CN103198300B - Parking event detection method based on double layers of backgrounds - Google Patents

Abstract

The invention discloses a parking event detection method based on double layers of backgrounds. The parking event detection method based on the double layers of backgrounds mainly comprises the steps of double-layer background modeling, secondary background replacement, parking detection, and state table updating. According to the parking event detection method based on the double-layer background, an absolute value of the pixel difference of corresponding points of a main background and the second background is used for judging whether a stopping object appears, detection is carried out according to the outline of the stopping object, if the stopping object is a vehicle, a parking state is calibrated and updated, and in addition, the background is replaced by judging whether the foreground in a background model is empty or not. The parking event detection method based on the double layers of backgrounds can well detect a parking event in real time, and is simple and capable of reducing influence of the environment on a detection result by means of the double layers of backgrounds and accurately recording a parking position and the parking state due to the establishment of a parking event state table. The parking event detection method based on the double layers of backgrounds is capable of being used in different occasions such as an expressway, a parking lot, and an urban road due to the setting of detection parameters and interested areas, high in accuracy and good in real-time performance.

Description

A Parking Event Detection Method Based on Two-layer Background

technical field

The invention relates to the field of video detection, in particular to a parking event detection method based on a double-layer background.

Background technique

With the rapid development of the national economy and the rapid increase of motor vehicles, my country's traffic problems are becoming increasingly serious. Traffic jams and secondary accidents caused by traffic incidents such as parking of motor vehicles, throwing objects, suspension of traffic accidents, and landslides caused by harsh environments are increasing. This accident is sudden and accidental, and once it occurs, it will cause great casualties and property damage. In the past, the detection of parking incidents was mainly through manual monitoring, traffic flow information collection and other methods, which greatly consumed the human, material and financial resources in traffic control. Therefore, it is particularly important and necessary to establish a video-based automatic parking event detection system.

At present, a variety of video-based parking event detection methods have been developed in the world, mainly based on virtual frame pixels, grayscale statistics, based on target speed and based on single background block segmentation parking event detection methods.

The method based on virtual frame pixels and grayscale statistics obtains moving pixels and static pixels through background difference method, and judges whether the vehicle stops according to the changes of pixels or grayscale in the delimited area; although this method is simple in algorithm, it is vulnerable to external light Conditions and other interference, and the virtual detection area needs to be manually set, so the detection practicability is poor. The parking event detection method based on the target speed needs to track the vehicle in real time, calibrate the scene, and calculate the moving speed of the target in real time; the algorithm is relatively complex and has a high false alarm rate. The block segmentation parking detection method based on a single background saves three different backgrounds and compares them pairwise to determine whether there is a suspicious block, and then detects the suspicious block to determine the parking event; this method needs to divide the image into blocks and is not suitable for road conditions In complex areas, and the three background images saved are from the same background model, the detection accuracy is not high. The above method lacks effective analysis for vehicles stopping or driving away, and the effectiveness and practicability of the algorithm for analyzing parking events can no longer meet the realistic requirements.

Contents of the invention

The object of the present invention is to provide a parking event detection method based on a double-layer background that is less affected by the environment and has accurate parking detection.

Technical solution of the present invention is:

A kind of parking incident detection method based on double-layer background, it is characterized in that: comprise the following steps: realize by following steps:

(1) Establish two different backgrounds—the main background and the secondary background, using the characteristics of the main background’s fast update speed and sensitivity to the stop target, and the secondary background’s slow update and slow response to the stop target. the difference between the two images;

(2) Make a difference to the pixels corresponding to the double-layer background and calculate the absolute value to obtain a stationary target, and then binarize the difference image of the double-layer background; suppress the shadow of the corresponding pixel by HSI, and eliminate the shadow of the corresponding pixel of the target. Perform a closed operation on the binary image to eliminate discontinuous holes;

(3) According to the focal length, height and angle of the camera, weight the target pixels in the image, increase the weight of the distant target pixels, and decrease the weight of the nearby target pixels, and calculate the weighted target pixel and value , when the threshold is reached, the parking event counter S is increased by 1, and when S is greater than the threshold, the binary image is saved;

