TMCnet News
Motion Object Detection Research Based on Texture Feature [Sensors & Transducers (Canada)](Sensors & Transducers (Canada) Via Acquire Media NewsEdge) Abstract: The result of motion object detection will be affected by illumination changes. This paper proposes a novel background model based on local texture feature, which is described by local binary pattern (LBP), and LBP is gray-insensitive. Every coming frame and background is converted into LBP at first. Foreground probabilistic frame is generated by operation of exclusive or current LBP and background's LBP. It represents the probability of a pixel belonging to foreground. Then background can be updated adaptively and fast by foreground probabilistic frame. Experimental results demonstrate the effectiveness of the proposed approach in dealing with sudden illumination changes. It can relative precisely obtain moving objects and update background model quickly. Copyright ©2014 IFSA Publishing, S. L. Keywords: Motion object detection, Illumination change, Local binary pattem, Background model. (ProQuest: ... denotes formulae omitted.) 1. Introduction Moving object detection is an important research topic in the field of computer vision, it has been widely used in many areas, such as intelligent video surveillance, human-computer interaction, visual navigation. It is in the bottom of the entire video processing system, and it is the basis of a variety of subsequent advanced processing, such as target tracking, target classification, behavior recognition, scene understanding and so on. Current detection methods of the surface target commonly have light flow method, inter-frame difference, background modeling method in motion, there is computing capacity in optical flow method [1], and it is more sensitive to noise, there are high hardware requirements, timeliness and usefulness are poor, illumination changes can't be handled, especially when there is the light mutation in scene, its effect is poor. Frame difference method [2] is the difference in time-domain, which is subtraction between two to three frame images with different moments, this method is very effective in dynamic scenarios, illumination changes can be handled slowly to a certain extent, when there is a sudden change in the light, it can't effectively detect the target. Background modeling method is that moving targets are detected by a differential between current image and background image [3-6], the calculation speed is fast, a more complete characterization data can often be provided, but the background image with modeling is more sensitive to light changes, there will be a lot of pseudo-target point. As Cheng et al. [7] proposed a light-sensitive background model (ISBS), the algorithm illumination changes slowly with good results, but it can't handle the situation of the light mutation. Hu et al. [8] built background modeling by the image histogram gradient, there is certain robustness in illumination change, Toth et al. [9] introduced homomorphic filtering to reduce the influence of illumination change, Aach et al. [10] build a simple model by Bayesian theory, the sensitivity is reduced to light illumination change. Local Binary Pattem (LBP) was first proposed by Ojala et al. [11], the basic idea is that the gray value of the center pixel is set into threshold, compared the pixel points with its circular neighborhood, binary code is used to describe local texture, it is less susceptible to gray linear change of the whole image, when the gray of the image has uniform changes, the LBP texture feature is constant, and therefore LBP texture features has a gray-invariant. There is a simple calculation, classification capabilities are strong, the described texture feature extraction has a significant effect. This paper mainly researches that background modeling method can't adapt to sudden changes in illumination, LBP texture features are insensitivity to light, the adaptive background update strategy is designed, in conditions of light sudden changes, LBP texture features have robustness, for maintaining background, and the prospect location and outline can be well positioned, a moving target detection algorithm is presented, which is robust to illumination changes. 2. The Basic Framework LBP texture features are with gray invariance, a robust background illumination model is built. Firstly, the prospect probability plots are gotten by XOR of the current frame and the background frame LBP texture features, the current context is calculated for re-use of prospect probability plots, the initial target is obtained by a further difference, the moving target is gotten the initial target morphological processing, the algorithm processes is shown in Fig. 1. 