Moving cast shadow detection of vehicle using combined color models

Transcription

1 Moving cast shadow detection of vehicle using comined color models Bangyu Sun 1 Shutao Li 1 1. College of Electrical Information Engineering, Hunan University, Changsha, 41008, China jinjin135@16.com; shutao_li@yahoo.com.cn Astract: Moving cast shadow detection of vehicle is important to vehicle detection tracking. In this paper, a novel moving cast shadow detection approach using comined color models is proposed. Firstly, the ratio of hue over intensity in HSI color model is used to detect the right oject pixels in foreground regions. Secondly, we employ the theory of photometric color invariants in c 1 c c 3 color model to distinguish the dark (similar to shadow) colorful oject pixels from shadow pixels. Finally, to improve the accuracy of shadow detection, post processing is used to correct failed shadow oject detection. Experimental results indicate that the proposed method can detect moving cast shadows accurately. Key Words: traffic monitoring system, shadow detection, HSI model, post processing 1 INRODUCION Moving ojects detection plays an important role in intelligent visual surveillance systems. However, shadows of moving ojects often cause serious errors in image analysis ecause of the misclassification of shadows moving ojects [1]. During the past years, many methods of moving cast shadow detection have een proposed [] [3] [4] [5]. enerally, shadow detection methods can e divided into two categories: shape-ased methods spectrum-ased methods. Shape-ased methods use the priori geometric information for the scenes, ojects, light-source location to solve the shadow detection prolem [3]. Hsieh et al. utilized lane features proposed a line-ased algorithm to separate shadows from the moving vehicles [4]. Yoneyama et al. estalished a two-dimensional joint-vehicle/shadow model to represent the ojects their cast shadows in order to separate the shadows from the ojects [6]. Because these methods depend on the geometrical relationship of the ojects in the scenes, they lose effectiveness when the geometrical relationship changes. In the existing shadow detection method, shadow spectral characteristics are more popular than geometric features. Spectrum-ased techniques employ spectral information of oject regions shadow regions to detect shadows [7] [8]. Compared with the shape-ased approaches, the information ased on spectrum-ased approaches is more explicit, ecause the spectral relationship etween oject regions shadow regions is affected only y illumination, rather than light source directions or oject shapes. In [9], rightness chromaticity distortions are defined normalized, so a pixel is separated as shaded ackground or shadow if it has similar chromaticity ut lower rightness than the same ackground pixel. In [], Cucchiara et al. proposed an improved shadow detection in moving oject detection with HSV color information. In [10], a classification method aout color edges applying photometric invariant features into shadow geometry edges, highlight edges material changes is proposed. he differences of scenes may result in the difference of the luminance. In addition, the vehicles also have different color, size speed. Most algorithms can not adapt to these changes well. In this paper, a novel moving cast shadows detection approach using comined color models is proposed. In the first step, the ratio of hue over intensity is employed to determine whether the pixel is a shadow pixel or not. In the second step, three aussian models in color model c 1 c c 3 are estalished to detect shadows. he c 1 c c 3 invariant color features can e adaptive to variale illumination conditions. In the third step, we get the rough result of shadow detection y synthesizing the aove two results. Finally, to improve the accuracy of shadow detection, two types of spatial analysis are designed to verify actual shadow pixels. In the process of shadow detection, we deal with the foreground region instead of the whole image, thus it reduces the computation time. he rest of this paper is organized as follows. Section riefly reviews two color models. Section 3 introduces the process of shadow detection. Experimental results discussions are provided in Section 4, the conclusions are drawn in Section 5. WO COLOR MODELS In this section, we will introduce the two color models employed in this paper..1 Color model HSI Because the intensity of shadow region is lower than that of oject region, the HSI color model can reflect this prolem etter than other models, such as RB, YUV. Here, we apply the ratio of the hue over the intensity. he following equation is aimed to transform RB into HSI color model /10/$ IEEE

