光源変化シミュレーションと深層学習による特徴量変換を用いたカメラ位置姿勢推定

全文

(1)Vol.2017-CVIM-207 No.19 2017/5/10. ৘ใॲཧֶձ‫ڀݚ‬ใࠂ IPSJ SIG Technical Report. ޫ‫ݯ‬มԽγϛϡϨʔγϣϯͱਂ૚ֶशʹΑΔ ಛ௃ྔม‫׵‬Λ༻͍ͨΧϝϥҐஔ࢟੎ਪఆ ਖ਼ຬ ૑ଠ1,a). ؒԼ Ҏେ1,2,b). Photchara Ratsamee1,2,c) ஛ଜ ࣏༤1,2,f). Ӝ੢ ༑थ1,2,d). ਗ਼઒ ਗ਼3,e). ֓ཁɿ֦ு‫࣮ݱ‬ʢAugmented Reality; ARʣͰόʔνϟϧ‫ڥ؀࣮ͱڥ؀‬ͷ‫ز‬Կֶత੔߹ੑΛͱΔͨΊͷํ ๏ͷҰͭʹɼࣄલʹ࡞੒ͨ͠ಛ௃ྔσʔλϕʔεͱ࣮ը૾ͷಛ௃ྔΛϚονϯάͤͯ͞ΧϝϥͷҐஔ࢟੎ Λਪఆ͢Δࣗ‫ݾ‬Ґஔਪఆ͕͋Δɽ͔͠͠ৗʹมԽ͍ͯ͠Δ࣮ੈքͰࣗ‫ݾ‬ҐஔਪఆΛߦ͏৔߹ɼ࣌ؒଳ΍ఱ ީʹΑͬͯޫ‫͕ڥ؀ݯ‬มԽ͢Δɽͦͷ݁Ռɼೖྗը૾ͱσʔλϕʔε಺ͷಛ௃ྔͷϚονϯάͷࣦഊʹΑ Δࣗ‫ݾ‬Ґஔͷਫ਼౓ͷ௿Լ͕໰୊ͱͳΔɽ͜ͷ໰୊ʹରͯ͠ɼγϛϡϨʔγϣϯʹΑͬͯଟ༷ͳޫ‫Ͱڥ؀ݯ‬ ͷόʔνϟϧը૾Λ༻͍ͯσʔλϕʔεΛߏங͢Δ͜ͱͰޫ‫ڥ؀ݯ‬ͷมԽʹରͯ͠‫͢ʹ݈ؤ‬Δख๏͕͋Δ ͕ɼγϛϡϨʔγϣϯͱ‫࣮ݱ‬ͷဃ཭͕՝୊Ͱ͋Δɽຊ‫Ͱڀݚ‬͸࣮ը૾͔Β෮‫ ͨ͠ݩ‬3 ࣍‫ݩ‬ϞσϧΛ‫͠ʹج‬ ͨγϛϡϨʔγϣϯͱɼࣗ‫߸ූݾ‬Խ‫ث‬ɼϥϯμϜϑΥϨετʹΑΔ‫ػ‬ցֶशΛ༻͍ͨޫ‫ڥ؀ݯ‬ͷมԽʹ‫ؤ‬ ݈ͳࣗ‫ݾ‬Ґஔਪఆख๏ΛఏҊ͢Δɽࣗ‫߸ූݾ‬Խ‫࣮Ͱث‬ը૾ͷಛ௃ྔΛόʔνϟϧը૾ͷಛ௃ྔʹม‫͢׵‬Δ ͜ͱʹΑͬͯɼγϛϡϨʔγϣϯͱͷဃ཭Λղܾ͢Δɽ·ͨϥϯμϜϑΥϨετʹΑΔϚονϯάΛߦ͏ ͜ͱͰϚονϯάਫ਼౓Λ޲্ͤ͞Δɽ࣮‫ݧ‬ͷ݁ՌɼఏҊख๏Ͱ͸ैདྷख๏ͱൺֱͯ͠গͳ͍൑ఆ࣌ؒͰਫ਼ ౓ͷߴ͍ࣗ‫ݾ‬Ґஔਪఆ͕ߦ͑Δ͜ͱΛ֬ೝͨ͠ɽ Ωʔϫʔυɿ֦ு‫࣮ݱ‬ɼࣗ‫ݾ‬Ґஔਪఆɼޫ‫ݯ‬มԽɼࣗ‫߸ූݾ‬Խ‫ث‬ɼϥϯμϜϑΥϨετ. 1. ͸͡Ίʹ ֦ு‫࣮ݱ‬ʢAugmented Reality; ARʣٕज़ͷதͰɼΧϝϥ ͰࡱӨͨ͠ը૾ʹίϯϐϡʔλάϥϑΟοΫεʢComputer. Graphics; CGʣΛਖ਼͍͠Ґஔʹදࣔ͢Δɼ‫ز‬Կֶత੔߹ੑ ʹؔ͢Δ‫͕ڀݚ‬੝ΜʹߦΘΕ͍ͯΔɽ‫ز‬Կֶత੔߹ੑΛಘ ΔͨΊʹ͸ɼࡱӨͨ͠ΧϝϥͷҐஔ࢟੎Λਖ਼֬ʹਪఆ͢Δ ࣗ‫ݾ‬Ґஔਪఆ͕ඞཁͰ͋Δɽࣗ‫ݾ‬Ґஔਪఆ͸ॳ‫ظ‬ϑϨʔϜ ͱτϥοΩϯάϩετ͔Βͷ෮‫ߦʹ࣌چ‬ΘΕΔͨΊɼ਺ඵ Ҏ಺ʹॲཧΛऴ͑Ε͹֦ு‫༻࣮͍͓ͯʹ࣮ݱ‬తͱ‫͑ݴ‬Δɽ ࣗ‫ݾ‬Ґஔਪఆख๏ʹ͸ਓ޻తͳϚʔΧΛ༻͍Δख๏ [3]ɼ. GPS ΍ Bluetooth ͳͲͷηϯαΛ༻͍Δख๏ [4], [5]ɼσʔ. ਤ 1. ޫ‫ݯ‬มಈʹΑΓมԽ͢ΔಉҰ෺ମͷ֎‫؍‬. λϕʔεʢDatabase; DBʣΛ༻͍Δख๏ [6], [7], [8] ͳͲ. ϯதͷಛ௃ྔͱͦΕʹରԠ͢Δ‫ੈ࣮ݱ‬քͷ 3 ࣍‫࠲ݩ‬ඪͷ૊. ༷ʑͳख๏͕͋Δɽ͜ͷ͏ͪɼDB Λ༻͍Δख๏͸༧Ίγʔ. Λ‫ه‬࿥ͨ͠ DB Λ࡞੒͠ɼೖྗը૾ͷಛ௃ྔͱϚονϯά Λߦ͏͜ͱͰࣗ‫ݾ‬ҐஔਪఆΛ࣮‫͢ݱ‬Δɽ͔͠͠ɼ԰֎ͳͲ. 1 2 3 a) b) c) d) e) f). େࡕେֶ ৘ใՊֶ‫ڀݚ‬Պ େࡕେֶ αΠόʔϝσΟΞηϯλʔ ಸྑઌ୺Պֶٕज़େֶӃେֶ ৘ใՊֶ‫ڀݚ‬Պ [email protected] [email protected] [email protected] [email protected] [email protected] [email protected]. ⓒ 2017 Information Processing Society of Japan. ͷর໌৚͕݅มԽ͢Δ‫Ͱڥ؀‬͸ਤ 1 ͷΑ͏ʹর໌৚݅ʹ Αͬͯγʔϯͷ‫͕ํ͑ݟ‬มԽ͢Δɽ͜ͷ৔߹ɼΧϝϥҐஔ ͕ಉҰͰ͋ͬͯ΋ DB ʹଘࡏ͠ͳ͍ಛ௃ྔ͕‫ݕ‬ग़͞ΕͨΓ ಉ͡ 3 ࣍‫ݩ‬ҐஔʹରԠ͢Δಛ௃͕มԽͨ͠Γ͢ΔͨΊɼಛ ௃ྔϚονϯά͕ࠔ೉ʹͳΔޫ‫ݯ‬มಈ໰୊͕‫͖ى‬Δɽ ޫ‫ݯ‬มಈ໰୊ʹରͯؒ͠ԼΒ͸γϛϡϨʔγϣϯΛ༻͍. 1.

(2) Vol.2017-CVIM-207 No.19 2017/5/10. ৘ใॲཧֶձ‫ڀݚ‬ใࠂ IPSJ SIG Technical Report. ৾ಆ. ሺ‫ݔ‬ଵ ǡ ‫ݕ‬ଵǡ ‫ݖ‬ଵ ሻ. ৻਀્ඉ୤भ৭ল. ীഘ. ३঑গঞ‫ش‬३ঙথ. ॹ‫ॱش‬ঋ‫ش‬५. ્ඉ୤ীഘभেਛ. ং‫ॳش‬কঝ્ඉ୤द ছথॲ঒ইज़ঞ५ॺ॑৾ಆ. ীഘ୞ಀ. ਖ਼৒. োৡ઺൸. ৰ્ඉ୤॑ং‫ॳش‬কঝ ્ඉ୤ष૗ఌ. ਤ 2. . . ীഘ. ‫ڭ‬. ং‫ॳش‬কঝ઺൸. ঽഞ࿋ಀ৲ஓभ৾ಆ. ሺ‫ݔ‬ଶ ǡ ‫ݕ‬ଶ ǡ ‫ݖ‬ଶ ሻ. . ૘ਯभৰ઺൸ ৰ્ඉ୤पৌૢघॊ ং‫ॳش‬কঝ્ඉ୤भ৭උ. ং‫ॳش‬কঝ ્ඉ୤ীഘ. . ৌૢघॊ્ඉ୤ীഘभ ௓৒. ৰ્ඉ୤ীഘ 5$16$&द3Q3ਖ਻॑ ੰऎ. ఏҊख๏ͷֶश࣌ɼ൑ఆ࣌ͷྲྀΕ. ͯ DB Λ࡞੒͢Δख๏ΛఏҊͨ͠ [1]ɽ͜ͷΑ͏ͳख๏Ͱ. ‫ث‬ͷֶशσʔλΛੜ੒͠ɼֶशΛߦ͏ɽϥϯμϜϑΥϨε. ͸࣮ੈքͷର৅෺ͷ CG ϞσϧΛར༻ͨ͠όʔνϟϧ‫ڥ؀‬. τ͸ DB ͷಛ௃ྔΛֶशσʔλʹֶͯ͠शΛߦ͏ɽ൑ఆ࣌. Ͱଟ༷ͳޫ‫Ͱڥ؀ݯ‬ͷը૾Λੜ੒͢Δ͜ͱʹΑΓɼޫ‫ݯ‬ม. ͸ೖྗը૾ͷಛ௃ྔΛࣗ‫߸ූݾ‬Խ‫ʹث‬Αͬͯόʔνϟϧը. ಈʹରͯ͠‫ͳʹ݈ؤ‬Δ [2]ɽ·ͨόʔνϟϧը૾͸Χϝϥ. ૾ͷಛ௃ྔʹม‫͠׵‬ɼϥϯμϜϑΥϨετͰಛ௃ྔϚον. ͱ CG ϞσϧͷҐஔ͕‫طʹڞ‬஌Ͱ͋ΔͨΊɼ֤ըૉʹରԠ. ϯάΛߦ͏ɽͦͷϚονϯάͷத͔Β RANSAC[10] Λ༻. ͢Δ 3 ࣍‫఺ݩ‬Λਖ਼֬ʹಛఆͰ͖Δɽ͜ΕʹΑΓಉҰͷ 3 ࣍. ͍ͯ‫ޡ‬ରԠΛഉআ͠ɼPnP ໰୊Λղ͘͜ͱʹΑͬͯࣗ‫ݾ‬Ґ. ‫ʹ఺ݩ‬ରԠ͢Δಛ௃ྔू߹Λ 1 ͭͷΫϥεͱ‫͕ͱ͜͢ͳݟ‬. ஔਪఆΛߦ͏ɽ. Ͱ͖ɼಛ௃఺ͷରԠ෇͚໰୊ΛଟΫϥε෼ྨ໰୊ͱͯ͠ղ ͘͜ͱ͕Ͱ͖ΔɽͦͷҰํͰɼγϛϡϨʔγϣϯͱ‫ͱ࣮ݱ‬. 