Ship Target Detection Algorithm for Maritime Surveillance Video Based on Gaussian Mixture Model

Transcription

1 Journal of Physics: Conference Series PAPER OPEN ACCESS Ship Targe Deecion Algorihm for Mariime Surveillance Video Based on Gaussian Mixure Model To cie his aricle: Zuohuan Chen e al 08 J. Phys.: Conf. Ser View he aricle online for updaes and enhancemens. This conen was downloaded from IP address on 0/0/09 a :55

2 IOP Conf. Series: Journal of Physics: Conf. Series (08) 00 doi :0.088/ /098//00 Ship Targe Deecion Algorihm for Mariime Surveillance Video Based on Gaussian Mixure Model Zuohuan Chen, Jiayuan Yang, Zhining Chen and Zhen Kang Navigaion College, Dalian Mariime Universiy, Dalian, 606, China Srais College of Engineering, Fujian Universiy of Technology, Fuzhou, 35008, China Absrac. The paper presens a vessel arge deecion algorihm o achieve he mariime visual surveillance, which aims o reduce he influence of cluer ha exiss in he bacground and improve he reliabiliy of ship arge deecion. In he proposed deecor, he main seps including bacground modeling, model raining and updaing and he segmenaion of foreground, are all based on Guassian Mixure Model (GMM). By exploiing he characerisics of GMM, we simply deermine wheher new pixels, in he video, belong o he foreground. Having modeled surveillance region, we roll ou he moving ship deecion using bacground subracion, segmening he ship arge by he coninuiy of he adjacen frames. The arge precision rae of he algorihm is 97.9% and he false alarm probabiliy is.83% in he experimens. Comparing wih oher algorihms, he resuls show ha his algorihm can no only improve arge precision rae, bu also reduce false alarm probabiliy, and grealy overcome he influence of large amoun of cluer on he deecion of moving ship objecs in video bacground, effecively resraining he influence of he noise from he dynamic scenario ransformaion.. Inroducion The curren mariime surveillance sysems are based on space-born Synheic Aperure Rader (SAR), High Frequency Surface Wave Rader (HFSWR), regular ship-based radars, and air-/space-borne opical sensors []-[3]. The SAR equipmen can operae coninuously under all-weaher condiions a he expense of limied image. However, mariime video surveillance, an imporan par of he vessel raffic service, can provide more inuiive image includes moving ship informaion and maes up he deficiency of AIS (Auomaic idenificaion Sysem) and radar, which has become an effecive and feasible mode for mariime auhoriies o manage ship raffic [4][5]. Ship deecion and coninuous racing are he basis of many follow-up sudies, such as video moion analysis and behavior undersanding, which has resuled in he visual mariime surveillance become a research hospo of inelligence raffic. A presen, he curren mariime surveillance ools are mainly insalled he channel, cruise ships, bridges, near he waer-gae o acquire surveillance video image which can be used for ship arge deecion. Comparing wih oher monioring sysems, he mariime video surveillance has he following characerisics: he mariime images conain horizon consiss of sea surface and sy; Bacground cluer: mariime video surveillance includes a lo of shaing ripples and ligh spos; Camera shaing resuls in a non-linear change in he sae of he arge; Illuminaion variaions: differen seasons, ime periods, illuminaion inensiy and direcion affec he deecion of ship arges direcly. All hese facors mae i difficul o ship arge deecion. Conen from his wor may be used under he erms of he Creaive Commons Aribuion 3.0 licence. Any furher disribuion of his wor mus mainain aribuion o he auhor(s) and he ile of he wor, journal ciaion and DOI. Published under licence by Ld

