A NOVEL MOTION DETECTION METHOD USING BACKGROUND SUBTRACTION MODIFYING TEMPORAL AVERAGING METHOD

Transcription

1 International Journal of Computer Engineering and Applications, Volume XI, Issue IV, April 17, ISSN A NOVEL MOTION DETECTION METHOD USING BACKGROUND SUBTRACTION MODIFYING TEMPORAL AVERAGING METHOD P. Poornima 1, G. Rithin Surya SaiNadh 2 1 Assistant Professor, Department of Computer Science & Engineering, Mahatma Gandhi Institute of Technology, Hyderabad, India 2 Department of Computer Science & Engineering, Mahatma Gandhi Institute of Technology, Hyderabad, India ABSTRACT: The detection of moving objects in an image sequence is very important low-level task for many computer applications now-a-days, such as video surveillance and traffic monitoring. Many of these methods use background subtraction techniques to separate the foreground objects from the background. A mere approach in background subtraction techniques named as Temporal averaging is modified by overcoming some of the drawbacks for detection of moving foreground objects from a nonstationary background. The traditional temporal averaging method works perfectly in a stationary background and fails for small object detection and nonstationary background. Two types of variants are shown. One is by adaptive threshold and other is by adaptive updating speed by pixel difference. The proposed method is more reliable, less complex and not very expensive to implement. The comparison of results of traditional temporal averaging and proposed method are clearly shown with F-measure analysis. Keywords: Background Subtraction, Temporal averaging, Motion detection [1] INTRODUCTION In recent years, Motion detection is one of the challenges of image processing. But later, there is a rapid development in the field of digital image processing which lead to the improvement of techniques of motion detection. There are many challenges for motion detection. Some of them are Traffic monitoring, video surveillance system in heavily crowded places like shopping malls and banks. The existing motion detection techniques require a lot of computational time and a bit expensive to implement. So, there is a need for versatile and mere approach with a less cost of implementation. There are many methods existing to detect the moving objects like Kernel Density Estimation, Mixture of Gaussians method, Eigen backgrounds method and Background subtraction method. As the background subtraction method is one of the most popular methods, simple to implement, P. Poornima and G. Rithin Surya SaiNadh 35

2 A NOVEL MOTION DETECTION METHOD USING BACKGROUND SUBTRACTION MODIFYING TEMPORAL AVERAGING METHOD effective detection and not so expensive compared to the other methods. We used background subtraction method which is also known as foreground detection method. The cameras are stationary, the moving objects can be detected by the comparison of the new frame obtained with the background and this process is called background subtraction and the scene representation is called the background model. The background subtraction method uses the difference between the current pixel and background image to detect the motion of objects. In real-time static cameras, the background subtraction method is mainly used to detect moving objects in a video sequence. The captured video is processed by a MATLAB program to detect the motion. Figure.1: Original Image Figure.2: Its motion image The existing Temporal averaging method is good at detecting the moving objects in a stationary background. To increase the efficiency of the traditional Temporal averaging method we have modified it to cover some of the drawbacks to detect the moving objects in a nonstationary background. [2] METHODOLOGY The proposed methods are shown in two variants those are the following. The first one is adaptive updating the speed of background. The other one is adaptive threshold which can be given by the user. Variant (i): The first step of the proposed method is same as in the traditional Temporal Averaging Method. The background rate where x, y, λt is adaptive learning rate. The further step is to estimate the difference between the current pixel frame and the background. It can be determined as. 36

