┾㧒 ㍍ Ỗ㿲₆㧊┺. 㧊 㞢ἶⰂ㯮㦖 ⁎Ⱂ 3[2]ὒ ṯ㧊 䞒㼦 ⱋ㦚 ㆧ㞚⌊⓪ ὒ㩫₢㰖 䞮⋮㦮 ⍺䔎㤢䋂 ῂ㫆⯒ Ṗ㰖⓪ 㞢ἶⰂ㯮㧊┺. SSD ⓪ 䞒㼦 ⱋ㦚 㡂⩂
Ṳ㦮 䋂₆⪲ Ⱒ✺㠊㍲, 䋆 䞒㼦 ⱋ㠦㍲⓪ 㧧㦖 ⶒ㼊 㦮 Ỗ㿲㦚 䞮ἶ 㧧㦖 䞒㼦 ⱋ㠦㍲⓪ 䋆 ⶒ㼊㦮 Ỗ㿲 㦚 䞮㡂 㥚䂮 㿪㩫 㧛⩻ 㧊⹎㰖㦮 㨂䚲㰧㦚 㠜㞶 Ⳋ㍲ 㩫䢫☚ ⏨㦖 ἆὒ⯒ ☚㿲䞲┺. 㡂⩂ 䞒㼦 ⱋ㦖 䞲 㧊⹎㰖⯒ ┺㟧䞲 ⁎Ⰲ✲⪲ 㩧⁒䞮㡂 ┺㟧䞲 䋂₆ 㦮 ⶒ㼊✺㦚 Ỗ㿲䞶 㑮 㧞☚⪳ 䞮Ⳇ[2], 䋆 ⶒ㼊⓪ ゚ 㥾㧊 ╂⧒㪎☚ 㧮 㺔⓪ 㧻㩦㦚 Ṭἶ 㧞┺. 䞮㰖Ⱒ, 㡂
⩂ Ṳ㦮 䞒㼦 ⱋ㦮 Ỗ㿲㦚 ┺ Ἒ㌆䞮⸖⪲ Ἒ㌆ ゚㣿 㧊 㯳Ṗ 䞶 㑮 㧞㦒Ⳇ 㧛⩻ ㌂㧊㯞Ṗ 䋂㰖 㞠┺Ⳋ 㾲
㌗㥚 Ἒ䂋㠦㍲䎆 㩫⽊Ṗ 㩗₆ ➢ⶎ㠦 㧧㦖 ⶒ㼊⓪ 㧮 㺔㰖 ⴑ䞲┺⓪ ┾㩦㦚 Ṭἶ 㧞┺.
(⁎Ⱂ 3) SSD 㦮 㞚䋺䎣㻮
3. 洢橎穞垚 CCTV 焮峏 凊朞 愯憛
⽎ 㧻㠦㍲⓪ 㞴㍲ 2 㧻㠦㍲ ㍺ⳛ䟞▮ ὖ⩾ 㡆ῂ✺
⽊┺ 㔺㔲Ṛ ṳ㼊 㧎㔳 㞢ἶⰂ㯮 㭧㠦㍲ ⽎ 㡆ῂ㠦 㩲㧒 㩗䞿䞮┺ἶ 䕦┾♮⓪ YOLOv3 㞢ἶⰂ㯮㦚 䢲㣿䞮 㡂 CCTV 㠦㍲ 㹾⨟ 㧎㔳 Ἒ㑮 㞞㠦 ὖ䞮㡂 㩲 㞞䞮⓪ ⻫㦚 ㍺ⳛ䞲┺. Ⲓ㩖, 3.1 㩞㠦㍲⓪ YOLOv3 㞢ἶⰂ㯮㠦 ╖䞮㡂 ㍺ⳛ䞮ἶ, 3.2 㩞㠦㍲⓪ YOLOv3 㞢 ἶⰂ㯮㦚 㧊㣿䞲 㹾⨟ 㧎㔳 Ἒ㑮 㞞㠦 ╖䞮㡂
㍺ⳛ䞲┺.