3 IOP Conf. Series: Journal of Physics: Conf. Series (08) 00 doi :0.088/ /098//00 Some scholars have proposed deecion echniques for ship based on mariime video surveillance images. These mehods mainly include AFDM (adjacen frame difference mehod) [6], opical flow mehod [7], bacground modeling and subracion [8]. While AFDM has a srong adapabiliy o he change of illuminaion, scale and oher esing environmen, i is easy o deec holes in he ship arges deecion when dealing wih he exure of he arge pixel of he moving ship and he change of he color gradaion. Opical flow mehod can independenly deec he objecs of he moving ship, don' need o now in advance video ship arges moion scene, bu his mehod needs a large number of calculaions and has poor ani-noise performance, which is difficul o mee he requiremens of realime deecion. Bacground subracion mehod can effecively obain he complee region of he ship arges. The ey of his is o obain he bacground image of video surveillance scene and mainenance updaes. Bacground modeling plays a significan role in ship arge deecion by using he bacground subracion. N Friedman esablished a single Gaussian disribuion model [9], believing ha each pixel value in he video image approximaely obeyed Gaussian disribuion and by using he obained pixel value o esimae Gaussian disribuion parameers. The new pixels, in he video, which did no mach he Gaussian disribuions were regarded as bacground, oherwise he foreground. In he video scene of he arge deecion of he acual moving ship, he bacground exiss he waer ripples wih dynamic shaing, and his single-gaussian disribuion model has no significan effec on dynamic scenes deecion. C Sauffer proposed he GMM (Gaussian Mixure Model) o replace he single Gaussian disribuion o esimae he bacground [0], and successfully solved he approximae periodic moion of he bacground, such as he shaing leaves and corrugaion, and could also adap o he change of illuminaion. W C Hu consruced he bacground image using he median of he firs n frames of images, and proposed he corresponding bacground updaing mehod o realize he auomaic deecion of illegal arge inrusion in mariculure zone []. A Borghgraef and W Y Wang preinroduced arge informaion o be deeced [][3]. Based on he pixel subracion of he bacground o obain he ship arge image, and hen o updae and assessmen. Running Mean Updaing Bacground (RMUB) was used o obain he foreground image by consrucing he bacground model [4], reducing he "empy" phenomenon in moving arge deecion, bu i was sensiive o illuminaion change, while iniializing he updae bacground phase was slow, bacground learning was lagged behind, causing he foreground arge be deeced o bacground. In view of he above analysis, GMM can adap o he changing of complex bacground. Therefore, GMM is used o deec he arge of moving ship. Finally, he reliabiliy of he algorihm is verified by experimens.. Moving ship arge deecion The mariime surveillance video images are colleced by he infrared hermal imager or infrared high definiion camera insalled on he ship, gae, bridge e al. I usually includes four pars: sy, ship, sea surface and fixed objec, among which he sy, sea and fixed objecs are mared as he bacground, while ship is mared as he foreground. In his paper, GMM is used o deec he arge of moving ship. The deecion processes include bacground modeling, model raining and updaing and he segmenaion of foreground... Bacground modeling GMM can approximae he probabiliy densiy disribuion of arbirary shape, fiing he bacground disribuion of he real scene, and i is widely used o moving arge deecion [5]. The change of pixel value x capured in he image of ime is considered as a sochasic value. K Gaussian funcions are used o fi he gray disribuion of he pixel value x. The probabiliy of x is wrien as: K P( X ) = ωη ( X, µ, ) () =