3 International Journal of Computer Engineering and Applications, Volume XI, Issue IV, April 17, ISSN where S: learning rate of the speed of updating x y, λt C: A user set parameter. λt+1(x,y) = μ. S.Dt+1(x,y) N + (1 μ). λt(x,y) C determines the range of x, y, λt. N is the dynamic range of the processed signal in levels. The method aims to adapt the speed of the updating background when there is a rapid change in the false positives or a fast-moving background. The disturbances for detecting the motion could be the false positives and the windy climate or repeating backgrounds. Noise can be cleared by unmoving object subtraction. In the last step, the estimation of the foreground mask is done, Mt x, y as same as the above equation. Variant (ii): In this phase, we make threshold T of the foreground mask M t(x, y) adaptive for each pixel of the current frame. The false positive alarm is prevented by the threshold T t(x,y) should be greater than D t(x,y). Pixels that represent dynamic backgrounds assume frequently high levels of difference D t(x, y). Updating the threshold T for a current frame t at a current pixel location (x,y) is given by the rule: Tt+1(x,y) = (λt(x,y)/c) (Dt(x,y) + ΔT) + (1 λt(x,y) C ). Tt(x,y) where k j, Dt is the current level of the difference, ΔT is constant threshold summed over the current value of k j, Dt. The function of ΔT is increasing the threshold over the difference and reducing the number of false positives pixels. The other part of the method is same as the proposed Variant i. The background Bt (x,y). is updated according to above equation. The absolute difference Dt and the learning rate x y, λ t are given by the below equation. The foreground mask is estimated by the modified equation. Ft(x,y) = 0,Dt(x,y) Tt(x,y) Ft(x,y) = 1,Dt(x,y) > Tt(x,y) There are also some adjustable set of parameters and user set parameters like learning rate S and the parameter C. The C parameter, determines the range of λ. The values of the adjustable parameters depend upon the speed of the updating background or the dynamic background, high speed of moving corresponds to high levels of λ. By executing the videos by changing the values of the set of parameters we can obtain a range of values of the parameters where the correct or required output is clearly obtained. λ usually varies from 0.1 to 0.001, the C parameter is in range of 0.2 to 0.4 for the most successful outcomes. Again, the low levels of λ, C results in the failure of the moving detection. [3] RESULTS AND DISCUSSIONS All the videos are processed by using a MATLAB program in the version R2013a. The proposed method is tested for three different scenarios. All the videos are set with a resolution of 720x576 P. Poornima and G. Rithin Surya SaiNadh 37

4 A NOVEL MOTION DETECTION METHOD USING BACKGROUND SUBTRACTION MODIFYING TEMPORAL AVERAGING METHOD pixels and with a frame rate of 25 fps. The shooting camera is fixed or stationary for all the three videos. But the background is not stationary. The first video is the day light video with traffic and a group of cars moving before the signal in a city. The second video sequence is the night recorded same location as day video. The third video is a bit different from the above two, where some people are moving casually and some with cycle. But it is very windy climate video. The detection of moving objects can be effected by the false positive pixels which are not in foreground but are mistakenly detected. The traditional Temporal Averaging Method and the proposed adaptive threshold approach, adaptive updating speed method variants are executed and analyzed. Figure.3: Dark Background with Intensity Foreground Figure.4: Highlighted Motion Objects An example of the executed day light video footage with the intensity foreground is shown in the below image. The same method of execution is done for the night and windy climate videos. Figure.5: Original Footage and Intensity Foreground The evaluation of performance of the existing method and the proposed methods are compared using the accuracy metrics like recall, precision and F-measure. 38

5 International Journal of Computer Engineering and Applications, Volume XI, Issue IV, April 17, ISSN Recall: It is the ratio of detected true positives pixels to the total number of true positives in the ground truth. It is also known as detection rate. Recall = TP/(TP+FN) Where, TP is the total number of true positives pixels, which are correctly classified foreground pixels. FN is the total number of false negatives, which are actually foreground pixels, wrongly segmented as background. And (TP + FN) combined sum indicates the total number of pixels present in the ground truth. Only recall is insufficient to compare and analyze different complex methods, so it is generally combined with another parameter named precision. Precision: It is the ratio of detected true positives to the total number of items detected by the method, which is also known as positive prediction. Precision = TP/(TP+FP) Here, FP is the total number of false positive pixels, these pixels are actually background pixels which are wrongly classified as a foreground. And (TP + FP) combined sum indicates the total number of detected pixels from the output mask. F-measure: It is the harmonic mean of Precision and Recall values, also known as Figure of Merit. It can be written as. F-measure = (2* Recall * Precision)/(Recall + Precision) The value of F-measure varies between 1 and 0. Which are the higher and the lower values respectively. Fig.6: Recall-Precision. The adaptive Temporal averaging and the classical temporal averaging are implemented for the two variants and then the results are compared. The F-measure is calculated when the foreground image is subtracted to a ground truth image. P. Poornima and G. Rithin Surya SaiNadh 39

6 A NOVEL MOTION DETECTION METHOD USING BACKGROUND SUBTRACTION MODIFYING TEMPORAL AVERAGING METHOD Fig.7: F-measure of learning rate λ Variant (i). In the first video footage that is in the day light the adaptive threshold method is more accurate by 1% than the traditional. The difference in accuracy is minor because in daylight there are no change in any conditions, as the light will be dominated by the sunlight so no dynamic background will be available. The variant(i) is more accurate than the other variant in the first video footage. The next video that is in the night footage the adaptive threshold is more accurate that is 4% improvement is shown in the F-measure than the classical temporal averaging method. This is because of the dynamic background effect, low contrast and non-constant lighting conditions due to night effect. The proposed method adapts the change in background very fast than the existing method and also the reduction in false positives is observed. Thus, there is increase in F-measure and precision. In the third footage that is the windy climate, the adaptive threshold method stood in the first place as the weather condition is windy there are waving trees in the video sequence. The resulting false positives are very low and there is an increase in F-measure nearly a 5% than the traditional temporal averaging method. Fig.8: F-measure of learning rate λ Variant(ii). 40

