Analysis of advantages and disadvantages of--vibe algorithm based on background modeling

First, the advantages of Vibe algorithm

Vibe background Modeling is a new, fast and effective moving target detection algorithm for moving target detection in neighborhood. Its advantages have the following two points:

1, the idea is simple, easy to realize. Vibe usually randomly select a neighborhood of 20 samples to establish a sample based background model for each pixel. With the advantages of high initialization speed, less memory consumption and less resource utilization, a quadratic sampling factor φ is used to make the finite sample cardinality approximate the infinite time window, that is, under less sample conditions, To ensure the accuracy of the algorithm, finally, a domain communication mechanism is used to ensure the spatial consistency of the algorithm.

2. High efficiency of operation. There are two reasons: first, the vibe background model is based on a small number of samples of the background model, the second is to optimize the background model of similarity matching algorithm. The vibe background model is a background model based on n samples, in order to get the best n value, we selected N as 5, 15, 20 and 25 respectively to compare the experiments, as shown in Figure 3.3, the experimental results show that N takes 20, 25 o'clock, the result is ideal, the calculation load is considered, and N is optimal. Compared with the calculated matching of the 3-5 Gaussian models with Gaussian mixture, the calculation of background model based on 20 samples has the advantages of low computational cost and fast detection speed. Vibe's background model similarity matching function is only related to R and Umin, the sample in the background model and the Euclidean distance of the pixels to be classified are less than the number of R Umin, match the background model, judge the background, in the actual implementation process, the optimization algorithm, once found umin matching samples, stop the calculation, To judge the background of the pixel to be classified, this method greatly reduces the computation of the whole algorithm, thus improving the computational efficiency.

3, sample attenuation optimal. Some people handle complex scenes by increasing the sample cardinality (up to 200), and some combine two sub models to handle both fast and slow updates. In fact, the selection of replaced samples to update the background model, in essence, is the sample life problem, the traditional way to use the first out of the replacement strategy. The probability of each sample being selected as a replacement sample is equal in the vibe background model, which is independent of the length of the sample's existence time, which guarantees that the sample life in the background model is exponentially attenuated and the model update reaches its optimal state.

Disadvantages of Vibe algorithm

Vibe background modeling has the advantages of simple thinking, easy implementation and high computational efficiency, but the algorithm has its own limitations. The main ghosts, stationary targets, shadow foreground and incomplete motion target are the problems.

First question: Ghost-shadow problem. As shown in Figure 3.4, Vibe background modeling uses the first frame to initialize the background model of the initial frame, figure (d) is the result detected by the VIBE algorithm in Fig. (c), because there is a moving target in the first frame (as shown in figure (b)), not a real background image (as shown in figure (a)), and there is a phantom in the test result. As shown in Figure 3.4 (d). The causes of the ghost Shadow can be summed up in two categories: (1) The initial frame of the background model has the moving target, (2) The moving target's state transition, from the static to the motion, the update rate of the background model cannot be same as the background rate, and the false target is the Phantom.

Second question: The problem of stationary targets. As shown in Figure 3.5, figure (a) the people in the red box are waiting for the subway, from figure (a) to figure (c) After 498 frames, long time to stay in the movement, the person's moving target gradually absorbed by the background. In this paper, a moving target region with no significant displacements above 450 frames is defined as a stationary target region. The causes of static target problem can be classified into two categories: (1) Moving target from movement to Static, (2) moving target moving slowly. When the update speed of the vibe background model is too fast, the stationary or slow moving target is absorbed as a part of the background, at which time the two moving targets are not detected.

Third question: The shadow foreground problem. As shown in Figure 3.6, fig. (b) and fig (d) are the results of the detection of human motion target (a) and vehicle body moving Target (c) by using the VIBE algorithm respectively, because the light is obscured by body or body moving target, the background of projecting shadow area is mistakenly detected as the foreground of moving target. The existence of shadow leads to the inaccurate shape of the detected moving object, which affects other intelligent video processing modules, such as classification, tracking, recognition and analysis of subsequent targets. The source of the shadow foreground problem is: The light is obscured by the foreground of the moving target, the color of the projection shadow area is darker than the background color, that is, the distance between shadow and background color value is large, and the background difference is mistaken for the moving target foreground.

Question Fourth: Incomplete motion objectives. As shown in Figure 3.7, there is a void inside the person in figure (a), a fault occurs among the people in figure (b), and the upper body of the figure (c) appears to have a marginal remnant, and the windshield of the bodywork appears empty. Moving target can be divided into rigid objects and rigid objects, people belong to non rigid objects, cars belong to rigid objects, these two common detection objects of detection results have been incomplete phenomenon. Summing up the experimental results of FIG. 3.7, the incomplete phenomena of the moving target can be divided into three categories: a There are a lot of voids inside the moving target. (b) The edge of the moving target is incomplete, showing C-shaped sag. (c) There is a fault in the middle of a moving target.

The main causes of the problem of incomplete motion target are the following two points: (1) The VIBE algorithm has its own defects. The VIBE sample model based on the statistical principle is limited to the sample number of the model, and when the sample tends to infinity, the scene can be accurately described, which is impossible to be achieved in practical application, and (2) the complexity and variability of the scene or the moving target. There are three kinds of cases: 1 instantaneous ray mutation, the background model is too late to update, 2 foreground and background color is similar, the foreground is mistaken for background, 3 noise is disturbed, and the isolated noise point and the connected noise region appear.