2.2 ࣗ‫߸ූݾ‬Խ‫ث‬ͷԠ༻ʹΑΔဃ཭໰୊ͷղܾ. ͷؒʹဃ཭͕ଘࡏ࣮͠ը૾ͷಛ௃ྔͱόʔνϟϧը૾ͷಛ. ࣗ‫߸ූݾ‬Խ‫ͱث‬͸ೖྗ x ͱग़ྗ y ͕ಉ͡ʹͳΔΑ͏ʹ. ௃ྔʹͣΕ͕ੜ͡Δ͜ͱ͕༧૝͞ΕΔɽ͜ͷ৔߹ɼ࣮ը૾. ֶश͞Εͨॱ఻ൖ‫ܕ‬χϡʔϥϧωοτϫʔΫʢNeural Net-. ͱ DB ͱͷಛ௃ྔϚονϯάʹࣦഊ͢ΔͨΊɼ࣮ը૾Λର. workʣͰ͋Δɽࣗ‫߸ූݾ‬Խ‫ث‬͸࣍‫ݩ‬ѹॖ΍σΟʔϓχϡʔ. ৅ͱ͢Δࣗ‫ݾ‬Ґஔਪఆͷਫ਼౓͕௿Լ͢Δɽ. ϥϧωοτϫʔΫʢDeep Neural Networkʣͷࣄલֶशʹ. ຊ‫Ͱڀݚ‬͸ޫ‫ݯ‬มಈ໰୊ͱဃ཭໰୊Λಉ࣌ʹղܾ͢Δɼ. ΋༻͍ΒΕΔɽࣗ‫߸ූݾ‬Խ‫ث‬ͷԠ༻ʹೖྗը૾ʹਓ޻తʹ. γϛϡϨʔγϣϯͱ‫ػ‬ցֶशΛ‫ͨ͠ʹج‬Χϝϥࣗ‫ݾ‬Ґஔਪ. ϊΠζΛ෇༩ͨ͠Γม‫ͨ͠ܗ‬Γ͢ΔͳͲͷલॲཧΛՃ͑. ఆख๏ΛఏҊ͢Δɽ·ͣχϡʔϥϧωοτϫʔΫͷҰͭͰ. ͯɼલॲཧΛ͔͚Δલͷը૾͕ग़ྗը૾ʹಘΒΕΔΑ͏ʹ. ͋Δࣗ‫߸ූݾ‬Խ‫ث‬ʢAutoencoderʣΛԠ༻࣮͠ը૾ͷಛ௃. ֶश͢ΔσϊΠδϯάʢDenoisingʣ͕͋Δɽ. ྔʹରԠ͢Δόʔνϟϧը૾ͷಛ௃ྔ΁ͷࣸ૾Λֶशͤ͞. ຊ‫Ͱڀݚ‬͸ࣗ‫߸ූݾ‬Խ‫ث‬ͷσϊΠδϯάͰͷ࢖͍ํΛࢀ. Δ͜ͱͰɼ࣮ը૾ͷಛ௃ྔʹ΋ରԠͰ͖ΔΑ͏ʹ͢Δɽ࣍. ߟʹͯ͠ɼ࣮ը૾͔ΒಘΒΕͨಛ௃ྔΛόʔνϟϧը૾͔. ʹϥϯμϜϑΥϨετʢRandom ForestʣΛར༻ͯ͠ಛ௃. ΒಘΒΕͨಛ௃ྔʹม‫͢׵‬ΔΑ͏ʹֶशͤ͞ΔɽҎ߱ɼ࣮. ྔϚονϯάΛߦ͍ɼಛ௃ྔϚονϯάੑೳͷ޲্ͱॲཧ. ը૾ͷಛ௃ྔΛʮ࣮ಛ௃ྔʯͱ‫ͼݺ‬ɼόʔνϟϧը૾ͷಛ. ࣌ؒͷ୹ॖΛਤΔɽ. ௃ྔΛʮόʔνϟϧಛ௃ྔʯͱ‫Ϳݺ‬ɽγϛϡϨʔγϣϯͰ. 2. ఏҊख๏ 2.1 ύΠϓϥΠϯ. ಘΒΕͨಛ௃ྔ෼෍ Dk ͸γϛϡϨʔγϣϯۭؒʹ͓͚Δ ̏࣍‫࠲ݩ‬ඪ pk ͱόʔνϟϧಛ௃ྔྻ {f ki } Ͱߏ੒͞ΕΔɽ ࣗ‫߸ූݾ‬Խ‫ث‬͸·ͣόʔνϟϧಛ௃ྔ f ki ಉ࢜Λֶशσʔ. ఏҊख๏ͷύΠϓϥΠϯΛਤ 2 ʹࣔ͢ɽ·ͣ‫ط‬ଘख๏ [1]. λ (f ki , f ki ) ͱͯ͠ࣄલֶशͤ͞ɼͦͷ‫ޙ‬ɼ࣮ಛ௃ྔ g j. ͱಉ༷ʹγϛϡϨʔγϣϯΛߦ͏͜ͱͰଟ༷ͳޫ‫ڥ؀ݯ‬Լ. ͱͦΕʹରԠ͢Δόʔνϟϧಛ௃ྔ f kσ(j) Λֶशσʔλ. ͷόʔνϟϧը૾Λऔಘ͢Δɽͦͷόʔνϟϧը૾ͷಛ௃. (g j , f kσ(j) ) ͱֶͯ͠शͤ͞ΔɽࣄલֶशΛߦ͏ͷ͸ɼόʔ. ྔͷத͔Β