The first step is initializing, it allows the observation and preparation of new frames for the detection-monitoring. For this we are based on several techniques like Meanshift (motion estimation) algorithm, GraphCuts (segmentation) under Blob tracking. We can decompose the initialization phase in two phases: preprocessing and prediction .  
The first frame initialization is based on a simple background subtraction, which we will use Gaussian mixture MoG. In this phase we will make two classifications of pixels: objects (blobs) pixels and others pixels. When the blob is detected for the first time, it stores its histograms in a list of blobs; this list will be to update each suspension or detection of objects. From the second frame, the initialization phase is to predict objects (blobs), it functions using energy minimization algorithm in which GraphCuts warranty segmentation at a modest cost, blobs obtained segmentation by this, we will compare the histogram with that of previous frame, by updating a list of blobs.
Blob tracking and Shape Feature extraction: In the traditional blob tracking feature two are taken to model an object by contour modeling and by color modeling. In this article we will add more features while getting a reliable and accurate detection. – Moment: The statistical concept of the moment is defined by the following formula: mpq =  (u,v)∈R I(u, v).upvq (1) It describes the moment of the order p,q for a discrete (image) function I(u, v) ∈ R. All the following definitions are also generally applicable to regions in grayscale images. This moment is used by central moment. – Central moments: To compute position-independent (translationinvariant) region features, the region’s centroid,which can be determined precisely in any situation, can be used as a reference point. In other words, we can shift the origin of the coordinate system to the region’s centroid x¯ = (¯x, y¯) to obtain the central moments of order p,q: upq(R) =  (u,v)∈R I(u, v).(u − x¯) p(v − y¯) q (2) For a binary image I(u, v)=1. – Eccentricity: Similar to the region orientation, moments can also be used to determine the "elongatedness" or eccentricity of a region. We adopt the following definition because of its simple geometrical interpretation: Ecc(R) = a1 a2 = μ20+μ02+ √(μ20−μ02)2+4.μ2 11 μ20+μ02− √(μ20−μ02)2+4.μ2 11 (3) where a1 = 2λ1, a2 = 2λ2 are multiples of the eigenvalues λ1, λ2 of the symmetric 2 × 2 matrix : A =  μ20 μ11 μ11 μ02 formed by the central moments μpq of the region R. • GraphCuts: In practice, there are several methods of segmentation interactive of foreground/background classical tools of image segmentation that use either texture (color) or edge (contrast). In [7] an approach based on optimization by graph cut has been successfully developed, which combined the two types of information. Which shows that the graph Cuts is the best tool for making a good discrete optimization. In cases where finding a global minimum is not possible, we have seen there exist many efficient algorithms based on Graph Cuts [9, 10].

4 Algorithm and Theory

The approach is typically able to react according to the content of the captured scene. In which we deal with a lot of problems namely: lighting problem and luminance change issues such as movement changes the background and changes in small claims. For this we proceed by a able to make a simultaneous segmentation and tracking based on storing blobs for each iteration. For this reason the system must predict the existence of the blob based on a list of blobs where comparing their histograms, this is going to reduce the processing time. After this step it may have new birth of blobs, blobs if the list is updates (correction step).Input sequences are in the form of segmented images (by GraphCuts), each blob is identified by a unique number that is identical in all sequences of images on this number will be displayed above each detected blob. At first we select the image that contains the pixel values objects that were used to define the blobs (removal of objects from the background).When the blob is detected, you must check that the blob is already featured in the previous images, why we do what we call: verification of similarity which is checked: The location (pixel coordinates), the size, color (texture), orientation (time computation of second-order) form (we calculate the eccentricity) and histogram (color distribution in the blob) with other blobs in the previous frames. This criterion of similarity varies in functions of all the properties mentioned above by representing a percentage. As the percentage is large it is more likely to have a similar blob. If we found a similar blob is recovered easily from a list of already detected blobs, if the blob is considered new and not connected to a previous blob.In this case it creates the new blob that must Identify (associate an id) and extract its features which will store the list of blobs. When detected blobs disappear in an image sequence, they may reappear in the following images in the hope of preserving the id. For this we define a threshold which is the minimum number of images for that we should start tracking this blob. To avoid small movements that does not correspond to the target object. We also define another threshold: the minimum number of pixels containing a blob for it to be considered a blob active, to avoid false detections dedicated noise.

5 Conclusion  

In this article we presented a new approach that combines several principles and methods to achieve the goal of ensuring a better detection and monitoring of vehicle in video sequences, on which we are based at the beginning of the first modeling approach background. In an initialization step, we invoked blobs to identify patterns of body parts through a contour analysis and segmentation region (Mean-shift). After initialization, based on a cost minimization (GraphCuts) approach is used for matching patterns of several related components and monitoring.This minimization of energy is applied to an area of the image containing parts of the body. The results showed very good performance in terms of computational cost and quality of precision in the detection-monitoring. This method is not limited only to the detection of vehicles, it can be also applied in other utility such as video surveillance.

References

1. Comaniciu, D., Ramesh, V., Meer, P.: Real Time Tracking of Non-Rigid Objects using Mean Shift. In: IEEE Computer Vision and Pattern Recognition II, pp. 142–149 (2000)
2. Collins, R.T.: Mean-shift Blob Tracking through Scale Space. In: IEEE, pp. 234–240 (2003)
3. Liang, D., Huang, Q., Jiang, S., Yao, H., Gao, H.: Mean shift blob tracking with kernel histogram filtering and hypothesis testing, pp. 605–614 (2005)
4. Fukunaga, K., Hostetler, L.: The estimation of the gradient of a density function, with applications in pattern recognition. IEEE Trans. on Information Theory 21(1), 32–40 (1975)
5. Comaniciu, D., Ramesh, V., Meer, P.: Real-time tracking of non-rigid objects using mean shift. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, vol. II, pp. 142–149 (2000)
6. Liang, D., Huang, Q., Jiang, S., Yao, H., Gao, W.: Mean-Shift Blob Tracking with Adaptive Feature Selection and Scale Adaptation. In: IEEE International Conference on Image Processing, ICIP 2007, vol. 13(3), pp. 369–372 (2007)
7. Zhao, L., An, G., Zhang, F., Wang, H., Dai, G.: Scale Adaptation of Mean Shift Based on Graph Cuts Theory. In: 12th International Conference on ComputerAided Design and Computer Graphics, pp. 202–205. Chinese Academy of Sciences (September 2011)
8. Bugeau, A., Pérez, P.: Track and Cut: Simultaneous Tracking and Segmentation of Multiple Objects with Graph Cuts. EURASIP Journal on Image and Video Processing, Ref: 2008:317278 (2008)
9. Boykov, Y., Veksler, O., Zabih, R.: Fast approximate energy minimization via graph cuts. IEEE Transactions Pattern Analysis and Machine Intelligence 123, 1222–1239 (2001)
10. Boykov, Y., Funka-Lea, G.: Graph Cuts and Efficient N-D Image Segmentation. International Journal of Computer Vision 70(2), 109–131 (2006)