CN103700118A - Moving target detection method on basis of pulse coupled neural network - Google Patents

Abstract

The invention discloses a moving target detection method on the basis of a pulse coupled neural network, which comprises the following steps of a, sensing a video image sequence by utilizing the pulse coupled neural network and extracting global features of a video image; b, establishing a global feature histogram of each pixel; c, for each pixel, utilizing the global feature histograms corresponding to first K frames to establish an initial background model of the pixel; d, for each pixel, calculating similarity of a global feature histogram of a current frame image and the corresponding global feature histogram in the background model and detecting whether the pixel is a moving target; e, for each pixel, utilizing the global feature histogram of the current frame image to update the background model of the pixel. The moving target detection method uses the integral characteristics of human visual perception for reference, utilizes the pulse coupled neural network to extract the global features of the image and is beneficial for inhibiting the negative influence of disturbance of a dynamic background on detection of the moving target so as to improve accuracy of detecting the moving target.

Description

Moving Target Detection Method Based on Pulse Coupled Neural Network

technical field

The invention relates to a moving target detection method based on a pulse-coupled neural network, belonging to the technical field of video image processing.

Background technique

In the intelligent video surveillance system, moving target detection technology is the basis of other post-processing (such as target tracking, target recognition and behavior analysis, etc.). In order to accurately and effectively segment moving target areas in various monitoring environments, it is of great significance to study moving target detection methods that are robust to complex dynamic scenes (such as illumination changes, background disturbances, etc.).

In order to solve the difficulties brought by complex dynamic scenes to moving target detection, at present, people mainly use the local neighborhood features of pixels to describe the background model of pixels, and extract the local area features of the image to improve the feature for the area. Robustness to illumination changes and background perturbations. The limitation of this method is that because it fails to make good use of the global features of the image, it can only effectively suppress the background disturbance when the dynamic background motion is limited to a local area, and it is not suitable for dynamic background environments with violent motion. (The term “limited to the local area” here means limited to the area where the local features are calculated. Qualitatively speaking, when the dynamic background moves in the area where the local features are calculated, the calculated local area features tend to be stable; When the dynamic background motion exceeds the area where the local features are calculated, the obtained local features are not easy to stabilize, and the background model established by it is also less robust to the dynamic background with violent motion). Therefore, there is an urgent need for a dynamic object detection method that can be applied to a dynamic background environment with severe motion and can improve detection accuracy and robustness at the same time.

Contents of the invention

The purpose of the present invention is to provide a moving target detection method based on a pulse-coupled neural network, which can effectively solve the problems in the prior art, especially when the dynamic background movement is limited to a local area, the background can be effectively suppressed. Perturbation is not suitable for the problem of dynamic background environment with heavy motion.

In order to solve the above-mentioned technical problems, the present invention adopts the following technical solutions: the moving target detection method based on pulse-coupled neural network comprises the following steps:

a. Use the pulse-coupled neural network to perceive the video image sequence and extract the global features of the video image;

b. Establish a global feature histogram for each pixel;

c. For each pixel, use the global feature histogram corresponding to the previous K frames as the initial background model of the pixel;

d. For each pixel, calculate the similarity between the global feature histogram of the current frame image and the corresponding global feature histogram in the background model, and detect whether the pixel is a moving target;

e. For each pixel, use the global feature histogram of the current frame image to update its background model.

Preferably, step a specifically includes: a neuron in the pulse-coupled neural network corresponds to a pixel in the video image; for a neuron located at (i, j) in the pulse-coupled neural network, it receives external stimulus information S at time n After _ij and other neurons in the adjacent k×l neighborhood are affected by the pulse information {Y _kl } at time n-1, feedback input F _ij , linear connection input L _ij , internal activity item U _ij , membrane potential dynamic threshold θ _ij , The pulse output Y _ij of the pulse generator and the global feature Q _ij of the extracted video image are respectively:

F _ij (n) = S _ij ;

${\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>& \underset{\mathrm{<span\; class="notranslate">\; kl</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; kl</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; 0</span>}\\ \mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; otherwise</span>}\end{array}<span\; class="notranslate">\; ;</span>\right.$

U _ij (n)=F _ij (n)·(1+βL _ij (n));

${\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}}{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ;</span>$

${\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>& {\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; otherwise</span>}\end{array}\right.<span\; class="notranslate">\; ;</span>$

Q _ij (n)=Q _ij (n-1)+(iter-n)Y _ij (n);

Among them, iter is the total number of iterative perception of the pulse coupled neural network; S _ij is the pixel gray value of any band image perceived by the neuron at (i,j); α _θ is the decay time constant of the dynamic threshold θ _ij ; V _θ is the threshold amplification factor; β is the connection strength.