Kurz Βͷख๏ [9] ͰΧϝϥࢹ఺มಈʹ‫ͳ݈ؤ‬. νϟϧಛ௃ྔΛࣗ਎΁ม‫ͤ͞׵‬Δͱ͖ͷ‫਺ؔࠩޡ‬Λ࠷খʹ. ୅දಛ௃ྔΛબग़͠ɼ3 ࣍‫࠲ݩ‬ඪ͕͍ۙಛ௃ྔΛάϧʔϓ. ͢ΔॏΈͱɼ࣮ಛ௃ྔΛόʔνϟϧಛ௃ྔʹม‫ͤ͞׵‬Δͱ. Խͯ͠ಛ௃ྔ෼෍Λ࡞੒ͯ͠ DB Λߏங͢ΔɽDB ͷಛ௃. ͖ͷ‫਺ؔࠩޡ‬Λ࠷খʹ͢ΔॏΈ͸͍ۙҐஔʹ͋Δͱ༧૝͞. ྔͱগ਺ͷ࣮ը૾͔ΒಘΒΕΔಛ௃ྔΛ༻͍ͯࣗ‫߸ූݾ‬Խ. ΕΔ͔ΒͰ͋Δɽ͜ͷΑ͏ʹֶशͤ͞Δͱࣗ‫߸ූݾ‬Խ‫ث‬ͷ. ⓒ 2017 Information Processing Society of Japan. 2.

(3) Vol.2017-CVIM-207 No.19 2017/5/10. ৘ใॲཧֶձ‫ڀݚ‬ใࠂ IPSJ SIG Technical Report. Virtual. Ȱ௩௜௥௧௨௔௟. Virtual ਈ੺ວ১. Ȱ௥௘௔௟. ஜ౦৲. Ԅ Ʌ. Real. ୭୆. ୭୆. ୭୆. Real. ୭୆. ୭୆. ‫ܭ‬ਈ੺ວ১. ਤ 3. ࠷ۙ๣๏ʹΑΔಛ௃ྔͷબ୒ʢࠨʣͱന৭Խͱ K ࠷ۙ๣๏ʹ ΑΔಛ௃ྔͷબ୒ʢӈʣ ͷιࠃΊ. ग़ྗ y(g j ) ͸όʔνϟϧಛ௃ྔͱಉ͡Α͏ʹѻ͏͜ͱ͕Ͱ ͖Δɽ ਤ 4. ޫ‫޲ํݯ‬ʢࠨ্ʣ ɼ࣮‫͍༻ʹݧ‬Δޫ‫ڥ؀ݯ‬ʢӈ্ʣ ɼΧϝϥҐஔ ʢࠨԼʣ ɼ࠶౤Ө‫ࠩޡ‬Λௐ΂Δ఺ʢӈԼʣ. 2.3 ࣮ಛ௃ྔʹରԠ͢Δόʔνϟϧಛ௃ྔͷબ୒ ࣗ‫߸ූݾ‬Խ‫ث‬ͷֶशʹ͸࣮ಛ௃ྔʹରԠ͢Δόʔνϟϧ ಛ௃ྔ͕ඞཁͰ͋Δɽ͜͜Ͱ࣮ը૾ In ͱɼ֤࣮ը૾ʹର Ԡ͢Δ‫͚͓ʹۭ࣮ؒݱ‬Δ 3 ࣍‫࠲ݩ‬ඪΛը૾࠲ඪ΁౤Ө͢Δ. ఆ͞ΕΔͷͰɼ‫ط‬ଘख๏ [1] ͱൺֱͯ͠ޫ‫ݯ‬มಈʹରͯ͠. ౤Өߦྻ P n ͕‫ط‬஌ͱԾఆ͢Δɽ͜ΕΒ͔Β࣮ಛ௃ྔ g j. ΑΓ‫ͳʹ݈ؤ‬Δͱ‫ظ‬଴Ͱ͖Δɽ. ΛಉҰ 3 ࣍‫ʹ఺ݩ‬ରԠ͢Δ΋ͷಉ࢜ͰάϧʔϓԽ͠ɼόʔ. ͔࣮͠͠ը૾ͷ௿ίϯτϥετͳྖҬ͔Βಛ௃఺Λ‫ݕ‬ग़. νϟϧಛ௃ྔ෼෍ Dk ʹରԠ͚ͮΒΕΔɽ͔࣮͠͠ಛ௃ྔ. ͢Δ৔߹ɼಛ௃ྔ‫ه‬ड़ࢠͷᮢ஋Λऑ͘ઃఆ͢Δඞཁ͕͋Δɽ. g j ͕Ͳͷόʔνϟϧಛ௃ྔ f ki ʹਅʹରԠ͢Δ͔͸ෆ໌. ͦͷΑ͏ʹͨ͠৔߹ DB ʹଘࡏ͠ͳ͍ಛ௃఺΋‫ݕ‬ग़͞ΕΔ. Ͱ͋ΔͨΊɼಛ௃ྔͷରԠ෇͚͕໰୊ʹͳΔɽ୯७ʹ࠷ۙ. ͕ɼϥϯμϜϑΥϨετΛ༻͍Δͱ͍ͣΕ͔ͷಛ௃ྔ෼෍. ๣ͷόʔνϟϧಛ௃ྔΛબ୒͢Δͱਤ 3 ࠨͷΑ͏ʹ෼෍ͷ. ʹϚονϯά͢Δɽ͜ΕΒͷϚονϯά݁ՌΛ RANSAC. Ґஔ͕ͣΕ͍ͯͨ৔߹ɼDk ͷ‫ڥ‬ք෇ۙ΁ࣸ૾͞ΕΔͨΊɼ. ΁ೖྗ͠ͳ͍Α͏ʹ͢ΔͨΊɼϥϯμϜϑΥϨετͰಘΒ. ‫ޙ‬ஈͷಛ௃ྔϚονϯά͕ࠔ೉ʹͳΔɽͦ͜Ͱຊख๏Ͱ͸. ΕͨϚονϯάͷ͏ͪ M ‫ݸ‬Λબग़͢ΔɽϥϯμϜϑΥϨ. ࣮ಛ௃ྔ෼෍ͱόʔνϟϧಛ௃ྔʹന৭ԽΛߦͬͯಛ௃ྔ. ετ͸ೖྗ g j ͕ಛ௃ྔ෼෍ Dk ʹରԠ͢Δ֬཰ P (Dk |g j ). ͷରԠ෇͚Λߦ͏ɽന৭ԽʹΑͬͯฏ‫ͱۉ‬෼ࢄ͕ਖ਼‫ن‬Խ͞. Λ‫ٻ‬ΊΔ͜ͱ͕Ͱ͖Δɽͦͷ࠷େ໬౓ maxk P (Dk |g j ) Ͱ. Εۭͨؒʹ͓͍ͯ࠷ۙ๣ͷಛ௃ྔΛ‫ٻ‬ΊΕ͹ɼਤ 3 ӈͷΑ. શͯͷϚονϯάΛ߱ॱʹฒͼସ͑ɼ্Ґ M ‫ݸ‬ͷϚον. ͏ʹ֤࣮ಛ௃ྔ g j ͸όʔνϟϧಛ௃ྔ෼෍ Dk ʹຬวͳ. ϯάΛ࠾༻͢Δɽ. ࣸ͘૾͞ΕΔɽ͞Βʹ K ࠷ۙ๣๏Λ༻͍ͯ 1 ͭͷ࣮ಛ௃ྔ. 3. ࣮‫࡯ߟͱݧ‬. ʹ͖ͭ K ‫ݸ‬ͷόʔνϟϧಛ௃ྔΛରԠͤ͞Δ͜ͱͰɼࣗ ‫߸ූݾ‬Խ‫ث‬ͷֶशσʔλΛΑΓଟ֬͘อͰ͖Δɽ·ͨ K. 3.1 γϛϡϨʔγϣϯ‫ڥ؀‬ γϛϡϨʔγϣϯ༻ͷϞσϧʹ͸ DTU Robot Image. ͷ஋Λ࣍ࣜͷΑ͏ʹόʔνϟϧಛ௃ྔ෼෍ Dk ͕࣋ͭόʔ. Data Set ͷ MVS Set[11] Λར༻࣮ͯ͠‫ݧ‬Λߦ͏ɽ͜ͷσʔ. νϟϧಛ௃ྔ਺ |Dk | ʹԠͯ͡ௐ੔͢Δɽ. K = r. |Dk |. ληοτʹ͸ 49 ͔ॴͷΧϝϥࢹ఺͔ΒࡱӨͨ͠ 7 छྨͷ. (1). ͜͜Ͱ r ͸ൺ཰Λද͢ύϥϝʔλͰ͋Δɽલஈͷ୅දಛ௃ ྔͷબग़ʹΑΓɼ|Dk | ͕େ͖͍΄ͲΧϝϥࢹ఺มಈʹ‫݈ؤ‬ ͳಛ௃ྔͰ͋Δͱ‫͑ݴ‬ΔͨΊɼϚονϯά͠΍͍͢ಛ௃ྔ Λॏ఺తʹֶश͢Δ͜ͱ͕Ͱ͖Δɽ. 2.4 ϥϯμϜϑΥϨετʹΑΔࣝผੑೳͷ޲্ ຊख๏Ͱ͸ϥϯμϜϑΥϨετΛ༻͍ͯಛ௃ྔϚονϯ άΛߦ͏ɽϢʔΫϦου‫཭ڑ‬΍ϚϋϥϊϏε‫[ ཭ڑ‬1] Λ༻ ͍ͨಛ௃ྔϚονϯάͱൺֱͯ͠ɼϥϯμϜϑΥϨετʹ ΑΔϚονϯάͰ͸ಛ௃ྔ෼෍ͷ‫ܗ‬ঢ়͕ෳࡶͳ৔߹Ͱ΋ਖ਼ ͍ܾ͠ఆ‫ڥ‬քΛֶशͯ͠Ϛονϯά͢Δ͜ͱ͕‫ظ‬଴Ͱ͖ Δɽޫ‫ݯ‬มಈʹΑͬͯಛ௃ྔ෼෍ͷ‫ܗ‬ঢ়͸ෳࡶʹͳΔͱ૝ ⓒ 2017 Information Processing Society of Japan. ޫ‫͚͓ʹڥ؀ݯ‬Δը૾σʔλͱɼͦͷը૾͔Β Structure. from MotionʢSfMʣΛ༻͍ͯ෮‫܈఺ͨ͠ݩ‬σʔλ͕‫·ؚ‬ Ε͍ͯΔɽ఺‫܈‬σʔλʹ͸৭৘ใ͕ܽଛ͍ͯͨͨ͠Ίɼը ૾σʔλΛ༻͍֤ͯ௖఺ʹରԠ͢Δըૉ஋ͷதԝ஋Λ௖఺ ͷ৭ͱͯ͠࠾༻ͨ͠ [12]ɽ ޫ‫ݯ‬มԽγϛϡϨʔγϣϯ͸ Unity 5*1 ্Ͱߦ͏ɽޫ‫ݯ‬ ʹ͸ฏߦޫ‫͋Ͱݯ‬Δ Directional Light Λ༻͍Δɽਤ 4 ࠨ ্ͷΑ͏ʹର৅෺Λத৺ͱͨ͠൒‫্ٿ‬ͷҐஔ (θ, φ) ΛมԽ ͤ͞ɼͦͷҐஔ͔Βର৅෺த৺Λ޲͘ํ޲Λޫ‫͢ͱ޲ํݯ‬ ΔɽDB Λੜ੒͢ΔͨΊͷը૾͸ਤ 4 ӈ্ʹࣔ͢ θ ∈ {10}ɼ. φ ∈ {30, 60, · · · , 150} ͷޫ‫Ͱڥ؀ݯ‬ɼਤ 4 ࠨԼʹࣔ͢ 16 Օ ॴͷ஍఺͔Β֯౓Λ 5 ౓ͣͭมԽͤͯ͞ࡱӨ͢Δɽ *1. https://unity3d.com. 3.

