Research on Vehicle Detection Method Based on Background Modeling

The statistical average method, also called the mean method, is essentially a statistical filter idea. In a period of time, the collected images are added together, and the average value is taken as the reference background model. It is to obtain the gray average value ofN frames of images in the sequence image as the estimated value of the background image to weaken the interference of the moving object to the background. The specific calculation formula is shown in (1) A v g k = 1 N ( f k + f k − 1 + f k − 2 + … + f k − N + 1 )

Avg_k is the background model established for the acquisition of the Frame K image system; N is the average number of frames; f_k, f_k-1, …, f_k-N+1 are the continuous sequences frame.

The statistical average method is simple and fast, but if easily causes noise accumulation and mixing. This method is more suitable for a small number of continuous motion objects in the scene, and the background is visible in the most of the time. While there are a large number of moving objects, especially when they are moving slowly, there will be a large deviation in the background of this situation.

The Gaussian distribution background model was first proposed by N. Friedman et.al and is divided into background models with single Gaussian distribution and mixed Gaussian distribution. The single Gaussian model method regard the change in the gray value of each pixel in the background image as a Gaussian random process, and establishes a Gaussian model for each pixel in the image, which is achieved by continuously updating the Gaussian model background image. The mixed Gaussian model uses K (basically 3~5) Gaussian models to characterize the features of each pixel in the image. After the new image is acquired, the mixed Gaussian model is updated. Each pixel in current image is mixed with a Gaussian mixture model. Matching to determine the pixel belongs to a background point or a front sight.

The section focuses on the hybrid Gaussian background modeling method. This modeling method uses the statistical information such as the probability density of a large number of sample values for a long time to represent the background. Use the statistical difference (such as 3σ principle) to judge the target pixel. This method can model complex dynamic backgrounds with a large amount of computation. Suppose that any pixel (x,y) in the background obeys a Gaussian model composed of K Gaussian distributions, as shown in follow.

η(I_{(x, y)}, μ_{(x, y)}, σ_(x,y,j)) means that the j-th Gaussian probability density, means value is μ_(x,y,j), the variance is σ_(x,y,j). And the ω_j is the weight of the j-th Gaussian distribution.

The pixel values observed at the current moment are compared with the current K Gaussian distribution functions in descending order of weights to obtain the best match. If there is no match, the pixel is a foreground spot.Otherwise it is a background point. The Gaussian distribution background model has a large amount of calculation, many storage parameters, and takes a long time, which is not conducive to practical application.

By using the statistical average method, the background of the mixture Gaussian model and the method proposed in this paper, the background model is obtained by modeling the background image of a video of a traffic video surveillance video. The video of this video screen is 1080 * 960, the frequency is 25fps, and the time length is 10 seconds.The experimental results are shown are shown in Fig. 1 below.

Figure 1.

It can be seen that in the background frame extraction process, the average statistical method extracts the background [Fig. 1 (b-1)] and the blurred background is affected by the camera shake of the video camera, and the effect is worst.

The parameters of the mixed Gaussian model are set as follows: the number of pixel models is 5, the initial variance is 30, and the learning rate of model weights is α = 0.005, T = 0.7;

The resulting background [Fig. 1 (c-1)] is clearer, but the background extracted at the lower left corner of the screen is still somewhat blurred; the corresponding parameter setting of this algorithm is: T_h = 15, α = 0.2, threshold = 12. The background [Fig. 1 (d-1)] obtained by the method proposed in this paper is of good quality and closer to the real background.

Finally, after the difference, binary processing, and morphological processing, the image is finally obtained. The statistical mean method [Fig. 1 (b-5)] is noisy, and the extraction result is not clear. The result graph obtained by using the Gaussian method [Fig. 1(c-5)] works well, but there is still noise. The result graph obtained by using this algorithm [Fig. 1(c-5)] has the best effect, the noise is very small, the extracted vehicle is clearer and the connectivity is better.

The final conclusion of the comparison experiment is not only drawn from the experimental results, but also from the performance side. Table 1 compares the performance ofthe three methods, including the time used for the entire inspection process, the memory footprint, and the CPU usage. It can be seen that the algorithm proposed in this paper compared with the other two methods, it consumes less time, consumes less memory, and has a small CPU occupancy rate.

TABLE I.