4 IOP Conf. Series: Journal of Physics: Conf. Series (08) 00 doi :0.088/ /098//00 Where: K is he oal number of Guassian disribuions, wih he ranging from 3 o 5; ω K ( 0 ω, ω = = )is he normal weigh of componen a ime ; µ and are he mean and covariance marix a ime. The Guassian probabiliy densiy funcion η is defined as: T ( X µ ) ( X µ ) η = π exp () (( ) ).. Model raining and updaing Model maching is o compare he pixel poin I( xy, ) observed in he curren frame image wih he esablished GMM. If he equaion (3) is saisfied, i is considered ha he pixel poin I( xy, ) of he curren frame maches he esablished GMM, oherwise, he pixel does no mach. I xy, µ.5σ (3) ( ) Where: σ is he variance of componen a - ime. Model maching mus updae model parameers in real ime. Updaing mehod sees he equaion (4). + µ = ( -α ) µ + σx + σ = ( -ρ) σ + ρ ( µ -x ) (4) + ω = ( r) ω + rm Where: α is updae parameer, according o [3], he value is 0.0; ρ αω; r is updaed parameer for weigh; M is a value ha model mached is, oherwise 0. Keeping 3 Gaussian componens arrange from big o small. When he model does no mach, he curren frame pixel appears new disribuion, and he mean µ of he curren frame is used; besides, iniializing a larger variance and a smaller weigh. The new Gaussian disribuion model replaces he las Gaussian disribuion in he original sequence, and he remaining wo Gaussian disribuions eep he same mean and variance. Updaing he weigh according o equaion (5). ω + = r ω (5) ( ).3. The segmenaion of foreground In he GMM model parameer learning phase, he Gaussian disribuion wih a large weigh describes a large probabiliy of bacground pixel value, oherwise a small probabiliy of foreground pixel value. The Gaussian disribuion is sored by he value of ω σ from mid-size o small. If i, i,..., in,..., i m means ha each Gaussian disribuion are arranged according o he value of ω σ from big o small a ime and he firs n Gaussian disribuions are saisfied wih equaion (6), hen firs n Gaussian disribuions are he bacground disribuion and he res are he foreground disribuions. in ω T (6) = Where: T general value is 0.7~0.8, here is The experimenal resuls and analysis In order o evaluae he deecion performance of GMM, he arge precision rae (TPR) and false alarm probabiliy (FPR) are seleced as he evaluaion crierion [3]. In view of differen complex sea condiions, in his paper, a large number of experimenal ess are carried ou o compare GMM wih convenional deecion mehod: AFDM and RMUB, o verify he reliabiliy of GMM in he deecion of moving ship arges. All experimens are using MATLAB programming language, he compuer configuraion: Windows0, Inel(R).8GHz, memory 8GB of memory. The resoluion of he image:

5 IOP Conf. Series: Journal of Physics: Conf. Series (08) 00 doi :0.088/ /098//00 In he experimens, a oal of 037 images of wo scenarios are seleced as experimenal es samples. Scenario one is a surveillance video wih a oal of 07 frames, which indicaes he deecion bacground wih he change of illuminaion and dynamic swaying waer ripples during he day, anoher one conains boh dynamic swaying waer ripples and fixed objecs wih a oal of 830 frames. In order o deeply analyze he segmenaion rules of he arge pixels and he bacground pixels in surveillance video, his paper presens a mehod of pixel value analysis for a specific pixel poin I( xy, ). Using X o represen RGB color componen a ime, X ( x R, x G, x B ) =. The change of surveillance video image means he change of hree RGB color componen values, which means RGB values a he same locaion are replaced by new values. This paper chooses wo differen observaion poins in scenario, I ( 35,398 ) and I ( 69,45). Table indicaes he characerisics of variaion a I and I. In he scenario, here are no moving ship arge hough he poin I, and only one passes hrough poin I. The RGB componen changes and RGB 3D disribuion of poin I and I.are shown in figure and figure respecively. As is shown able, he mean, variance and sandard deviaion of he sample poin are relaively small, when here is no moving ship arge hough he poin I. The mean value increased suddenly, when here is moving ship arge passes by he sample poin, while he variance and sandard deviaion are larger han hose wihou moving arges. Table. Mean, variance and sandard deviaion of pixel sample poins Scenario Poins Coordinaion Frames Targe sae Mean Variance Sandard deviaion Scenario I (35,398) -69 No Scenario I (69,45) -86 No Yes No () No moving ship arge in scenario The change of he pixel value of he observaion poin eeps sable when here is no moving ship arge in scenario, as is shown in figure (a), and he change rend of RGB 3D disribuion presens a ellipsoid shape, as is shown in figure (a). () A few number of moving ship arges RGB color componens are suddenly increasing and hen reurn o he original value when he moving arge appears in scenario, as is shown in figure (b). RGB color componens of arge poins are far away from he cener of he model, as is shown in figure (b) (a) Changes of RGB componen a I (b) Changes of RGB componen a I Figure. Changes of RGB componen a I and I in scenario 4

