Microsoft Word - 11-王思明_oee180008_xeQzyx.doc

Size: px

Start display at page:

Download "Microsoft Word - 11-王思明_oee180008_xeQzyx.doc"

服绶花
5 years ago
Views:

Opto-Electronc Engneerng 光电工程 Artcle 2018 年, 第 45 卷, 第 10 期复杂动态背景下的运动目标检测王思明, 韩乐乐 * 730070 摘要 : 为实现复杂动态背景下快速准确地检测运动目标, 提出一种改进二进制鲁棒不变尺度特征 (BRISK) 算法的运动目标检测方法首先对图像进行分块, 利用图像熵对图像块进行筛选 ;

7% 的特征点, 并且匹配效率提高了 75%, 处理速度比以往算法快, 并且具有较强的抗噪性能关键词 : 运动目标检测 ;BRISK 算法 ; 图像熵 ; 欧氏距离中图分类号 :TP391.41 文献标志码 :A 引用格式 : 王思明, 韩乐乐. 复杂动态背景下的运动目标检测 [J].

1 Opto-Electronc Engneerng 光电工程 Artcle 2018 年, 第 45 卷, 第 10 期复杂动态背景下的运动目标检测王思明, 韩乐乐 * 摘要 : 为实现复杂动态背景下快速准确地检测运动目标, 提出一种改进二进制鲁棒不变尺度特征 (BRISK) 算法的运动目标检测方法首先对图像进行分块, 利用图像熵对图像块进行筛选 ; 然后针对特征匹配过程中存在大量误匹配的问题, 采用 k 近邻算法与欧氏距离进行特征匹配 ; 最后通过改进的顺序抽样一致性算法进行特征点提纯, 进一步完成背景运动补偿, 从而利用形态学处理分割运动目标采用多组视频图像进行验证, 本文算法在原 BRISK 算法的基础上去除了 32.7% 的特征点, 并且匹配效率提高了 75%, 处理速度比以往算法快, 并且具有较强的抗噪性能关键词 : 运动目标检测 ;BRISK 算法 ; 图像熵 ; 欧氏距离中图分类号 :TP 文献标志码 :A 引用格式 : 王思明, 韩乐乐. 复杂动态背景下的运动目标检测 [J]. 光电工程,2018,45(10): Movng object detecton under complex dynamc background Wang Smng, Han Lele * School of Automaton & Electrcal Engneerng, Lanzhou Jao Tong Unversty, Lanzhou, Gansu , Chna Abstract: In order to realze fast and accurate detecton of movng targets under complex dynamc background, a movng object detecton method based on BRISK (bnary robust nvarant scalable keyponts) algorthm s proposed. Frstly, the mage s dvded nto blocks, and the mage blocks are fltered by usng mage entropy. Then, amng at the problem of large number of msmatch n the process of feature matchng, the k-nearest neghbor algorthm and Eucldean dstance are used to perform feature matchng. Fnally, the mproved sequental samplng consstency algorthm s used to refne the feature ponts and further completes the background moton compensaton, and morphologcal processng s used to segment the movng target. Through the verfcaton of multple vdeo mages, the proposed algorthm removes 32.7% of the feature ponts based on the orgnal BRISK algorthm and mproves the matchng effcency by 75%. The proposed algorthm has faster processng speed than prevous algorthms and strong ant-nose performance. Keywords: movng object detecton; BRISK algorthm; mage entropy; Eucldean dstance Ctaton: Wang S M, Han L L. Movng object detecton under complex dynamc background[j]. Opto-Electronc Engneerng, 2018, 45(10): 收稿日期 : ; 收到修改稿日期 : 基金项目 : ( ) 作者简介 :(1964-)E-mal @qq.com 通信作者 :(1992-)E-mal hlldyx12@163.com

1 引言 [1] [2-3] [4] [5] SIFT(scale-nvarant feature transform) SIFT [6] SURF(speeded up robust features) [7] ORB(orented fast and rotated BRIEF) ORB PROSAC(progressve sample consensus) BRISK(bnary