3. Background Modeling 3.1. LBP Texture Features LBP describes the image texture through the joint distribution of P pixels on the annular neighborhood of each pixel and its radius R. Wherein gc represents the gradation value of the local neighborhood center, gp(jj =0,1,- -,P-l) corresponds to a gray value of the P dividing points P on the ring with the radius R, which is different (P, R) in combination, LBP operator is not the same, Fig. 2 is three kinds of different LBP operator. In order to achieve gray invariance of this texture operator, the gray values of the P equal points gp (p = 0,1, * * *, P -1 ) with annular neighborhood minus the center point gray value gc, joint distribution T is converted to ... (1) Assume that gc and gp are each independent, Formula (1) is approximately decomposed into ... (2) In the Formula (2), t(gc) describe the gray distribution of the entire image, there is no impact on local texture distribution of the image, so the image texture features can be described by the joint distribution of the differential, i.e. ... (3) When the light additive changes of the image occur, the center pixel gray value and its ring neighboring pixel gray values generally do not change the relative size, i.e., gp-gc is not affected by light additive changes, therefore, sign function of difference between the central pixel and its neighborhood pixels can be used in place of specific numerical values, which can describe the image texture. ... (4) ... (4) where s is the sign function. ... (5) The results, which are obtained joint distribution T, are sorted by specific order on annular neighborhood pixels, which is constituting a 0/1 sequence, by the counterclockwise sort in the paper, the right neighboring pixels of the center pixelas is as a starting point to calculate pixel, binomial factor 2P is given for every s(gp - gc ) , the local spatial texture pixel may be represented as a unique decimal number, the decimal number is called LBPp, r number, which is reasons that the texture operator is referred to as local binary mode (Local Binary Pattem), LBPp, r number can be calculated by Formula (6). ... (6) Specific LBP texture feature calculation process is shown in Fig. 3, the left side in Fig. 3 is the template of the threshold value, so that the neighboring pixels and the central pixel is compared, it is greater than 0, 1 is set, it is less than 0, it is set to 0, which are obtained in Fig. 3, by counterclockwise, 0/1 sequence constructor is sequence (10100101), the final corresponding decimal number (165) is calculated, LBP texture feature of the pixel value is 165, LBP eigenvalue is found for each pixel in the image, LBP texture characteristic graph of an image is obtained, which are shown in Fig. 4. Because LBP texture features at the edge are affected little by the neighborhood, these pixels of the image edges retain the original pixel gray value. 3.2. Background Update Policy When illumination changes occur in moving target detection process, the background is constantly changing, so the target detection is good or bad, which often depends on whether the background update is accurate and timely, an accurate background model is built, which can effectively suppress illumination changes to bring on the impact in the target detection, the adaptive background updating strategy is proposed in this paper, when a new frame arrives, the background is updated by using Equation (7). ... (7) where Bt (x, y) represents background of the current frame, Bt_x{x,y) represents the previous frame background, It{x,y) represents the current frame, P,(x,y) represents the probability of the current foreground pixels, £0(x,y) is initialized with the first frame. ... (8) where L5P(/r(x,y)) is the characteristic graph of the current frame LBP texture, LBP{Bt_x (x, y)) represents an LBP background texture feature map of previous frame, ones(x) represents the number of 1 in the binary representation of x, P is the number of pixels on the neighborhood. According to LBP gray invariance, when certain pixel It{x,y) is a background, after XOR between 7,(x,y) and Bt_x(x,y), the number of 1 is less, Pt(x,y) is small, whereas the number of 1 is large, 7^(x,y) is large. According to Formula (7), when a pixel is background, P,(x,y) is small, Bt (x,y) tends to 7<(x,y) , the change of light can effectively suppressed, and the background is updated in the true background for a short time. When a pixel is foreground, Pt{x,y) is big, Bt(x,y) tends Bt_j (x, y ), the target can effectively detected. 4. Moving Target Detection After generating the background model, to calculate the difference between background and the current frame. ... (9) When À,(x,y) is greater than the threshold value Tt{x,y), the pixel is the prospect otherwise it is considered to be the background, T0{x,y) is initialized to 20, Tt(x,y) uses the update strategy in the literature [7]. ... (10) Since this algorithm is based on texture features, when the background textures and the foreground texture are similar, the detected target will be many voids. For this case, after obtaining prospects, but also the prospects are needed for further processing, voids are eliminated. In this paper, morphological operations are used, the first operation is to expand target, and then the operation is etching target, voids can be eliminated to some extent. 