Preferably, step b specifically includes: according to the characteristics of the three band images of the color image R, G, and B, using a pulse-coupled neural network to perceive the three band images of the color image R, G, and B respectively, and after iter times, respectively Obtain the features QR, QG, and QB of the three bands; for each pixel, statistically obtain the feature histograms HR, HG, and HB of the three bands in its neighborhood, and concatenate these histograms as the global feature histogram H of the corresponding pixel , so that the moving target detection of color photos can be realized; in addition, the feature histograms of the three bands of color images R, G, and B are connected in series, thereby reducing the dimensionality and sparsity of feature representation, which is conducive to reducing the storage of background models space.

Specifically, in step c, the establishment of the initial background model specifically includes: for each pixel, use the previous K frames of images to obtain K global feature histograms {H ₁ ,...,H _K } through step a and b extraction; Each global feature histogram is assigned a weight w _k , and

Step d of the present invention specifically includes: for each pixel, if the similarity between the global feature histogram H _c of the current frame image and any of the previous B background histograms corresponding to the background in the background model is greater than the threshold T _s , then the pixel is the background; otherwise, the pixel is the foreground, that is, the moving object.

The updating of the background model described in step e of the present invention specifically includes:

e1. For each pixel, use steps a and b to extract the current frame image to be detected to obtain its global feature histogram H _c ; use the histogram intersection to calculate the global feature histogram H _c of the current frame image and K histograms in the background model Similarity between graph H _i :

其中，N表示直方图bin的个数；i=1，2,......，K；

Among them, N represents the number of histogram bins; i=1, 2,..., K;

e2. Set the similarity measurement threshold as T _S , if the similarities between the global feature histogram H _c of the current frame image and the K histograms in {H ₁ ,…,H _K } are all lower than the threshold T _S , then The current histogram cannot match the background model, so replace the histogram with the smallest weight in the background model with the global feature histogram H _c of the current frame image; if the global feature histogram H _c of the current frame image is consistent with the background model {H ₁ ,...,H _K }, the similarity between some histograms in H K } is higher than the threshold T _S , then select the background model histogram H _m with the largest similarity as the best matching histogram, and the histogram H of the current frame image _c updates the best matching histogram H _m and its corresponding weight w _m , namely:

H _m =α _h H _c +(1-α _h )H _m ;

w _m =w _m +α _w (1-w _m );

Among them, α _h is the histogram learning factor, α _w is the weight learning factor; at the same time, the weight w _i of other histograms in the background model is also adjusted accordingly, namely:

w _i =(1-α _w )w _i , where i≠m;

e3. Reorder the K histograms in the background model according to the weight of each histogram from large to small, and select the first B histograms corresponding to the background; wherein, B<K.

In the present invention, the first K frames are used to initialize the background model, and the moving target is detected from the K+1 frame; The current frame updates the background model.