2 1 引言 [1] [2-3] [4] [5] SIFT(scale-nvarant feature transform) SIFT [6] SURF(speeded up robust features) [7] ORB(orented fast and rotated BRIEF) ORB PROSAC(progressve sample consensus) BRISK(bnary robust nvarant scalable keyponts) k PROSAC 2 改进的 BRISK 特征匹配算法 2.1 特征点检测 1) n octave ( c ) n ntra-octave ( d ) c 0 octave c 0 2 ntra-octave c c d c 0 w wc ( ) = 2, wd ( ) = (1) 2) octave ntra-octave FAST9-16 ( ) n=4 8 c 0 FAST5-8 ( d(-1) ) 9 3) 9 (8 )( 18 ) 26 FAST 4) 3 FAST FAST 2.2 特征描述 N 1 x/pxel y/pxel 图 1 采样模式 Fg. 1 Samplng mode N N(N-1)/2 A, (2) A = p p R R < N j< j N 2 2 {(, j), } I( p, σ ) I( pj, σ j ) σ I( pj, σ j) I( p, σ ) g( p, pj) = ( pj p) (3) 2 pj p S L [8] S = {( p, p ) A p p < δ } A, (4) j j max

3 L= {( p, p ) A p p > δ } A, (5) j j δ max = 9.75t δ mn = 13.67t δ t gx 1 g = = g( p, pj) (6) g y L ( p, pj) L 512 = arctan( g y, g x ) 512 bt bt 1, I( pj, σ j) > I( p, σ) b, ( p, = pj ) S (7) 0, otherwse Str 1 Str = 1 mn D( Str, Str ) = x y (8) 2.3 基于图像熵的图像块筛选 [9] [6] [10] f 255 H = f log f (9) = 0 1) 8 2) 基于 k 近邻与欧氏距离的特征匹配算法 2 [7] k k 1) k P 1 P 2 P 3 P 1 P 2 P 3 d 1 d 2 2) r d 1 /d 2 <r P 2 P 1 3) S 1 S 2 S 1 S 2 S 1 S 2 S 3 S 3 m 4) S 3 m S 1 S 2 P 1 P 一种基于改进的 PRAOSAC 算法的特征点提纯 [11] [7] PROSAC PROSAC RANSAC(random sample consensus) PROSAC 3 1)

22 y, (10) y a11 a12 t b21 b 22 x t y 2.

4 2) N 3 e s 3) 3 2) 4) 2.6 运动补偿 2.5 [12] x a t 11 a12 x x y = + b t 21 b 22 y, (10) y a11 a12 t b21 b 22 x t y 2.7 运动目标提取 It 1 ( xy, ) It ( xy, ) T 1 It 1( xy, ) It( xy, ) > T dt 1, t( x, y) =, (11) 0 otherwse T Otsu q ω 1 μ 1 ω 2 μ 2 ε 2 ε = ω1 ω2 ( μ2 μ1) (12) Otsu ε q q q 实验结果与分析 Intel core(tm) M 2.4 GHz PC 4 G C++ opencv Vso Studo2013 BRISK k 图 2 算法流程 PROSAC Fg. 2 Algorthm flowchart (a) (b) (c) 图 3 三组视频图像 (a) 第一组视频序列 V1;(b) 第二组视频序列 V2;(c) 第三组视频序列 V3 Fg. 3 Vdeo mages. (a) The frst set of vdeo sequences V1; (b) The second set of vdeo sequences V2; (3) The thrd set of vdeo sequences V

5 3 V1 V pxels 720 pxels 3 k r 特征点数目及算法实时性分析 SIFT SURF BRISK [7] 1 V1 V3 1 BRISK 32.7% 1.1 s SIFT 15 SURF 4 [7] ORB V1 V2 V3 ORB 0.38 s 3.2 算法匹配准确度分析 V1 V3 RANSAC k 2 2 RANSAC 50% k 75% 84% 3.3 算法检测效果对比 V1 V3 SIFT SURF [7] (a) 5(a) 6(a) SIFT 4(b) 5(b) 6(b) SURF 4(c) 5(c) 6(c)[7] ORB 4(d) 5(d) 6(d) BRISK 4 1 SIFT SURF SIFT SURF 4(c) 6(c) 4(d) 6(d)[7] 表 1 不同算法特征点数目及耗时对比结果 Table 1 Numbers of feature ponts and tme-consumng of dfferent algorthms Vdeo mages Feature ponts/(tme-consumng/s) SIFT SURF BRISK Proposed method Reference [7] V1 3639/ / / / /1.06 V2 2378/ / / / /0.94 V3 488/ / / / /0.95 表 2 特征匹配准确度对比结果 Table 2 Comparson of feature matchng accuracy Algorthms V1 V2 V3 Orgnal featureponts Brute-force +RANSAC 253/56.7% 181/51.6% 32/50% KNNMatch 98/83.2% 90/75.9% 7/89.1% Proposed method 83/85.8% 58/84.5% 5/92.2% (a) (b) (c) (d) 图 4 视频序列 V1 检测效果对比 (a) SIFT 算法 ;(b) SURF 算法 ;(c) 文献 [7] 算法 ;(d) 本文算法 Fg. 4 Comparson of vdeo sequence V1 detecton. (a) SIFT algorthm; (b) SURF algorthm; (c) Reference [7] algorthm; (d) Proposed algorthm paper