(4) Filter the binary image, perform segmentation and contour detection on the target in the image, set the detection sensitivity, draw a rectangular frame for the target whose pixel value is greater than the sensitivity threshold, and judge whether the target is a vehicle according to the aspect ratio , when the target is a vehicle, record the coordinates of the intersection of the diagonals of the rectangular frame and store it in the parking event state table;

(5) Process and analyze the pixel to determine whether the object is a vehicle; if the object is judged to be a vehicle, trigger a parking event alarm and assign the current frame of the main background to the secondary background, and continue to compare differences; When there is no target, keep the frame image as a pure background;

(6) When the foreground in the secondary background model is empty, store the current background image as a pure background. When a stop target is detected, store the current frame of the main background and compare it with the pure background image. If two frames When the difference in the image is less than the threshold, skip to step (2) and continue running. If the difference between the two frames of images is greater than the threshold, replace the secondary background with the current frame of the main background, skip to step (2) and continue, and detect parking again When an event occurs, compare the coordinates of the center point of the target, judge whether the target is leaving, and update the state table.

Compared with the parking detection method based on virtual coil statistics, the parking detection method based on target tracking and the block segmentation parking detection method based on single background, the parking event detection method based on the double-layer background of the present invention is less affected by the environment and does not need to set the image And image calibration, the algorithm is simpler and more feasible, which reduces the computational complexity of parking detection and improves the real-time performance of parking event detection. Eliminate interference through HIS shadow suppression and morphological filtering processing, making parking detection more accurate; establish a status table update, and accurately record the parking position and status.

Description of drawings

The present invention will be further described below in conjunction with drawings and embodiments.

Figure 1 is a flowchart of a parking detection method based on a double-layer background.

detailed description

In conjunction with the accompanying drawings and embodiments, a parking event detection method based on a double-layer background is provided. The method establishes a secondary background—a mixed Gaussian background model and a main background——RunningAvg background, and calculates the difference between the two layers of background, and calculates the difference The value image is binarized to obtain a static target or a low-speed moving target. Through the statistics and contour recognition of the binary image to detect whether there are stopped vehicles or leftovers, compare the current RunningAvg background with the pure background image to judge whether the road is smooth, update the parking status table and the secondary background, and proceed to the next step One-shot parking event detection. The specific implementation steps are as follows:

Step 1, grayscale processing is performed on the input video image I _n to obtain a grayscale image _Ingray , and a main background and a secondary background are established through the grayscale image _Ingray ;

Establishment and update of main background, secondary background and pure background:

The establishment of the main background model - the RunningAvg model, the RunningAvg model is shown in the following formula:

${\mathrm{<span\; class="notranslate">\; B</span>}}_{\mathrm{<span\; class="notranslate">\; Avg</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; \&PartialD;</span>}_{\mathrm{<span\; class="notranslate">\; avg</span>}}{\mathrm{<span\; class="notranslate">\; B</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>{<span\; class="notranslate">\; \&PartialD;</span>}_{\mathrm{<span\; class="notranslate">\; avg</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span>$

——Bavg(i,j) is the RunningAvg background model;

——B _n (i, j) is the updated background value of the nth frame,

——B _n-1 (i, j) is the background value of the n-1th frame,

——I _n (i, j) is the gray value of the current video frame,

—— is the update rate.

The present invention is The following improvements have been made: ${\&PartialD;}_{\mathrm{avg}}={\&PartialD;}_{\mathrm{avg}1}{m}_{\mathrm{no}}+{\&PartialD;}_{\mathrm{avg}2}(1-m_{\mathrm{no}})$

—— $m_{\mathrm{no}}=\left\{\begin{array}{cc}0& {\mathrm{D.}}_{\mathrm{no}}(i,j)<T\\ 1& {\mathrm{D.}}_{\mathrm{no}}(i,j)\&Greater\; Equal;T\end{array}\right.,$ M _n is a status bit;

D _n (i,j)=|I _n (i,j)-I _n-1 (i,j)|, Dn(i,j) is the residual image of adjacent frames;

are the variable weighting parameters respectively;

The establishment of the secondary background model - first initialize several pre-defined Gaussian models, initialize the parameters in the Gaussian model, and calculate the parameters that will be used later. Secondly, process each pixel in each frame to see if it matches a model, if it matches, put it into the model, and update the model according to the new pixel value, if it does not match, Then build a Gaussian model with this pixel, initialize the parameters, and represent the most unlikely model in the original model. Finally, select the most probable models in the front as background models to pave the way for background target extraction.