7 International Journal of Computer Engineering and Applications, Volume XI, Issue IV, April 17, ISSN The second variant, adaptive threshold method is executed and the results are compared. The number of false positives are minimized and the adaptive threshold value is high. Thus, precision increases and recall decreases. So, the F-measure is a trade-off between these two parameters. [4] CONCLUSION Using this method, we can not only adapt to sudden illumination changes but also slow changes are updated. The method requires less computational power, versatile, robust and less complex than the classical Temporal averaging method. Hence, Adaptive temporal averaging method justifies that the maximum performance of the method depends on the proper set of its parameters. In the results the F-measures are compared and shows that the proposed method is more accurate than the existing methods where a dynamic or non-stationary background is present. REFERENCES [1] Access and Privacy Service Alberta, Guide to Using Surveillance Cameras in Public Areas, Freedom of Information and Protection of Privacy June 2004, ISBN [2] Cheung, S. C. and Kamath, C. Robust Techniques for Background Subtraction in Urban Traffic Video, Visual Communications and Image Processing San Jose, January 2004, pp [3] DALLEY, G., MIGDAL, J., GRIMSON, W. Background subtraction for temporally irregular dynamic textures. In WACV Colorado (USA), January 2008, p [4] ELGAMMAL, A., DURAISWAMI, R., HARWOOD, D., DAVIS, L. Background and foreground modeling using nonparametric kernel density estimation for visual surveillance. Proceedings of the IEEE, July 2002, vol. 90, no. 7, p [5] Ghani, K.A. and Yosri, H. (2011) Application of Artificial Intelligent for Armour Vehicle Detection Using Digital Image Processing for Aerial Application. Proceedings of the International Conference on Advanced Science, Engineering and Information Technology, Putrajaya, January 2011, [6] HA, D. M., LEE, J. M., KIM, Y. D. Neural-edge-based vehicle detection and traffic parameter extraction. Image and Vision Computing, 2004, vol. 22, no. 11, p [7] Jabri, S., Duric, Z., Wechsler, H., Rosenfeld, A., Detection and Location of People in Video Images Using Adaptive Fusion of Color and Edge Information Pattern Recognition, Proceedings. 15th International Conference, Barcelona, Spain, 3 7 Sep 2000, pp vol. 4. [8] Koller, D., Weber, J., Huang, T., Ma, J., Ogasawara, G., Rao, B., Russell, S. Towards Robust Automatic Traffic Scene Analysis in Real-time. In Proc. of the 12th Int l Conference on Pattern Recognition (ICPR-94), pp , Jerusalem, Israel, October 9 13, pp [9] McFarlane N, Schofield C. Segmentation and Tracking of Piglets in Images, Machine Vision and Applications, (Vol. 8, No. 3). (1 May 1995), pp P. Poornima and G. Rithin Surya SaiNadh 41

8 A NOVEL MOTION DETECTION METHOD USING BACKGROUND SUBTRACTION MODIFYING TEMPORAL AVERAGING METHOD [10] S.-C. Cheung and C. Kamath, Robust techniques for background subtraction in urban traffic video, in Proc. EI- VCIP, 2004, pp [11] STAUFFER, C., GRIMSON, W. E. L. Adaptive background mixture models for real-time tracking. IEEE International Conference on Computer Vision and Pattern Recognition, June 1999, vol. 2, p [12] T. Agui, Y. Kokubo, H. Nagashashi, and T. Nagao, Extraction of FaceRecognition from Monochromatic Photographs Using Neural Networks, Proc. Second Int l Conf. Automation, Robotics, and Computer Vision, vol. 1, pp , [13] WANG, H., MILLER, P. Regularized online mixture of Gaussians for background subtraction. In IEEE International Conference on Advanced Video and Signal-Based Surveillance, AVSS Klagenfurt (Austria), September, 2011, p [14] Wu, J.P., Liu, Z.B., Li, J.X., Gu, C.D., Si, M.X. and Tan, F.Y. (2009) An Algorithm for Automatic Vehicle Speed Detection Using Video Camera. Proceedings of 4th International Conference on Computer Science & Education, Nanjing, July 2009,