3.1 YOLOv3㦮 㞚䋺䎣㻮
YOLOv3 ⓪ ⁎Ⱂ 4[8]㢖 ṯ㧊 ┾㧒 㔶ἓⰳ㦚 㩗㣿䞮 㡂 㧊⹎㰖⯒ 㧒㩫䞲 㡗㡃㦒⪲ 䞶䞮ἶ ṗṗ㦮 㡗㡃 㠦 Ṗ㭧䂮⯒ 㡂䞮⓪ 㔳㦚 ㌂㣿䞲┺. ṳ㼊⯒ 㡂⩂
⻞ 䢎㿲䞮⓪ ₆㫊㦮 ṳ㼊 㧎㔳 㞢ἶⰂ㯮ὒ⓪ ┺⯊Ợ 㧊⹎㰖⯒ 䞲 ⻞Ⱒ 䢎㿲䞮⸖⪲ ㏣☚ 䁷Ⳋ㠦㍲ R-CNN
⽊┺ 1000 ⺆ 㧊㌗ ザ⯊Ⳇ Fast R-CNN ⽊┺ 100 ⺆ ザ
⯎ ㏣☚㢖 ⏨㦖 㩫䢫☚⯒ ⽊㧊⓪ 㞢ἶⰂ㯮㧊┺.
FPN(Feature Pyramid Network)㦮 䄾㎟ὒ 㥶㌂䞲 㔳㦒⪲ ṗṗ 2 ⺆㝿 㹾㧊Ṗ ⋮⓪ 3 Ṳ㦮 ┺⯎ 㓺䅖㧒 㦮 㓺✺㦚 㡞䁷䞮ἶ ┺㟧䞲 㓺䅖㧒㠦㍲ 䔏㰫✺㦚 㿪㿲䞲┺. 3 Ṳ㦮 㤊❿ 㓺 * 3 Ṳ㦮 䞒㼦 ⱋ㦒⪲
㧊⬾㠊㰚 㽳 9 Ṳ㦮 㟋䄺 㓺⓪ k-means 䋊⩂㓺䎆Ⱇ 㦚 䐋䟊 ἆ㩫♲┺[3]. ┺⯎ 㞢ἶⰂ㯮✺㦮 ゚䟊 Ṛ┾䞲 㻮Ⰲ ὒ㩫㦒⪲ ㏣☚Ṗ ⰺ㤆 ザ⯊Ⳇ 㧊⹎㰖 㩚㼊⯒ 䞲
⻞㠦 ⧒⽊⓪ 㔳㦒⪲ 䋊⧮㓺㠦 ╖䞲 ⰻ⧓㩗 㧊䟊
☚Ṗ ⏨┺⓪ 㧻㩦㦚 Ṭἶ 㧞┺. 䞮㰖Ⱒ, 㧧㦖 㡺ぢ㩳 䔎㠦 ╖䞮㡂 ㌗╖㩗㦒⪲ ⌄㦖 㩫䢫☚⯒ ⽊㧊⓪ ┾㩦 㦚 Ṭἶ 㧞┺.
(⁎Ⱂ 4) YOLOv3 㦮 㞚䋺䎣㻮 2019년 추계학술발표대회 논문집 제26권 제2호 (2019. 11)
- 936 -
3.2 YOLOv3 ₆㦮 㹾⨟ 㧎㔳 Ἒ㑮 㞞
⽎ 㩞㠦㍲⓪ ゚❪㡺㠦㍲ YOLOv3 ₆ 㹾⨟ 㧎㔳
㞞㦚 㩲㞞䞲┺. 㩲㞞䞮⓪ ⻫㦖 YOLOv3 ₆㠦㍲
㹾⨟㦚 㧎㔳 Ἒ㑮䞮⓪ ┾Ἒ㠦 ╖䞮㡂 ㍺ⳛ䞲┺.
┾Ἒ⓪ 㽳 ㎎ ┾Ἒ⪲ ῂ㎇♲┺.
a) Step 1 (┾㧒 㔶ἓⰳ㦚 䐋䞲 㡗㡃 䞶 ┾Ἒ): ┾ 㧒 㔶ἓⰳ㦚 䐋䞲 㡗㡃 䞶 ┾Ἒ㠦㍲⓪ 㧛⩻
㦖 㧊⹎㰖⯒ ┾㧒 㔶ἓⰳ㦚 㩗㣿䞮㡂 㧒㩫䞮 Ợ 䞶䞲┺.
b) Step 2 (ἓἚ ㌗㧦 䢫⮶ 㡞䁷 ┾Ἒ): ἓἚ ㌗ 㧦 䢫⮶ 㡞䁷 ┾Ἒ㠦㍲⓪ ⪲㰖㓺䕇 䣢‖⯒
㧊㣿䞮㡂 ṗ ἓἚ ㌗㧦㠦 ╖䞲 ṳὖ㩗㧎 㩦㑮
⯒ 㡞䁷䞮ἶ 㟋䄺 㓺㢖 㔺㩲 㤊❿ 㓺㦮 㩲㧒 㫡㦖 IoU 㠦 Ṩ㦚 1 ⪲ ⚦㠊 Ἒ㌆䞮㡂 ἓ Ἒ ㌗㧦 㡞䁷㦚 䞲┺.
c) Step 3 (◆㧊䎆 㿪㿲 㩖㧻 ┾Ἒ): ◆㧊䎆 㿪㿲
㩖㧻 ┾Ἒ㠦㍲⓪ 㧊⹎㰖㠦㍲ 㿪㿲䞲 ṳ㼊㠦 ἓἚ 㓺 䢫⮶ 㡞䁷㧊 䚲㔲♮㠊 㩖㧻♲┺.