6 IOP Conf. Series: Journal of Physics: Conf. Series (08) 00 doi :0.088/ /098//00 (a) RGB 3D disribuion a I (b) RGB 3D disribuion a I Figure. RGB 3D disribuion a I and I in scenario Figure 3 shows some experimenal resuls in scenario, figure 3 (a, f) show he original he images of 0h and 98h in he mariime surveillance video, he second o fourh column represen ha he saliency maps of arge exraced by using AFDM, RMUB as well as GMM; see figure 3 (b, c, d, g, h, i). The fifh column indicaes he deecion resuls by using he GMM; see figure 3 (e, j). In he iniial deecion sage, he hree mehods achieve good deecion resuls. The waer ripples are deeced o foreground in he sea-surface; see figure 3 (b, c, d). Wih he deecion goes deep, AFDM and RMUB are difficul o avoid he bacground of waer ripples deec as foreground; see he figure 3 (b, c). However, he ship arge achieves beer deecion performance by using GMM in sea-surface, avoiding he complex boh he waer ripples and ligh spo are deeced o bacground; see figure 3 (i). (a) (b) (c) (d) (e) (f) (g) (h) (i) (j) Figure 3. Deecion resul under he change of illuminaion during he day. (a, f) Original images; (b, g) Saliency maps of AFDM; (c, h) Saliency maps of RMUB; (d, i) Saliency maps of GMM; (e, j) Deecion resuls of GMM Figure 4 shows some experimenal resuls in scenario. The firs column is he original image of frames 5 and 6 in he surveillance video. The second o he fourh column represens he saliency maps of moving ship exraced by using AFDM, RMUB and GMM under he environmen of dynamic swaying waer ripples and saionary objecs. The fifh column indicaes he resul of using GMM for he deecion of moving ship arge. In he iniial deecion sage, as shown in figure 4 (b, c, d), he deecion effec of GMM is similar o AFDM, bu he deecion resuls of GMM and AFDM are beer han he RMUB. In erms of reducing he influence of wave effec, GMM is similar o RMUB, boh of hem are obviously superior o ha of AFDM. Wih he deecion goes deeper, hough AFDM and RMUB can improve he deecion effec, here are massive loss of deecion arges, as shown in figure 4 (g, h). Alhough he coasal sea wave bacground is deeced as foreground by he GMM algorihm, he arge area of he moving ship is disinc and has a clear exure, as shown in figure 4 (i). The TPR and FAR of GMM algorihm is 94.58% and 4.36% respecively, which is superior o he oher wo mehods, as shown in able. 5

7 IOP Conf. Series: Journal of Physics: Conf. Series (08) 00 doi :0.088/ /098//00 (a) (b) (c) (d) (e) (f) (g) (h) (i) (j) Figure 4. Deecion resuls of hree inds of mehods under he environmen of dynamic swaying waer ripple and saionary objecs. (a, f) Original image; (b, g) Saliency maps of AFDM; (c, h) Saliency maps of RMUB; (d, i) Saliency maps of GMM; (e, j) Deecion resuls of GMM In order o verify he effeciveness of he GMM as a whole, using he AFDM, RMUB and GMM o deal wih he monior video in he hree scenarios separaely, which is conduced under he same simulaion environmen. In his paper, we calculae he TPR and FAR of differen mehods in differen scenarios respecively, he resuls in shown in able. Table. TPR and FAR of differen mehods in differen scenarios P TPR P FPR Sequence AFDM RMUB GMM AFDM RMUB GMM Scenario 87.40% 99.0% 00% 34.0% 67.0%.30% Scenario 77.87% 89.33% 94.58% 58.43% 3.80% 4.36% As shown in able, GMM is beer han AFDM and RMUB. From he saisical resuls, The TPR of GMM is 97.9%, higher han he oher wo mehods. The FPR of GMM is.83%, lower han he oher wo mehods. The reason why here is he FPR is ha he ship size is smaller and also here is some significan disurbance ha caused by he impac of he sea wave. In complex sea condiions, when bacground has fixed objec mars and shaing waer ripples, he performance of GMM algorihm is beer han ha of he oher wo mehods. 4. Conclusion In his paper, he moving arge is deeced by ship surveillance video under hree scenarios. The deecion resul shows ha he average deecion rae of he hree mehods, AFDM, RMUB and GMM, are similar. GMM is slighly higher han AFDM and RMUB. However, he FPR of he hree deecion mehods is quie differen. The FAR of GMM is.83%. The FAR of AFDM and RMUB is 46.3% and 49.45% respecively. The resuls show ha GMM has beer deecion resuls compared wih radiional deecion mehod, such as AFDM and RMUB, and has srong ani-inerference abiliy. GMM can achieve good resuls regardless of wheher he bacground is relaively fixed, or he noise inerference caused by ligh, or he ripple of a large amoun of shaing. The nex sep is o sudy he influence of weaher condiion, bloced moving arges under oher condiions, such as poor visibiliy, ec., o design and develop an easier and more pracical bacground model o reduce he influence of dynamic scenario change on he deecion of moving ship. Acnowledgemen Liaoning naural science foundaion of China (06084; 06008); Dalian mariime universiy eaching reform projec funding of China (05Y0) 6