2 1 1 1 I R V1 = (1) V 1 B S = V 1 + V () I 1 I V H = (3) 1 tan ( ), if V1 1 V1 H is undefined, othersise I 3 In the HSI color model, H I components denote the hue intensity components respectively.. Color model c 1 c c 3 A spectral property of shadows can e derived y considering photometric color invariants. Photometric color invariants are functions which descrie the color configuration of each image point discounted y shadows highlights. hese functions are demonstrated to e invariant to changes in viewing direction illumination condition. One of the typical photometric color invariants is color model c 1 c c 3, c 1 c c 3 is defined as follows [1]. c c c 1 3 R = arctan[ ] max(, B) = arctan[ ], (4) max( R, B) B = arctan[ ] max( R, ) where R,, B is the corresponding value of red, green, lue components of a pixel. 3 SHADOW DEECION In this section, we employ the two color models mentioned in Section to detect shadows. he process of shadow detection is descried as follows. Step 1: For the spectral characteristics of shadows, we detect shadows in HSI c 1 c c 3 color models, then get two shadow images. Step : etting the rough shadow image y synthesizing the aove two images using logical operation. Step 3: Post processing is used to o correct failed shadow oject detection in order to improve the accuracy of shadow detection. he lock diagram of the proposed shadow detection procedure is illustrated in Fig. 1. Input foreground images are passed through the system I i is shadow detection image after each step. I 4 Fig.1 Block diagram of the proposed shadow detection method 3.1 Rough detection As we all know, the intensity of oject region is usually larger that of shadow region, so we can find a threshold to separate shadows from foreground region. When detecting the shadows, we only deal with the foreground region segment the foreground region into one or more independent connected regions. By employing HSI model, H I components must e scaled to the range in [0, 1]. In this way, we can get the hue-equivalent image H e the intensity-equivalent image I e, respectively. he ratio r is defined as H e ( x, + 1 r( x, =, (5) I e ( x, + 1 where ( x, is coordinate of a pixel. H e ( x, I e ( x, represent the values H e I e at position ( x,, respectively. he term I e ( x, + 1 can avoid denominator y zero the ( H e ( x, + 1) /( I e ( x, + 1) ratio shall enhance the hue property of shadows with low luminance. i.e., pixels in shadowed regions will have higher values in the ( H e ( x, + 1) /( I e ( x, + 1) ratio than those pixels in oject regions. hen a threshold for each connected region separately can e found to detect shadows. A shadow map can e evaluated y 1, rxy (, ) > S1( x, =, (6) where is a threshold. If S 1 ( x, = 1 represents the shadow pixel at ( x,. In order to get the threshold, the following equation is exploited. = arg min( i= i= + 1 P( ( i μ ) + P( ( i μ ) ) (7)

3 where μ = ( ip( / W ), μ = ( ip( / W ). And 1 i = 0 1 = P i = i = = P( i = 1 W (, W, where P ( is the + proaility of the ratio value i in r map. According to the result, the method aove in HSI color model can only distinguish the large intensity pixels from the low intensity pixels. here are still many pixels having the similar low intensity to shadow pixels, e.g. the windshield of a vehicle the region of oject not eing irradiated, thus it may mistake the low intensity pixel of the oject for shadow pixel. In color model c 1 c c 3 this prolem can e solved well. As mentioned in section II-B, we define the ( P R, P ) as follows. PR = c1 c1 P = c c, (8) P = B c3 c3 where c 1, c, c 3 denote the values of c 1, c c 3 in the ackground image at the same position, respectively. In the coordinate systems ased on ( P R, P ), aussian models are exploited to represent the constant ( P R, P ). From the RB ratio aussian Shadow Model in [5], we can estalished three aussian models for P R, P, P B components to detect shadows. he models can e found y analyzing shadow samples taken from the shadow region in an image frame. u R, u, u B are the mean values σ R, σ, σ B are the stard deviation of P R, P, P B respectively. o cope with the variations, we employ the aussian distriution inside ± 1.5σ (88.6%) as a threshold. 