☯㔲㠦 䡚㨂 㔲Ṛὒ 䞾℮ 㡗㌗ ㏣㠦㍲ 㧎㔳♲
㹾⨟㦮 㑮Ṗ 䎣㓺䔎 䕢㧒⪲ 㩖㧻♲┺.
㧊➢, 㡗㌗ ㏣㠦㍲ 㧎㔳♲ 㹾⨟㦖 1 㽞㠦 㟓 20 ⻞ 㝿 䎣㓺䔎 䕢㧒⪲ 㩖㧻♲┺. 1 㽞 ╏ 㧎㔳♲ 㹾⨟㦮 㑮㠦 ╖䞲 㡺㹾 ⻪㥚⯒ 㭚㧊₆ 㥚䟊 1 㽞 ╏ 㧎㔳♲
㹾⨟㦮 㑮㠦 ╖䞲 䘟‶ Ṩ㦚 㿪㿲䞮㡂 ㌞⪲㤊 䎣㓺䔎 䕢㧒⪲ ┺㔲 㩖㧻䞮㡖┺. 㧊⯒ 䐋䟊, 㔲Ṛ㠦 ➆⯎ 㹾
⨟ Ἒ㑮 ἆὒ⯒ 䢫㧎䞮㡖㦒Ⳇ, ἆὒ 㧊⹎㰖⓪ 4.2 㩞 㔺䠮 ἆὒ ㍳㠦㍲ 䢫㧎䞶 㑮 㧞┺.
4. 柪竞
⽎ 㧻㠦㍲⓪ Ṳ 䢮ἓὒ 㔺䠮 ἆὒ㠦 ╖䟊 ㍺ⳛ䞲
┺. Ⲓ㩖, 4.1 㩞㠦㍲⓪ Ṳ 䢮ἓ㦚 ㍺ⳛ䞮ἶ, 4.2 㩞㠦
㍲⓪ 㔺䠮 ἆὒ ㍳㦚 ㍺ⳛ䞲┺.
4.1 Ṳ 䢮ἓ
⽎ 㡆ῂ㠦㍲⓪ ἓ㺆㼃 ☚㔲ᾦ䐋㩫⽊㎒䎆㠦㍲ 㩲Ὃ 䞮⓪ CCTV[7]⯒ ㌂㣿䞮㡖┺. 㔺䠮 㡗㌗㦖 㹾⨟㦮 㑮 Ṗ 㩗ἶ Ⱔ㦚 ➢㢖 㡗㌗㧊 ἶ 㠊⚦㤎 ➢㠦 ➆⯎ 㧎 㔳⮶㦚 ゚ᾦ䞮₆ 㥚䟊 㿲䑊⁒ 㔲Ṛ ➢㧎 㡺㩚 7 㔲
䎆 㡺㩚 9 㔲㢖 㡺䤚 6 㔲䎆 㡺䤚 8 㔲₢㰖 2 Ṗ㰖
㌗䢿㦒⪲ ⋮㠊 㰚䟟䞮㡖┺. Ṳ㦚 㰚䟟䞲 䅊䜾䎆
㌂㟧㦖 Intel® CoreTM i7-2000 CPU @ 3.40GHz㧊ἶ ⁎⧮
䞓 䃊✲⓪ GeForce GTX 1080 Ti ⯒ ㌂㣿䞮㡖┺.