(4) Vol.2017-CVIM-207 No.19 2017/5/10. ৘ใॲཧֶձ‫ڀݚ‬ใࠂ IPSJ SIG Technical Report. 3.2 ϥϯμϜϑΥϨετͷੑೳධՁ. . ͷͨΊʹɼόʔνϟϧը૾Λ༻͍ͨࣗ‫ݾ‬Ґஔਪఆ࣮‫ݧ‬Λߦ ͏ɽςετʹ༻͍Δը૾͸ DB ੜ੒༻ͷը૾ͱ͸ผʹ༻ҙ ͢ΔɽΧϝϥҐஔΛϥϯμϜʹબ୒͠ɼର৅෺த৺Λ޲͍ ͯ 1 ޫ‫ͨ͋ڥ؀ݯ‬Γ 100 ຕͷը૾Λੜ੒͢Δɽࣗ‫ݾ‬Ґஔਪ ఆͷ੒ޭɼࣦഊ͸ਤ 4 ӈԼʹࣔ͢ 12 ͔ॴͷ੨‫ٿ‬ͷ࠶౤Ө. ঽഞਜ਼઼௓৒ਛ඘૨. ϥϯμϜϑΥϨετʹΑΔಛ௃ྔϚονϯάͷੑೳධՁ. ‫ࠩޡ‬ʢRe-projection ErrorʣͰධՁ͢Δɽ. 1 E= 12. 12 . . ৰ્ඉ୤ँॉ. . ৰ્ඉ୤ऩख. . ||K[R|t]ui −. x∗i ||. . . . . ⃦க>SL[HO@. (2). i=1. ਤ 5. ࣮ಛ௃ྔͷ༗ແʹΑΔࣗ‫ݾ‬Ґஔਪఆ੒ޭ཰ͷࠩҟ. ͜͜Ͱ K ͸Χϝϥͷ಺෦ύϥϝʔλߦྻɼ[R|t] ͸ਪఆ͞ ΕͨΧϝϥҐஔ࢟੎ɼui ͸੨‫ٿ‬ͷ 3 ࣍‫࠲ݩ‬ඪɼx∗i ͸֤ς ετը૾ʹ͓͚Δ੨‫ٿ‬ͷ 2 ࣍‫࠲ݩ‬ඪͷਅ஋Ͱ͋Δɽຊ࣮‫ݧ‬ Ͱ͸ 640 × 480 ͷը૾ʹରͯ͠ E < 5.0[pixel] ͳΒ͹ࣗ‫ݾ‬ Ґஔਪఆʹ੒ޭͨ͠ͱ‫͢ͳݟ‬ɽ ൺֱख๏ͱͯ͠ϢʔΫϦου‫ʹ཭ڑ‬ΑΔϚονϯάͱɼ ࠷ۙ๣๏ʢNearest Neighbor; NNʣʹΑΔϚονϯάΛߦ ͏ɽϢʔΫϦου‫ʹ཭ڑ‬ΑΔϚονϯά͸ɼςετը૾ͷ ֤ಛ௃ྔͱ DB ಺ͷ֤ಛ௃ྔ෼෍ͷฏ‫ͱۉ‬ͷϢʔΫϦου ‫࠷͕཭ڑ‬΋͍ۙ෼෍ʹϚονϯά͢Δɽ࠷ۙ๣๏͸ςετ ը૾ͷ֤ಛ௃ྔͱ DB ಺ͷશಛ௃ྔͱͷϢʔΫϦου‫཭ڑ‬ ͕࠷΋͍ۙ෼෍ʹϚονϯά͢Δɽ ࣮‫ݧ‬ͷ݁Ռɼද 1 ͷΑ͏ʹఏҊख๏͸͍ͣΕͷޫ‫ڥ؀ݯ‬. ਤ 6. ࣮ಛ௃ྔ͋Γʢ্ʣͱ࣮ಛ௃ྔͳ͠ʢԼʣͷϚονϯά݁Ռ. ʹ͓͍ͯ΋ 90 %Λ௒͑Δਫ਼౓Λग़͓ͯ͠ΓɼϢʔΫϦο υ‫ʹ཭ڑ‬ΑΔϚονϯά΍࠷ۙ๣๏ͱൺ΂ͯ୹͍࣌ؒͰ൑. Δ͜ͱͰɼ‫ͱ࣮ݱ‬ͷဃ཭ʹରͯ͠‫ͳʹ݈ؤ‬Δ͜ͱΛ֬ೝ͢. ఆͰ͖ͨɽ‫ ڥ؀‬1ɼ5 ͸র໌͕ର৅෺ͷҰ෦ʹ͔͠౰ͨΒ. ΔͨΊ࣮ը૾Λ༻͍ͨࣗ‫ݾ‬Ґஔਪఆ࣮‫ݧ‬Λߦ͏ɽ30 ຕͷ࣮. ͳ͍ͨΊɼಛ௃ྔ෼෍ͷ‫ܗ‬ঢ়͕ෳࡶʹͳ͍ͬͯΔ͜ͱ͕૝. ը૾Λࣗ‫߸ූݾ‬Խ‫ث‬ͷֶशσʔλੜ੒ʹ༻͍ͯɼͦͷ 30 ຕ. ఆ͞ΕΔɽϢʔΫϦου‫ʹ཭ڑ‬ΑΔϚονϯάͰ͸ͦͷෳ. Λ‫ؚ‬Ή 60 