6 (a) (b) (c) (d) 图 5 视频序列 V2 检测效果对比 (a) SIFT 算法 ;(b) SURF 算法 ;(c) 文献 [7] 算法 ;(d) 本文算法 Fg. 5 Comparson of vdeo sequence V2 detecton. (a) SIFT algorthm; (b) SURF algorthm; (c) Reference [7] algorthm; (d) Proposed algorthm paper (a) (b) (c) (d) 图 6 视频序列 V3 检测效果对比 (a) SIFT 算法 ;(b) SURF 算法 ;(c) 文献 [7] 算法 ;(d) 本文算法 Fg. 6 Comparson of vdeo sequence V3 detecton. (a) SIFT algorthm; (b) SURF algorthm; (c) Reference [7] algorthm; (d) Proposed algorthm paper [7] ORB FAST AGAST( ) 4 结论 BRISK BRISK k 参考文献 [1] Yang K W, Duan W W, Yang X Z. Infra-vsble vdeo fuson object trackng under rany nght condton[j]. Chnese Journal of Scentfc Instrument, 2016, 17(5): 吴克伟, 段伟伟, 杨学志. 雨夜条件下的红外可见光视频融合目标跟踪 [J]. 仪器仪表学报, 2016, 17(5): [2] Akshay S. Sngle movng object detecton and trackng usng Horn-Schunck optcal flow method[j]. Internatonal Journal of Appled Engneerng Research, 2015, 10(11): [3] Lu H B, Chang F L. Movng object detecton by optcal flow method based on adaptve weght coeffcent[j]. Optcs and Precson Engneerng, 2016, 24(2): 刘洪彬, 常发亮. 权重系数自适应光流法运动目标检测 [J]. 光学精密工程, 2016, 24(2): [4] Yu H F, Lu W, Yuan H, et al. Movng object detecton based on sub-block moton compensaton[j]. Acta Electronca Snca, 2017, 45(1): 于红绯, 刘威, 袁淮, 等. 基于子块运动补偿的运动目标检测 [J]. 电子学报, 2017, 45(1): [5] Ge H Y, Sun J H, Ln N, et al. A movng target detecton approach to remove jtter by fusng wavelet transform and SIFT algorthm[j]. Research and Exploraton n Laboratory, 2016, 35(2): , 174. 葛鹤银, 孙建红, 林楠, 等. 融合小波变换及 SIFT 算法的去抖动运动目标检测 [J]. 实验室研究与探索, 2016, 35(2): , 174. [6] Tang J L, Zheng J F, L X Y, et al. Research on detectng algorthm of movng target n aeral vdeo[j]. Computer Scence, 2017, 44(S2): , 183. 唐佳林, 郑杰锋, 李熙莹, 等. 航拍视频中运动目标检测算法研究 [J]. 计算机科学, 2017, 44(S2): , 183. [7] Lu W, Zhao W J, L C, et al. Detectng small movng target based on the mproved ORB feature matchng[j]. Opto-Electronc Engneerng, 2015, 42(10): 刘威, 赵文杰, 李成, 等. 基于改进 ORB 特征匹配的运动小目标检测 [J]. 光电工程, 2015, 42(10): [8] Leutenegger S, Chl M, Segwart R Y. BRISK: bnary robust nvarant scalable keyponts[c]// Proceedngs of 2011 IEEE Internatonal Conference on Computer Vson, 2011: [9] He L Y, Lu J H, L G, et al. Fast mage regstraton approach based on mproved BRISK[J]. Infrared and Laser Engneerng, 2014, 43(8): 何林阳, 刘晶红, 李刚, 等. 改进 BRISK 特征的快速图像配准算法 [J]. 红外与激光工程, 2014, 43(8): [10] Krmnger E, Cobb J T, Príncpe J C. Onlne actve learnng for automatc target recognton[j]. IEEE Journal of Oceanc Engneerng, 2015, 40(3): [11] Zhang W Y, L Z, Wang Y J. UAV scene matchng algorthm based on CenSurE-star feature[j]. Chnese Journal of Scentfc Instrument, 2017, 38(2): 张闻宇, 李智, 王勇军. 基于 CenSurE-star 特征的无人机景象匹配算法 [J]. 仪器仪表学报, 2017, 38(2): [12] Cu S H, Xe Z, Wang G, et al. Feature matchng electronc mage stablzaton based on bnary robust nvarant scalable keyponts[j]. Optcs and Precson Engneerng, 2015, 23(9): 崔少辉, 谢征, 王刚, 等. 二进制鲁棒不变尺度特征匹配电子稳像 [J]. 光学精密工程, 2015, 23(9):