8 IOP Conf. Series: Journal of Physics: Conf. Series (08) 00 doi :0.088/ /098//00 References [] Tello M, López-Marínez C, Mallorqui J J. A novel algorihm for ship deecion in SAR imagery based on he wavele ransform[j]. IEEE Geoscience and remoe sensing leers, 005, (): [] Zhu C, Zhou H, Wang R, e al. A novel hierarchical mehod of ship deecion from spaceborne opical image based on shape and exure feaures[j]. IEEE Transacions on Geoscience and Remoe Sensing, 00, 48(9): [3] Zhang Y, Li Q Z, Zang F N. Ship deecion for visual mariime surveillance from non-saionary plaforms[j]. Ocean Engineering, 07, 4: [4] Sun W, Veraschisch L, Sahr J D. Developmen design and demonsraion of very high-speed muli-anenna digial receiver[j]. IET Radar, Sonar & Navigaion, 08, (5): [5] Sun W. High Speed Passive Radar Receiver wih Applicaion o Digial Television Signals[D]., 07. [6] LI Cong-sheng, BAI Jun, ZHOU Guang-lu. Vehicle deecion of ciy road a nigh base on frame difference[j]. Manufacuring auomaion, 0, 33(4):-4. [7] ZHANG Shui-fa, ZHANG Wen-sheng, DING Huang. Bacground modeling and objec deecing based on opical flow velociy field[j]. Journal of image and graphics, 0, 6(): [8] Sun Shou-qun, Liu Kang-ya, Liu Shou-yan, e al. Moving arge deecion in complex environmen of railway saion[j]. Journal of Traffic and Transporaion Engineering, 03, 3(3): 3-0. [9] Friedman N, Russell S. Image segmenaion in video sequences: A probabilisic approach[c]//proceedings of he Thireenh conference on Uncerainy in arificial inelligence. Morgan Kaufmann Publishers Inc., 997: [0] Sauffer C, Grimson W E L. Adapive Bacground Mixure Models for Real-Time Tracing[J]. Proc Cvpr, 998, :46. [] HU W C, Yang C Y, Huang D Y. Robus real-ime ship deecion and racing for visual surveillance of cage aquaculure[j]. Journal of Visual Communicaion & Image Represenaion, 0, (6): [] Borghgraef A, Barnich O, Lapierre F, e al. An Evaluaion of Pixel-Based Mehods for he Deecion of Floaing Objecs on he Sea Surface[J]. Eurasip Journal on Advances in Signal Processing, 00, 00():5. [3] WANG Wei-ye. Moving Objec Deecion and Behavior Recogniion in Inelligen Visual Surveillance[D]. Xian: Xidian Universiy, 03. [4] DING Wei. Research And Realizaion of Vehicle Deecion Algorihm in Inelligen Transporaion[D]. Nanjing: Nanjing Universiy of Poss Telecommunicaions, 0. [5] REN Ke-Qiang, ZHANG Pang-hua, XIE Bin. Adapive learning algorihm for moving arge deecion based on Gaussian mixure model[j]. Compuer Engineering and Design, 04, 35(3):