ຕͰࣗ‫ݾ‬ҐஔਪఆΛߦ͏ɽఏҊख๏ͷ࣮ಛ௃ྔ. ࡶ͞ʹରԠͰ͖ͣ௿͍੒ޭ཰ͱͳ͍ͬͯΔ͕ɼఏҊख๏͸. Λόʔνϟϧಛ௃ྔ΁ม‫͢׵‬Δࣗ‫߸ූݾ‬Խ‫ث‬Λ༻͍ͨ৔߹. ͦͷෳࡶ͞ʹରԠͰ͖ͨͨΊߴ͍੒ޭ཰͕ಘΒΕͨͱߟ͑. ͱɼֶशσʔλʹ࣮ಛ௃ྔΛ༻͍ͣʹόʔνϟϧಛ௃ྔΛ. ΒΕΔɽ·ͨɼϢʔΫϦου‫ʹ཭ڑ‬ΑΔϚονϯά͸ 1 ͭ. ࣗ਎ʹม‫͢׵‬ΔΑ͏ʹֶशͨࣗ͠‫߸ූݾ‬Խ‫ث‬Λ༻͍ͨ৔߹. ͷಛ௃ྔʹ͖ͭ DB ಺ͷશಛ௃ྔ෼෍਺ͱͷϚονϯάΛ. Ͱൺֱ͢Δɽͦͷ݁ՌΛਤ 5 ʹ·ͱΊΔɽਤ 5 ͸ 640 × 480. ߦ͍ɼ࠷ۙ๣๏͸ 1 ͭͷಛ௃ྔʹ͖ͭ DB ಺ͷશಛ௃ྔͱ. ͷը૾ʹର͢Δ࠶౤Ө‫ ࠩޡ‬E ͷᮢ஋ΛมԽͤͨ͞ͱ͖ͷࣗ. ͷϚονϯάΛߦ͏ɽҰํɼఏҊख๏Ͱ͸ϥϯμϜϑΥϨ. ‫ݾ‬Ґஔਪఆ੒ޭ཰Λදͨ͠άϥϑͰ͋ΔɽఏҊख๏ͷ࣮ಛ. ετΛ༻͍ΔͨΊɼ໦ͷਂ͞ͷ਺͚ͩԋࢉΛߦ͑͹Α͘ɼ. ௃ྔ͋Γͷࣗ‫߸ූݾ‬Խ‫ث‬Λ༻͍ͨ৔߹ɼ80 %Ҏ্ͷը૾͕. ୹͍࣌ؒͰ൑ఆ͕Ͱ͖Δɽ. ࠶౤Ө‫ࠩޡ‬Λ 10 pixel ະຬʹ཈͑ΒΕ͓ͯΓဃ཭ͷӨ‫͕ڹ‬ ड͚ʹ͘͘ͳ͍ͬͯΔ͜ͱ͕෼͔Δɽ. 3.3 ࣗ‫߸ූݾ‬Խ‫ث‬ͷੑೳධՁ. ਤ 6 ͸ಉҰը૾ʹର͢Δ࣮ಛ௃ྔ͋Γͷࣗ‫߸ූݾ‬Խ‫ث‬Λ. ࣗ‫߸ූݾ‬Խ‫࣮Ͱث‬ಛ௃ྔΛόʔνϟϧಛ௃ྔ΁ࣸ૾ͤ͞. ༻͍ͨ৔߹ͱ࣮ಛ௃ྔͳ͠ͷࣗ‫߸ූݾ‬Խ‫ث‬Λ༻͍ͨ৔߹ ͷϚονϯά݁ՌͰ͋Δɽ֤ਤͷࠨଆ͕ೖྗը૾Ͱ͋Γɼ. ද 1. όʔνϟϧը૾Ͱͷࣗ‫ݾ‬Ґஔਪఆ࣮‫ݧ‬ͷ݁Ռ ੒ޭ཰ ൑ఆ࣌ؒ [msec]. Env. Proposed Euclid. NN. Proposed. Euclid. ӈଆ͕ DB ಺ͷಛ௃ྔ෼෍ʹϚονϯάͨ͠ҐஔΛද͠ NN. ͍ͯΔɽ࣮ಛ௃ྔͳ͠ͷϚονϯά݁ՌΛ‫ݟ‬Δͱ΄ͱΜͲ ͷϚονϯά͕ࣦഊ͍ͯ͠Δɽ͜ͷ͜ͱ͔Βόʔνϟϧಛ. 1. 92%. 77%. 88%. 275.044. 638.552 17276.0. 2. 98%. 92%. 99%. 275.689. 717.094 20169.3. 3. 99%. 98%. 100%. 275.492. 1041.90 29407.4. ௃ྔΛਖ਼͘͠෼ׂ͢Δ͜ͱ͕Ͱ͖ͳ͍͜ͱ͕෼͔Δɽैͬ. 4. 98%. 97%. 100%. 277.485. 1296.85 37202.6. ͯɼ࣮ಛ௃ྔͷ෼෍ͱόʔνϟϧಛ௃ྔͷ෼෍ͱͷؒʹͣ. 5. 97%. 87%. 94%. 276.681. 1073.62 30180.2. Ε͕ଘࡏ͢Δͱ༧૝͞ΕɼͦͷͣΕΛന৭Խͱ࣮ಛ௃ྔΛ. ⓒ 2017 Information Processing Society of Japan. ௃ྔͰֶश͞ΕͨϥϯμϜϑΥϨετͷܾఆ‫ڥ‬քͰɼ࣮ಛ. 4.