5. Experimental Results and Analysis To verify the effectiveness of the proposed algorithm, the three video sequences are selected to validate the algorithm, they are intelligentjroom (IR), LightSwitch (LS) and TimeOfDay (TD). In Fig. 5, the first line is the video sequence IR, since there is the mirror-scene, when the target approaches the mirror, light localized mutations are caused. The second line is the video sequence LS, in a scene, due to light with cover, and the lamp suddenly is opened and closed, the illumination mutations are caused. The third line is the video sequence TD, a fixed value is artificially increased for the gray value of each pixel of a frame, the overall rate of mutation is caused by illumination, the gradation increments are selected into 7 in the herein experiments. The first column is the first image of video sequence, the second column is the illumination change video frame, 92th frame of IR section is selected, LS' 802th frame is selected, TD' 4737th Sensors & Transducers, Vol. 175, Issue 7, July 2014, pp. 279-283 frame is selected, the third column is the test results of GMM algorithm [4], the fourth column is the detection result of ISBS algorithm [7], the detection result of the algorithm in this paper is in the fifth column, the sixth column is the true background. The main hardware environment for experiments in this article: Intel Core (TM) 2 Duo CPU (2.00 GHz), DDR-2.00G RAM, NVIDA GeForce 405 graphics card, a video sequence size is 160^120, processing speed is 8.9 / s. As can be seen from the above results, the proposed algorithm is robust to illumination. In the first experiment, the partial illumination mutation, which mirror caused, has been effectively suppressed, while the enhanced light in the reflector is false detected into the prospects in GMM and ISBS algorithm. In the second experiment, because the target has a shielding effect of the light, which results in illumination mutation, a large area of the target are detected by GMM and ISBS, while the use of this algorithm has a gray-invariant with LBP texture features, the target can be detected better. In the third experiment, the mutational effects of light are artificially increased, the background can be not updated in GMM and ISBS algorithms, the not target scene is detected to become a target, which results in false detection, and the detection of our algorithm have better results. You can also learn from the experiment, by our algorithm, the mutations, which are caused by background illumination, are updated soon, because the background updating strategy in this article is to allow every updated background be as close to the real background. 6. Conclusions and Outlook This paper studies the moving object detection under illumination changes, the gray invariance of LBP texture features is used, a new background updating algorithm is proposed, the LBP texture features XOR between the current frame and the background frame is firstly used to get prospects probability plots, the current context is calculated by re-use of prospect probability plots, the prospects are gotten by further differential. By comparison with other algorithms experiments, the robustness of the algorithm is verified for illumination, the algorithm has a very good performance for partial illumination mutation and the overall lighting mutation. Dealing reflected light and shadows with complex background will be the direction of future research. References [1]. Alvarze L., Weichert J., Reliable Estimation of Dense Optical Flow Fields with Large Displacements, Computer Vision, 39, 1, 2000, pp. 41-56. [2]. Lipton A., Fujiyoshi H., Patil R., Moving target classification and tracking from real-time video, in Proceedings of the IEEE Workshop on Application of Computer Vision, 1998, pp. 8-14. [3]. Stauffer C., Grimson W. E. L., Learning Patterns of Activity Using Real-Time Tracking, IEEE Trans, on Pattern Analysis and Machine Intelligence, 22, 8, 2000, pp. 747-757. [4]. Kaew TraKulPong P., Bowden R., An Improved Adaptive Background Mixture Model for Realtime Tracking with Shadow Detection, in Proceedings of the 2nd European Workshop on Advanced Video Based Surveillance Systems, 2001. [5]. Maha M., Howida A., Ashraf S., A new Technique for Background Modeling and Subtraction for Motion Detection in Real-Time Videos, in Proceedings of the IEEE Int'l Conf. on Image Processing, 2010, pp. 3453-3456. [6]. Ming Y. Y., Yang B., Men A. D., Guo Y. L., Background Subtraction under Multiple Varying Illuminations in Different Background Luminance, in Proceedings of the IEEE Int'l Conf. on Computer Modeling and Simulation, 2010, pp. 189-193. [7]. Fan-Chieh Cheng, Shih-Chia Huang, Shanq-Jang Ruan, Illumination-Sensitive Background Modeling Approach for Accurate Moving Object Detection, IEEE Trans, on Broadcasting, 2011, pp. 1-8. [8]. Lei Hu, Weibin Liu, Bo Li, Weiwei Xing, Robust Motion Detection using histogram of Oriented Gradients for Illumination Variations, in Proceedings of the 2nd International Conference on Industrial Mechatronics and Automation, 2010. [9]. Toth D., Aach T., Metzler V., Illumination-invariant change detection, IEEE Southwest Symposium on Image Analysis and Interpretation, 2000, pp. 3-7. [10]. Aach T., Dumbgen L., Mester R., et al, Bayesian illumination-invariant motion detection, in Proceedings of the International Conference on Image Processing (ICIP01), 3,2001, pp. 640-643. [11]. Ojala T., Pietikäinen M., Mäenpää T., Multiresolution Gray-Scale and Rotation Invariant Texture Classification with Local Binary Pattern, IEEE Trans, on PAMI, 24, 7,2002, pp. 971-987. 'Shanshan Peng 1 School of Information Science and Engineering, Hunan International Economics University, Changsha, 410205, China 1 Tel:+86-0731-88140728 1 E-mail: [email protected] Received: 15 July 2014 /Accepted: 28 July 2014 /Published: 31 July 2014 (c) 2014 IFSA Publishing, S.L. |