The mixed Gaussian model p(x _N ) is the statistical value of the probability of occurrence of pixels, as shown in the following formula:

$\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; K</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}\mathrm{<span\; class="notranslate">\; \&eta;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; ;</span>{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>$

——w _j is the weight of the k-order Gaussian background, x _N is the input sample, and θ _j is the observation value;

——η(x;θ _k ) is the standard normal distribution of k-order Gaussian, and its expression is as follows:

$\mathrm{<span\; class="notranslate">\; \&eta;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ;</span>{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; \&eta;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ;</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&pi;</span>}<span\; class="notranslate">\; )</span>}^{\frac{\mathrm{<span\; class="notranslate">\; D.</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}{<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; |</span>}^{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}}{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span>}$

——μ _k is the mean value;

——∑ _k = σ ² I is the variance;

The background pixel judgment formula is as follows:

${\mathrm{<span\; class="notranslate">\; B</span>}}_{\mathrm{<span\; class="notranslate">\; Gauss</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; arg</span>}}_{\mathrm{<span\; class="notranslate">\; b</span>}}\mathrm{<span\; class="notranslate">\; min</span>}<span\; class="notranslate">\; (</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>}{\overset{\mathrm{<span\; class="notranslate">\; b</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; )</span>$

${\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; w</span>}}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; )</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; w</span>}}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; a</span>}\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span>$

——B _Gauss is the background pixel;

——a is the learning rate;

——w _k is the initialization weight, is w _k expected value;

——T is the background threshold;

The mixed Gaussian model uses K Gaussian models to represent the characteristics of each pixel in the image. After a new frame of image is obtained, the mixed Gaussian model is updated, and each pixel in the current image is matched with the mixed Gaussian model. If it is successful, the Points are background points, otherwise they are foreground points. Looking at the entire Gaussian model, it is mainly determined by the two parameters of variance and mean. For the learning of mean and variance, adopting different learning mechanisms will directly affect the stability, accuracy and convergence of the model. Since we are modeling the background extraction of moving targets, we need to update the two parameters of the variance and mean in the Gaussian model in real time. In order to improve the learning ability of the model, the improved method adopts different learning rates for the update of the mean and variance; in order to improve the detection effect of large and slow moving targets in busy scenes, the concept of weight mean is introduced, and the background image is established and Update in real time, and then classify the foreground and background of the pixel points by combining the weight, the mean value of the weight and the background image.

The so-called pure background, when the foreground value of the secondary background-mixed Gaussian background model is empty, save the current image B _{G_clear} as the pure background.

Step 2, initialize the parking state table, and mark whether there is a stopped vehicle in the parking scene.

Step 3, the value of the corresponding pixel of the double-layer background is differenced and the absolute value is obtained to obtain the image of a stationary target or a low-speed moving target D(i,j)=|B _Avg (i,j)-B _Gauss (i,j) |, perform binarization on the double-layer background difference image, and use the binarization threshold T _h . Through HSI shadow suppression, and eliminate the shadow of the corresponding pixel of the target, perform a closed operation on the image, eliminate discontinuous holes, and finally obtain the image D _det (i,j);

Step 4: Weight the white pixels of the image D _det (i, j) according to the focal length and angle of the camera. In this example, the image is roughly divided into 5 parts to weight the pixels. From far to near, the weights are 2.0, 1.6, 1.2, 1.1 and 1 respectively. Perform statistics on the weighted pixels. When it is greater than the threshold T _p , in this example T _p =220, the counter S is added by 1, otherwise S=0; when S is greater than or equal to 90, the binarized image D _obj is saved;