4.2 㔺䠮 ἆὒ㍳
⁎Ⱂ 5 ⓪ 㹾⨟㦮 㑮Ṗ 㩗ἶ Ⱔ㦚 ➢㢖 㡗㌗㧊 ἶ 㠊⚦㤎 ➢㠦 ➆⯎ 㧎㔳⮶㦚 ゚ᾦ䞮₆ 㥚䟊 㿲䑊⁒
㔲Ṛ ➢㧎 㡺㩚 7 㔲䎆 㡺㩚 9 㔲㢖 㡺䤚 6 㔲䎆 㡺䤚 9 㔲⪲ ⋮㠊 㔺䠮㦚 㰚䟟䞮㡖┺. ⁎Ⱂ 5 㦮 (a), (b)⓪ 㿲⁒ 㔲Ṛ ➢㧎 8 㔲 30 㸺㦮 㧊⹎㰖㧊ἶ
⁎Ⱂ 5 㦮 (c), (d)⓪ 䑊⁒ 㔲Ṛ ➢㧎 㡺䤚 8 㔲㸺㦮 㧊
⹎㰖㧊┺. ⁎Ⱂ 5 㦮 (a), (c)⓪ 㹾⨟㦮 㑮Ṗ 㩗㦖 ἓ㤆 㧊ἶ ⁎Ⱂ 5 㦮 (b), (d)⓪ 㹾⨟㦮 㑮Ṗ Ⱔ㦖 ἓ㤆㧊┺.
㔺䠮 ἆὒ, 㹾⨟ 㧎㔳㦖 㡗㌗ ㏣ 㹾⨟㦮 㑮, 㡗㌗㦮
₆㢖 ㌗ὖ㠜㧊 ṗ 㹾⨟㠦 ╖䞲 Ỗ㯳 㩫䢫☚Ṗ 䘟‶
90%㧊㌗㦒⪲ ⏨㦖 㧎㔳⮶㦚 ⽊㡖┺. ⁎ ἆὒ⯒ 1 㽞
╏ 㟓 20 ⻞㝿 㧎㔳♮⓪ 㹾⨟㦮 㑮㠦 ╖䞮㡂 㡺㹾 ⻪ 㥚⯒ 㭚㧊₆ 㥚䟊 1 㽞╏ 㧎㔳♲ 㹾⨟㦮 㑮㠦 ╖䞲 䘟‶ Ṩ㦚 㿪㿲䞮㡂 ㌞⪲㤊 䎣㓺䔎 䕢㧒⪲ 㩖㧻䞮㡂 㔲Ṛ㠦 ➆⯎ 㹾⨟ Ἒ㑮⯒ 䢫㧎䞮㡖┺. ⁎Ⱂ 6 㦮 㫢䁷 㦖 㤦⽎ 䎣㓺䔎 䕢㧒㧊Ⳇ 㤆䁷㦖 㧎㔳♲ 㹾⨟㦮 㑮㠦
╖䞲 䘟‶ Ṩ㦚 㿪㿲䞮㡂 㩖㧻䞲 䎣㓺䔎 䕢㧒㦮 ἆὒ 䢪Ⳋ㧊┺.
(⁎Ⱂ 5) 㡗㌗ ㏣ 㹾⨟ 㧎㔳
(⁎Ⱂ 6) 㹾⨟ 㑮㦮 䘟‶ Ṩ 䎣㓺䔎 䕢㧒
5. 冶嵦 愕 窫篊 櫶割
⽎ ⏒ⶎ㠦㍲⓪ CCTV 㦮 ┺㟧䞲 㡃䞶 㭧 ⶒ㼊⯒ 㧎 㔳䞮㡂 Ἒ㑮䞮⓪ 㡃䞶㦚 㹾⨟ὒ 㩧⳿㔲䅲 㡗㌗ ㏣ 㹾
⨟㦚 㧎㔳 Ἒ㑮䞮㡂 ᾦ䐋 䦦⯚ 㡞䁷㦚 䐋䟊 䡚╖
㌂䣢㦮 ᾦ䐋 㼊㯳㦚 䟊ἆ䞮₆㥚䟊 ᾦ䐋 䦦⯚ 㡞䁷㦚 䞮⓪ 㞞㦚 㩲㔲䞮㡖┺. ◆㧊䎆⓪ ☚㔲ᾦ䐋㩫⽊㎒䎆 㠦㍲ 㩲Ὃ䞮⓪ CCTV 㡗㌗㦚 㑮㰧䞮㡂 ㌂㣿䞮㡖㦒Ⳇ ṳ㼊 㧎㔳 㞢ἶⰂ㯮㦖 YOLOv3 ⯒ 㩗㣿䞮㡖┺.