1, if PR( x, ur < 1.5σ R P( x, u < 1.5σ S ( x, =, (9) PB( x, ub < 1.5σ B where ( x, is coordinate of a pixel. P R ( x,, P ( x,, P B ( x, are the input values of P R, P, P B at position ( x,. If S ( x, = 1 represents the shadow pixel at position ( x,. From the theory in HSI c 1 c c 3 models, the synthetic result can e represented as follows. 1, if S1( x, = 1 S ( x, = 1 S ( x, = (10) If S ( x, = 1 represents that the pixel at ( x, is classified as shadow pixel, otherwise, as oject pixel. 3. Post processing hrough the procedures of detecting shadows, two types of errors may occur: shadow detection failure oject detection failure. Shadow detection failure means that shadow pixels may e misclassified as moving ojects results from some shadow pixels having the similar informations to the oject pixels. his occurs especially at the edges of shadow regions. Oject detection failure is that some regions of ojects are misclassified as shadows. o increase the accuracy of shadow detection, a post-processing spatial analysis for shadow verification is employed. he spatial analysis is used to confirm the true ojects as well as the true shadows according to their geometric properties. he steps of post processing are descried as follows. Step 1- Correct shadow detection failure: In the process of detecting shadows, the true shadows sometimes reak into isolated shadow los. Usually, the sizes of these shadow los are smaller than the detected ojects in the sequence. hese small los are not considered as moving ojects. We should eliminate the small shadow los from the whole los of moving oject cidates. So the prolem can e easily corrected y morphological close operation. Step -Correct oject detection failure: If one part of the detected oject is misclassified as a shadow, most of the exterior pixels adjacent to the oundary of this region will e located inside the cidate foregrounds. iven this condition, oundary information can e used to confirm whether the shadow cidate is a true shadow or not. In this paper, each distinct cidate shadow region is determined y using a connected components laeling algorithm. hen, the oundary s exterior 8-neighoring pixels set E are evaluated y equation (11). E = MaskCmpt SE MaskCmpt, (11) where SE = MaskCmpt represents the component mask. Symol denotes morphological dilation. At last, the numer of pixels in the set E ( N E ) is counted, as well as the numer of the pixels detected as foreground in E ( N M ). If the in-equation N M > 0. 8N E holds, the component will e modified as foreground, otherwise, confirmed as shadow. 4 EXPERIMENAL RESULS his section presents the results of moving cast shadow detection that has een tested in different scenes. wo sequences are used in experiments. Sequence1 is downloaded at derives from a road crossing. Sequence comes from Jingzhu Highway. he proposed method is used to compare two well-known methods. Statistical Nonparametric (SNP) [11] approach Deterministic Nonmodel-Based (DNM) [] approach are chosen. 4.1 Metrics of performance evaluation In order to analyze the proposed method ojectively quantitatively, the shadow detection rate η shadow

4 discrimination rate ξ are widely used [8]. hey are defined as follows. Ps PF η =, ξ = (1) Ps + FN s PF + FN F In (1), the suscript symol S sts for shadow F for foreground, where P S P F respectively represent the numers of shadow pixels foreground pixels correctly recognized. FN S FN F respectively represent the numers of shadow pixels foreground pixels falsely recognized. 4. Comparison results he results of Sequence1 Sequence processed y three algorithms are demonstrated in Fig. Fig.3, white color corresponds to the detected shadow pixels. he DNM algorithm works in the HSV color model is the one with the most stale performance. Due to the limitation of this algorithm, it may not detect the oject regions which have the similar intensity to shadows. In Fig.- Fig.3-, sizale windshields of the vehicles are misclassified as shadows some other shadows are not extracted correctly. he SNP algorithm treats oject colors as a reflectance model from the Lamertian hypothesis. It employs the normalized distortion of rightness distortion of chrominance, computed from the difference etween the ackground color of a pixel its value in the current image, to verify whether a pixel is a shadow or not. he SNP algorithm is very effective in most of the scenes, ut with very variale performances. It achieves good detection performance in indoor scene, ut will e out of work in traffic system. In Fig.