Step 5: Filter the image D _obj , and perform segmentation and contour detection on the target in the image, set the detection sensitivity T _p , and determine the upper left edge and lower right corner of the target through contour detection for the target whose pixel value is greater than or equal to T _p The edge is marked with a box, and non-vehicle targets are removed by calculating the aspect ratio of the box and the number of white pixels in the box,

The coordinates of the upper left corner of the contour are (x ₁ ,y ₁ ), and the coordinates of the lower right corner are (x ₂ ,y ₂ ), then the length is l=max(|x ₁ -x ₂ |,|y ₁ -y ₂ |), The width is w=min(|x ₁ -x ₂ |,|y ₁ -y ₂ |). when When established, the target is a vehicle, where t ₁ and t ₂ take values according to the aspect ratio of the vehicle. When the target is a vehicle, trigger a parking event alarm and record the diagonal intersection coordinates of the rectangular frame, and store it in the parking event state table;

Step 6, the coordinates of the intersection of the diagonals of the box are (x _i1 , y _i1 ) where And store the coordinate information, when it is detected that the stopping target is a vehicle again, the coordinates (x _j1 , y _j1 ) of the diagonal intersection point of the next batch of vehicle boundary boxes are obtained, when max(|x _i1 -x _j1 |,|y If _i1 -y _j1 |)≤Tc, it is considered that there is a vehicle leaving, the vehicle coordinates stored in the state table are eliminated, and the number of stopped vehicles is updated, that is, the value of vehicle existence is reduced by 1. When max(|x _i1 -x _j1 |,|y _i1 -y _j1 |)>T _c , it is considered that there is a vehicle entering, increase the vehicle coordinates in the state table, update the number of stopped vehicles, that is, add 1 to the vehicle existence value, where T _c is the offset threshold, and w is the width of the vehicle;

Store the current RunnineAvg background and compare it with the pure Gaussian background image B _{G_clear} . If the difference between the two frames of images is less than the threshold T _d , skip to step 3 and continue. If the difference between the two frames of images is greater than or equal to the threshold T _d , use The RunnineAvg background replaces the Gaussian background, and skips to step 3 to continue execution;

Main background parameter setting, update speed parameter

Secondary background parameter setting, Gaussian distribution weight sum threshold T=0.7, background threshold T=2.5, learning rate Initial weight w _k =0.05, initial standard deviation ∑ _k =30;

In step three, the threshold T _h =25; in step five, the detection sensitivity T _p =200.

Claims (1)

1. a kind of parking event detecting method based on background double layer, is characterized in that：Comprise the following steps：

(1) set up the main background of the different background of two-layer and time background, using the renewal speed of main background fast, to stopping target More sensitive characteristic, and relatively slow, slower to the stopping target response characteristic of secondary context update, compare the difference of two width figures；

Main background model set up RunningAvg model, RunningAvg model is shown below：

$B_{Avg}(i,j)={\&part;}_{avg}{B}_{n-1}(i,j)+(1-{\&part;}_{avg})I_n(i,j)$

——B_Avg(i, j) is RunningAvg background model；

——B_n(i, j) is the background value after n-th frame renewal,

——B_n-1(i, j) is the background value of the (n-1)th frame,

——I_n(i, j) is the gray value of current video frame,

——For renewal rate；

RightCarry out following improvement：

—— $M_n=\{\begin{array}{cc}0& D_n(i,j)<T\\ 1& D_n(i,j)\&GreaterEqual;T\end{array},$ M_nFor mode bit；

D_n(i, j)=| I_n(i,j)-I_n-1(i, j) |, Dn (i, j) is consecutive frame residual image；

It is respectively variable weighting parameter；

The foundation of secondary background model initializes predefined several Gauss model first, and the parameter in Gauss model is entered Row initialization, and the parameter that will use after obtaining；Secondly, each of each frame pixel is processed, sees it Whether mate certain model, if coupling, be classified in this model, and this model is updated according to new pixel value, If mismatching, a Gauss model, initiation parameter being set up with this pixel, acting on behalf of model most unlikely in original model； Above several most possible models are finally selected as background model, to be that target context extraction is laid the groundwork；