(5) Vol.2017-CVIM-207 No.19 2017/5/10. ৘ใॲཧֶձ‫ڀݚ‬ใࠂ IPSJ SIG Technical Report. ༻͍ͨࣗ‫߸ූݾ‬Խ‫ʹث‬Αͬͯղܾ͢Δ͜ͱ͕Ͱ͖ͨͱߟ͑ ΒΕΔɽ. [12]. 4. ࠓ‫ޙ‬ͷ՝୊ ೖྗը૾͔ΒಘΒΕΔ࣮ಛ௃ྔͷத͔ΒɼDB ʹରԠ͢ Δόʔνϟϧಛ௃ྔ͕͋Δಛ௃ྔΛબ୒͢Δ࢓૊Έͷಋೖ. sion and Pattern Recognition (CVPR), pp. 406–413, 2014. Daniel N. Wood, Daniel I. Azuma, Ken Aldinger, Brian Curless, Tom Duchamp, David H. Salesin, and Werner Stuetzle, “Surface Light Fields for 3D Photography,” ACM Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pp. 287–296, 2000.. ͢Δɽ͜ΕʹΑΓɼߴ଎ͳࣗ‫ݾ‬Ґஔਪఆ͕ՄೳʹͳΔͱ‫ظ‬ ଴Ͱ͖Δɽ ँࣙ. ຊ ‫ ڀ ݚ‬ͷ Ұ ෦ ͸ JSPS Պ ‫ ݚ‬අ JP16H02858 ͱ. JP16K16100 ͷॿ੒Λड͚ͨɽ ࢀߟจ‫ݙ‬ [1]. [2]. [3]. [4]. [5]. [6]. [7]. [8]. [9]. [10]. [11]. Tomohiro Mashita, Alexander Plopski, Akira Kudo, Tobias Höellerer, Kiyoshi Kiyokawa, and Haruo Takemura, “Simulation based Camera Localization under a Variable Lighting Environment,” International Conference on Artificial Reality and Telexistence (ICAT), 2016. ޻౻ জ, Alexander Plopski, Tobias H¨ ollerer, ؒԼ Ҏେ, ਗ਼઒ ਗ਼, ஛ଜ ࣏༤, “ҟͳΔޫ‫͚͓ʹڥ؀ݯ‬Δը૾ಛ௃ͷ ‫݈ؤ‬ੑͷௐࠪ”ɼ৘ใॲཧֶձ‫ڀݚ‬ใࠂɼίϯϐϡʔλϏ δϣϯͱΠϝʔδϝσΟΞʢCVIMʣ, Vol. 2015, No. 65, 2015. Hirokazu Kato and Mark Billinghurst, “Marker Tracking and HMD Calibration for a Video-based Augmented Reality Conferencing System,” IEEE and ACM International Workshop on Augmented Reality (IWAR), pp. 85–94, 1999. Ronald Azuma, Bruce Hoff, Howard Neely III, and Ron Sarfaty, “A Motion-Stabilized Outdoor Augmented Reality System,” IEEE Virtual Reality (IEEE VR), pp. 252– 259, 1999. Javier J. M. Diaz, Rodrigo de A. Maués, Rodrigo B. Soares, Eduardo F. Nakamura, and Carlos M. S. Figueiredo, “Bluepass: An Indoor Bluetooth-based Localization System for Mobile Applications,” IEEE Symposium on Computers and Communications (ISCC), pp. 778–783, 2010. Yukiko Shinozuka, Francois De Sorbier, and Hideo Saito, “Specular 3D Object Tracking by View Generative Learning,” Irish Machine Vision and Image Processing Conference (IMVIP), pp. 9–14, 2014. Gilles Simon, “Tracking-by-Synthesis Using Point Features and Pyramidal Blurring,” IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pp. 85–92, 2011. ബ ॆ޹ ଞ, “ϥϯυϚʔΫσʔλϕʔεʹ‫ͮ͘ج‬੩ࢭը ૾͔ΒͷΧϝϥҐஔɾ࢟੎ਪఆ”, ೔ຊόʔνϟϧϦΞϦ ςΟֶձ࿦จࢽ, Vol. 13, No. 2, pp. 161–170, 2008. Daniel Kurz, Thomas Olszamowski, and Selim Benhimane, “Representative Feature Descriptor Sets for Robust Handheld Camera Localization,” IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pp. 65–70, 2012. Martin Fischler and Robert Bolles, “Random Sample Consensus: A Paradigm for Model Fitting with Applications to Image Analysis and Automated Cartography,” Communications of the ACM, Vol. 24, No. 6, pp. 381– 395, 1981. Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanaes, “Large Scale Multi-view Stereopsis Evaluation,” IEEE Conference on Computer Vi-. ⓒ 2017 Information Processing Society of Japan. 5.

(6)