-c Fig.3-c, there are ovious misclassifications etween shadows ojects. In our proposed method, comined color models are employed to detect shadows. From the purpose in HSI c 1 c c 3 color models, they are largely complementary. Oviously, the color of vehicles in Sequence1 is different from that in Sequence, the algorithm we proposed aove can remove shadows efficiently. Meanwhile, the size of moving ojects in Sequence is different in Sequence1, the proposed algorithm can also detect the moving shadows exactly. he experimental results show that the proposed method provides more reliale shadow detection results than the two chosen methods. c) SNP d) Proposed method Figure Comparison results of Sequence1 processed y three methods a) Original image ) DNM c) SNP d) Proposed method Figure 3 Comparison results of Sequence processed y three methods We have presented a quantitative comparison of our proposed method with the two chosen methods, the results are shown in ale 1. he results show that the proposed method excels the classical methods. ale 1: Quantitative evaluation comparison of the proposed method Sequence1 Sequence η (%) ξ (%) η (%) ξ (%) DNM SNP Proposed a) Original image ) DNM 5 CONCLUSIONS In this paper, a novel method for moving cast shadow detection of vehicle is proposed. By analyzing the shadow information, we employ the comined color models. From the purpose of the methods in these two models, we can see that they are largely complementary. Experimental results show that the proposed method can detect moving cast shadows

5 accurately, it outperforms two well-known shadow detection approaches oviously. For the further research of shadow detection, edge information can e used as a feature of the shadow, while the shadow region is smoother than the oject region. In addition, some potential properties of shadows will e further investigated to make systems more accurate effective. [14] Martel-Brisson, N. Zaccarin. A. Moving Cast Shadow Detection from a aussian Mixture Shadow Model. Proc. CVPR, 005, Vol., pp [15] V. J. D. sai. A comparative study on shadow compensation of color aerial images in invariant color models. IEEE rans. eosci. Remote Sens., 44(6): , 006. ACKNOWLEDEMENS his paper is supported y the National Natural Science Foundation of China (No ), the Ph.D. Programs Foundation of Ministry of Education of China (No ), the Key Project of Chinese Ministry of Education (009-10), the Open Projects Program of National Laoratory of Pattern Recognition. REFERENCES [1] E. Salvador, A. Cavallaro,. Erahimi. Cast shadow segmentation using invariant color features. Computer Vision Image Understing, (95): 38-59, 004. [] Rita. Cucchiara, Costantino. rana, M. Piccardi, Andrea Prati Stefano Sirotti. Improving Shadow Suppression in Moving Oject Detection with HSV Color Information. Proc. ICIS, Oakl (CA), USA, 001, pp: [3] J-W. Hsieh, W-F. Hu, C-J. Chang, Y-S. Chen. Shadow elimination for effective moving oject detection y aussian shadow modeling. Image Vision Computing, 1(): , 003. [4] J-W. Hsieh, S-H. Yu, Y-S. Chen, W-F. Hu. A shadow elimination method for vehicle analysis. Proc. ICPR, Camridge, UK, 004, Vol. 4, pp [5] K- Song, J-C ai. Image-ased traffic monitoring with shadow suppression. Proceedings of the IEEE, 95(): , 007. [6] A. Yoneyama, C. H. Yeh, C. C. J. Kuo. Moving cast shadow elimination for roust vehicle extraction ased on -D joint vehicle/shadow models. Proc. ICAVSBS, Miami, FL, 003, pp. 1-. [7] S. Nadimi B. Bhanu. Physical models for moving shadow oject detection in video. IEEE rans. Pattern Anal. Mach. Intell., 6(8): , 004. [8] A. Prati, I. Mikic, M. M. rivedi, R. Cucchiara. Detecting moving shadows: Algorithms evaluation. IEEE rans. Pattern Anal. Mach. Intell., 5(7): , 003. [9] Pankaj Kumar, Kuntal Sengupta Adrian Lee. A Comparative Study of Different Color Spaces for Foreground Shadow Detection for raffic Monitoring System. Proc. ICIS, Singapore, 00, pp [10]. evers, H. Stokman. Classifying color edges in video into shadow-geometry, highlight, or material transitions. IEEE rans. Multimedia, 5(): 37-43, 003. [11] I. Haritaoglu, D. Harwood, L. S. Davis. W4: Real-time surveillance of people their activities. IEEE rans. Pattern Anal. Mach. Intell., (8): , 000. [1]. evers, A. W. M. Color-ased oject recognition. Pattern Recognition, 3(3): , [13] R. Cucchiara, C. rana, M. Piccardi, A. Prati. Detecting moving ojects, ghosts shadows in video streams. IEEE rans. Pattern Anal. Mach. Intell, 5(10): , 003.