Compared with the prior art, the present invention draws lessons from the mature human visual perception system, and according to the overall characteristics of human visual perception (that is, the human visual perception system often recognizes the whole of the observed object before the recognition of the part, and this overall Cognition is not a simple summation of the individual cognitions that make up the observation object; each part of the individual in the observation object stimulates the visual perception system, and various stimuli interact, influence each other, and complement each other), using pulse-coupled neural networks to simulate human The eye visual nervous system, based on the impulse response of the neural network to perceive the color image, extracts the global feature information of the image, and builds a background model based on the global feature on this basis, so as to improve the robustness of moving target detection. In addition, the present invention draws on the overall characteristics of human visual perception and uses the pulse-coupled neural network to extract the global features of the image, which is conducive to suppressing the adverse effects of dynamic background disturbance on moving target detection, thereby improving the accuracy and F-measure value of moving target detection; And it is found through comparative experiments that, compared with the moving target detection method based on local features, the detection result of the method of the present invention has a better suppression effect on dynamic background disturbance. According to a large amount of data statistics, after removing the pre-processing and post-processing processes, compared with the existing moving target detection method in the same experimental environment, the detection accuracy of the existing moving target detection method is 78.3%, and the F-measure is 84.9%, while the detection accuracy rate of the present invention is 82.6%, F-measure is 86.7%, the accuracy rate has increased by 4.3%, and the F-measure has increased by 1.8%. In addition, according to the characteristics of the three band images of the color image R, G, and B, the present invention uses a pulse-coupled neural network to perceive the three band images of the color image R, G, and B respectively, and obtains the three bands after iter times. The features Q ^R , Q ^G , and Q ^B of the color image can be used to detect moving targets; the feature histograms of the three bands of the color image R, G, and B are connected in series, thereby reducing the dimensionality and sparsity of feature representation , which is beneficial to reduce the storage space of the background model.

Using the moving target detection method of the present invention and using the classic background subtraction method (GMM) based on Gaussian mixture model, the improved background subtraction method (IGMM) based on Gaussian mixture model and the background subtraction method (LBP) based on local binary pattern texture ) for moving target detection, the detection effect is shown in Table 1:

Table 1

The type of method used Accuracy F-measure The method of the invention 0.8260 0.8673 GMM 0.6578 0.7733 IGMM 0.7827 0.8485 LBP 0.6957 0.7823

The difficulty of the present invention is reflected in which method and method can be used to extract the information reflecting the global features of the image.

Description of drawings

Fig. 1 is a frame image in a group of video images;

Fig. 2 is the moving target region detected from Fig. 1 using the method of the present invention;

Fig. 3 is a flow chart of the method of the present invention.

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

Detailed ways

Embodiment 1 of the present invention: the moving target detection method based on pulse coupled neural network, as shown in Figure 3, comprises the following steps:

a. Utilize the pulse-coupled neural network to sense the video image sequence and extract the global features of the video image; wherein, a neuron of the pulse-coupled neural network corresponds to a pixel of the video image; for the neural network located at (i, j) in the pulse-coupled neural network After being affected by the external stimulus information S _ij and other neuron n-1 time impulse information {Y _kl } in the adjacent k×l neighborhood at time n, the feedback input F _ij , the linear connection input L _ij , the internal activity The term U _ij , the membrane potential dynamic threshold θ _ij , the pulse output Y _ij of the pulse generator, and the global feature Q _ij of the extracted video image are respectively:

F _ij (n) = S _ij ;

${\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>& \underset{\mathrm{<span\; class="notranslate">\; kl</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; kl</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; 0</span>}\\ \mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; otherwise</span>}\end{array}<span\; class="notranslate">\; ;</span>\right.$

U _ij (n)=F _ij (n)·(1+βL _ij (n));

${\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}}{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ;</span>$

${\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>& {\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; otherwise</span>}\end{array}\right.<span\; class="notranslate">\; ;</span>$

Q _ij (n)=Q _ij (n-1)+(iter-n)Y _ij (n);

Among them, iter is the total number of iterative perception of the pulse coupled neural network; S _ij is the pixel gray value of any band image perceived by the neuron at (i,j); α _θ is the decay time constant of the dynamic threshold θ _ij ; V _θ is the threshold amplification factor, which determines the degree of improvement of the threshold at the time of neuron firing, and plays an important role in regulating the firing cycle of neurons, so it usually takes a larger value; β is the connection strength, which determines the linear connection input The contribution of L _ij to the internal activity item U _ij , when β≠0, there is a coupling connection between the neurons of PCNN, and the ignition of a neuron will contribute to other neurons within its connection range; the connection strength is usually selected based on experience and let it be a constant;