7 Movng object detecton under complex dynamc background Wang Smng, Han Lele * School of Automaton & Electrcal Engneerng, Lanzhou Jao Tong Unversty, Lanzhou, Gansu , Chna Movng object detecton result under complex dynamc background Overvew: Movng object detecton has been the focus of research n the feld of machne vson and ntellgent transportaton. Its purpose s to segment the movng objects from the sequence of vdeo mages so as to make the next step for target recognton, trackng and navgaton. However, under the complex dynamc background, many factors such as lght changes, background nterference, camera moton and so on, make the detecton very poor. At present, the commonly used feature pont detecton algorthms nclude SIFT, SURF and ORB algorthm, but they cannot meet the requrements of movng target detecton. BRISK algorthm has better rotaton nvarance, scale nvarance and better robustness. BRISK algorthm s the best one n the mage regstraton wth larger blur, but the real-tme performance of BRISK algorthm s worse than ORB algorthm. Amng at the real-tme and accuracy of movng object detecton algorthm, ths paper proposes a movng object detecton algorthm based on mproved BRISK feature matchng. Frstly, the vdeo frames are dvded nto blocks, the entropy of each sub-block s calculated. The sub-blocks are fltered by usng the mage entropy, so that sub-blocks whose local nformaton s too concentrated can be removed so as to avod the nfluence of excessve local feature ponts. Secondly, the AGAST algorthm s used to detect the feature ponts of the remanng sub-blocks and generates the correspondng feature descrptors. Then, the feature matchng s performed accordng to the k-nearest neghbor algorthm, and the feature pont pars are further purfed by the Eucldean dstance. So as to acheve the purpose of further mprovng the accuracy of the algorthm, and provde relable data for calculatng the next moton parameters. An mproved PROSAC method s used to extract the optmal feature ponts to estmate the background moton parameters, and the background moton compensaton s completed by combnng the sx-parameter affne model. Fnally, the frame dfference method and morphologcal process to extract the movng target, and the Otsu method s used to obtan the optmal threshold to acheve a more complete segmentaton of the movng target. In order to evaluate the detecton effect of the algorthm, three groups of vdeo mages are used to verfy the algorthm. The proposed algorthm removes 32.7% of the feature ponts and mproves the runnng tme of 1.1 s based on the orgnal BRISK algorthm. The detecton effcency s better than the prevous ORB algorthm to some extent, at the same tme mproves the matchng effcency to more than 75%, and enhances the ant-nose performance of the algorthm. The expermental results show that the proposed algorthm can mprove the real-tme performance and ensure the robustness of the proposed algorthm. Compared wth the prevous detecton algorthms, ths algorthm s more sutable for the detecton of movng objects n the complex dynamc context. Ctaton: Wang S M, Han L L. Movng object detecton under complex dynamc background[j]. Opto-Electronc Engneerng, 2018, 45(10): Supported by Natonal Nature Scence Foundaton of Chna ( ) * E-mal: hlldyx12@163.com

Vol.39 No. 8 August 2017 Hyeonwoo Noh [4] boundng box PASCALV VOC PASCAL VOC Ctyscapes bt 8 bt 1 14 bt

Vol.39 No. 8 August 2017 Hyeonwoo Noh [4] boundng box PASCALV VOC PASCAL VOC Ctyscapes bt 8 bt 1 14 bt 39 8 Vol.39 No.8 2017 8 Infrared Technology August 2017 1,2,3 1,3 1,3 1. 2000832. 100049 3. 200083 4 1 TP391.41 A 1001-8891(2017)08-0728-06 Infrared Scene Understandng Algorthm Based on Deep Convolutonal