Mixed Gauss model p (x_N) it is pixel probability of occurrence statistical value, it is shown below：

$p\left(x_N\right)=\underset{j=1}{\overset{K}{\&Sigma;}}{w}_j\mathrm{\&eta;}(x_N;{\mathrm{\&theta;}}_j)$

——w_jFor the weight of k rank Gaussian Background, x_NFor input sample, θ_jFor observed value；

——η(x；θ_k) for k rank Gauss standard normal distribution, its expression formula is as follows：

$\mathrm{\&eta;}(x;{\mathrm{\&theta;}}_k)=\mathrm{\&eta;}(x;{\mathrm{\&mu;}}_k,c)=\frac{1}{{\left(2\mathrm{\&pi;}\right)}^{\frac{D}{2}}|{\mathrm{\&Sigma;}}_K{|}^{\frac{1}{2}}}{e}^{-\frac{1}{2}{(x-{\mathrm{\&mu;}}_k)}^T{\mathrm{\&Sigma;}}_k^{-1}(x-{\mathrm{\&mu;}}_k)}$

——μ_kFor average；

——∑_k=σ²I is variance；

Background pixel judgment formula such as following formula：

$B_{Gauss}(i,j)={\arg }_{b}min(\underset{j=1}{\overset{b}{\&Sigma;}}{w}_j>T)$

${\hat{w}}_k^{N+1}=(1-a){\hat{w}}_k^N+a\hat{p}\left(w_k\right|x_{N+1})$

——B_GaussFor background pixel point；

A is learning rate；

——w_kFor initializing weight,For w_kExpected value；

T is background threshold；

Mixed Gauss model carrys out the feature of each pixel in phenogram picture using K Gauss model, obtains in a new two field picture Update mixed Gauss model afterwards, mated with mixed Gauss model with each pixel in present image, if success, judge This point is background dot, otherwise for foreground point；Take an overall view of whole Gauss model, mainly have variance and two parameters of average to determine, right Average and the study of variance, take different study mechanisms, will directly influence stability, accuracy and the convergence of model； Due to being it is therefore desirable to variance in Gauss model and two parameters of average in real time more to the modeling of the background extracting of moving target Newly；For improving the learning capacity of model, improved method adopts different learning rates to the renewal of average and variance；For improving numerous Under busy scene, the Detection results of big and slow moving target, introduce the concept of weights average, set up background image in real time more Newly, then in conjunction with weights, weights average and background image, the classification of foreground and background is carried out to pixel；

(2) to background double layer, corresponding pixel is made difference and is obtained absolute value, obtains static target, to background double layer error image Carry out binaryzation；By HSI cast shadow suppressing, and eliminate the shade of target respective pixel point, closed operation is carried out to bianry image, disappears Except discontinuous cavity；

(3) according to focal length of camera, height and angle, target pixel points in image are weighted, the target pixel points weights of distant place Increase, target pixel points weights nearby reduce, and obtain the object pixel value preset after weighting, when a threshold is reached, Parking Enumerator S adds 1, when S is more than threshold value, then preserves this bianry image；

(4) this bianry image is filtered, and the target in image is split and contour detecting, setting detection is sensitive Degree, the target being more than the threshold of sensitivity to pixel value draws rectangle frame, judges whether this target is vehicle according to length-width ratio, works as mesh When being designated as vehicle, record the diagonal intersecting point coordinate of this rectangle frame, be stored in Parking state table；

(5) pixel is processed and analyzed, judged whether this target is vehicle；If this target is judged as vehicle, triggering Parking is reported to the police and main background present frame is assigned to time background, proceeds comparison in difference；When there is no target in secondary background, Retain image as pure background；

(6) when the prospect in secondary background model is space-time, store current background image as pure background, stop when having detected Only during target, it is compared by the storage of main background present frame and with pure background image, if difference is less than threshold in two field pictures Skip to step (2) during value to continue to run with, if difference is more than threshold value in two field pictures, replace time back of the body with this main background present frame Scape, skips to step (2) and proceeds, when Parking is detected again, comparison object center point coordinate, and judge whether target sails From, and update state table.