b. According to the characteristics of the R, G, and B band images of the color image, establish a global feature histogram for each pixel: use the pulse-coupled neural network to perceive the R, G, and B band images of the color image, and iterate After iter times, the features Q ^R , Q ^G , and Q ^B of the three bands are obtained respectively; for each pixel, the feature histograms ^HR , H ^G , and H ^B of the three bands are statistically obtained in its neighborhood, and these histograms Concatenated as the global feature histogram H of the corresponding pixel;

c. For each pixel, use the global feature histogram corresponding to the previous K frames as the initial background model of the pixel; the establishment of the initial background model specifically includes: for each pixel, use the previous K frame images to extract and obtain through steps a and b K global feature histograms {H ₁ ,…,H _K }; assign a weight w _k to each global feature histogram, and

d. For each pixel, calculate the similarity between the global feature histogram of the current frame image and the corresponding global feature histogram in the background model, and detect whether the pixel is a moving target. Specifically, for each pixel, if the current frame If the similarity between the global feature histogram H _c of the image and any of the first B histograms corresponding to the background in the background model is greater than the threshold T _s , then the pixel is the background; otherwise, the pixel is the foreground, that is, the moving target;

e. For each pixel, update its background model using the global feature histogram of the current frame image; the updating of the background model specifically includes:

e1. For each pixel, use steps a and b to extract the current frame image to be detected to obtain its global feature histogram H _c ; use the histogram intersection to calculate the global feature histogram H _c of the current frame image and K histograms in the background model Similarity between graphs H _i (i=1, 2, ..., K)):

$\mathrm{Sim}(h_c,h_i)=\underset{\mathrm{no}=0}{\overset{N-1}{\mathrm{\&Sigma;}}}\min (h_{\mathrm{cn}},h_{\mathrm{in}});$ Among them, N represents the number of histogram bins;

e2. Set the similarity measurement threshold as T _S , if the similarities between the global feature histogram H _c of the current frame image and the K histograms in {H ₁ ,…,H _K } are all lower than the threshold T _S , then The current histogram cannot match the background model, so replace the histogram with the smallest weight in the background model with the global feature histogram H _c of the current frame image; if the global feature histogram H _c of the current frame image is consistent with the background model {H ₁ ,...,H _K }, the similarity between some histograms in H K } is higher than the threshold T _S , then select the background model histogram H _m with the largest similarity as the best matching histogram, and the histogram H of the current frame image _c updates the best matching histogram H _m and its corresponding weight w _m , namely:

H _m =α _h H _c +(1-α _h )H _m ;

w _m =w _m +α _w (1-w _m );

Among them, α _h is the histogram learning factor, α _w is the weight learning factor; at the same time, the weight w _i of other histograms in the background model is also adjusted accordingly, namely:

w _i =(1-α _w )w _i , where i≠m;

e3. After completing the update of the background model histogram and corresponding weights, reorder the K histograms in the background model according to the weight of each histogram from large to small, and select the first B histograms corresponding to the background.

The above methods can realize the detection of moving objects in color photos.

Embodiment 2: The moving target detection method based on pulse coupled neural network, as shown in Figure 3, comprises the following steps:

a. Utilize the pulse-coupled neural network to sense the video image sequence and extract the global features of the video image; wherein, a neuron in the pulse-coupled neural network corresponds to a pixel in the video image; for the pulse-coupled neural network located at (i, j) After the neuron is affected by the external stimulus information S _ij and the impulse information {Y _kl } of other neurons in the adjacent k×l neighborhood at time n-1 at time n, the feedback input F _ij , the linear connection input L _ij , The internal activity item U _ij , the membrane potential dynamic threshold θ _ij , the pulse output Y _ij of the pulse generator, and the global feature Q _ij of the extracted video image are respectively:

F _ij (n) = S _ij ;

${\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>& \underset{\mathrm{<span\; class="notranslate">\; kl</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; kl</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; 0</span>}\\ \mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; otherwise</span>}\end{array}<span\; class="notranslate">\; ;</span>\right.$

U _ij (n)=F _ij (n)·(1+βL _ij (n));

${\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}}{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ;</span>$

${\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>& {\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; otherwise</span>}\end{array}\right.<span\; class="notranslate">\; ;</span>$