㔺䠮 ἆὒ, 㡗㌗ ㏣ 㹾⨟㦮 㑮 㡗㌗㦮 ₆㢖
㌗ὖ㠜㧊 㹾⨟㠦 ╖䞲 Ỗ㯳 㩫䢫☚Ṗ 䘟‶ 90%㧊㌗
㦒⪲ 㫡㦖 㧎㔳⮶㦚 ⽊㡖┺. 㧊⯒ 䐋䟊, 1 㽞╏ 㟓 20
⻞㝿 㧎㔳♲ 㹾⨟㦮 㑮㠦 ╖䞲 㡺㹾 ⻪㥚⯒ 㭚㧊₆ 㥚䟊 1 㽞╏ 㧎㔳♲ 㹾⨟㦮 㑮㠦 ╖䞲 䘟‶ Ṩ㦚 㿪 㿲䞲 ἆὒ⯒ 䎣㓺䔎 䕢㧒⪲ 㩖㧻䞮㡂 ᾦ䐋 䦦⯚ 㡞䁷 㦚 䞮ἶ㧦 䞮㡖┺. 䞮㰖Ⱒ 㡗㌗㦮 ⳾㍲Ⰲ 㴓㦒⪲ Ṟ 2019년 추계학술발표대회 논문집 제26권 제2호 (2019. 11)
- 937 -
㑮⪳ 㧎㔳㧊 㧮 㞞♲┺⓪ ⶎ㩲㩦ὒ 㹾⨟㦮 㑮Ṗ ỿ 䞮Ợ Ⱔ㞚㰖Ⳋ ṯ㦖 ㌟㦮 㹾⨟㧊 䞮⋮㦮 㹾⨟㦒⪲ 㧎 㔳♮⓪ ⶎ㩲㩦㠦 ╖䞮㡂 ⽊㢚 ⻫㧊 䞚㣪䞮┺.
䟻䤚 㡆ῂ⪲⓪ 㡗㌗ ⳾㍲Ⰲ 㴓㦮 㧎㔳⮶ Ṳ㍶ὒ 㹾
⨟㦮 㑮Ṗ Ⱔ㞚㰞 ἓ㤆 ṯ㦖 ㌟㦮 㹾⨟㧊 䞮⋮㦮 㹾
⨟㦒⪲ 㧎㔳♮⓪ ⶎ㩲 Ṳ㍶ 㡆ῂ ❇㦚 㰚䟟䞶 㡞㩫㧊
┺.
斲斲怾割
㧊 ⏒ⶎ㦖 2019 ⎚☚ 㩫(⹎⧮㺓㫆ὒ䞯)㦮 㨂㤦㦒⪲
㩫⽊䐋㔶₆㑶㰚䦻㎒䎆㦮 㰖㤦㦚 㞚 㑮䟟♲ 㡆ῂ㧚 (No.2016-0-00406. (₆ SW-㺓㫆㝾㞭 2 ┾Ἒ)SIAT 䡫 CCTV 䋊⧒㤆䔎 䝢⨁䙒 ₆㑶 Ṳ) ⁎Ⰲἶ 2019 ⎚☚ 㩫(ὒ䞯₆ 㑶㩫⽊䐋㔶)㦮 㨂㤦㦒⪲ 㩫⽊䐋㔶₆㑶㰚䦻㎒䎆㦮 㰖㤦㦚
㞚 㑮䟟♲ 㡆ῂ㧚 (No.2018-0-00225, ὒ䞯㩗 㩫㺛 㑮Ⱃ㦚 㥚䞲 ☚㔲䟟㩫 ❪㰖䎎䔎㥞 䟋㕂 ₆㑶 Ṳ)
焾処怾竒
[1] B. Benjdira, T. Khursheed, A. Koubaa, A. Ammar and K.Ouni, ³Car detection using unmanned aerial vehicles:
Comparison between faster r-cnn and yolov3, ´ in Proc.
1st Int. Conf. Unmanned Vehicle Syst.-Oman (UVS), 2019.
[2] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, and S. Reed,
³SSD:Single shot multibox detector, ´ arXiv:1512.02325, 2015.
[3] J. Redmon and A. Farhadi, ³Yolov3: An incremental improvement," arXiv preprint arXiv:1804.02767, 2018.
[4] K. Simonyan and A. =LVVHUPDQ ³Very Deep Convolutional Networks for Large-Scale Image Recognition´ ,QWHUQDWLRQDO &RQIHUHQFH RQ /HDUQLQJ
Representations (ICRL), 2015.
>@5*LUVKLFN ³Fast R-CNN´LQ3URFHHGLQJVRIWKH,(((
International Conference on Computer Vision, 2015.
[6] S. Ren, K. He, R. Girshick and J. Sun, ³Faster r-cnn:
Towards real-time object detection with region proposal networks, ´ In NIPS, 2015.
[7] CCTV㡗㌗ 㩲Ὃ,
https://www.utic.go.kr:449/main/main.do
[8] 㧦☯㹾 ㌂㰚, https://news.joins.com/article/21793292
2019년 추계학술발표대회 논문집 제26권 제2호 (2019. 11)