Q _ij (n)=Q _ij (n-1)+(iter-n)Y _ij (n);

Among them, iter is the total number of iterative perception of the pulse coupled neural network; S _ij is the pixel gray value of the gray image perceived by the neuron at (i, j); α _θ is the decay time constant of the dynamic threshold θ _ij ; V _θ is the threshold amplification factor; β is the connection strength;

b. Establish a global feature histogram H for each pixel;

c. For each pixel, use the global feature histogram corresponding to the previous K frames as the initial background model of the pixel; wherein, the establishment of the initial background model specifically includes: for each pixel, use the previous K frame images to pass steps a and b Extract K global feature histograms {H ₁ ,…,H _K }; assign a weight w _k to each global feature histogram, and

d. For each pixel, calculate the similarity between the global feature histogram of the current frame image (starting from K+1 frame) and the corresponding global feature histogram in the background model, and detect whether the pixel is a moving target; if the current frame image The similarity between the global feature histogram Hc of the background model and the first B histogram corresponding to the background in the background model is greater than the threshold Ts, then the pixel is the background; otherwise, the pixel is the foreground, that is, the moving target;

e. For each pixel, update its background model using the global feature histogram of the current frame image; the updating of the background model specifically includes:

e1. For each pixel, the current frame image to be detected (starting from K+1 frame) is extracted using steps a and b to obtain its global feature histogram Hc; use the histogram intersection to calculate the global feature histogram Hc of the current frame image and Similarity between K histograms Hi in the background model:

其中，N表示直方图bin的个数；i=1，2,......，K；

Among them, N represents the number of histogram bins; i=1, 2,..., K;

e2. Set the similarity measurement threshold as T _S , if the similarities between the global feature histogram H _c of the current frame image and the K histograms in {H ₁ ,…,H _K } are all lower than the threshold T _S , then The current histogram cannot match the background model, so replace the histogram with the smallest weight in the background model with the global feature histogram H _c of the current frame image; if the global feature histogram H _c of the current frame image is consistent with the background model {H ₁ ,...,H _K }, the similarity between some histograms in H K } is higher than the threshold T _S , then select the background model histogram H _m with the largest similarity as the best matching histogram, and the histogram H of the current frame image _c updates the best matching histogram H _m and its corresponding weight w _m , namely:

H _m =α _h H _c +(1-α _h )H _m ;

w _m =w _m +α _w (1-w _m );

Among them, α _h is the histogram learning factor, α _w is the weight learning factor; at the same time, the weight w _i of other histograms in the background model is also adjusted accordingly, namely:

w _i =(1-α _w )w _i , where i≠m;

e3. Reorder the K histograms in the background model according to the weight of each histogram from large to small, and select the first B histograms corresponding to the background. Wherein, B<K, B can be set as a fixed value or adaptively changed according to the prior art.

Next, for each pixel, calculate the similarity between the global feature histogram of the K+2 frame image and the corresponding global feature histogram in the background model, and judge whether the pixel is a moving target; then use the global feature histogram of the K+2 frame image The feature histogram updates the background model, and so on, calculates the K+3 frame image.....

The above method can realize the moving target detection on the gray scale image.

Experiment example: Figure 1 is a frame of images in a group of video images (in Figure 1, the character's coat 1 is purple, shirt 2 is green, hair 3 is black and yellow, branch 4 on the left is displayed in green, branch 5 on the right It is displayed in black, the wall brick 6 of the building is khaki, the wall brick 6 has sunlight 7 shining through the branches, and the sky 8 is blue), the monitoring background of this group of video images is dynamic because it contains tree branches swaying in the wind Changeable, which brings difficulties to moving target detection.

Use method of the present invention to carry out moving target detection to Fig. 1, specifically comprise the following steps:

(1-1) Use the pulse-coupled neural network to extract the global features of the image: for a group of 120×160 color video images like Figure 1, a neural network composed of 120×160 pulse-coupled neurons can be used for perception; for the neural network The neuron located at (i, j) in the middle is affected by the external stimulus information S _ij and the impulse information {Y _kl } of other neurons in the adjacent k×l neighborhood at time n-1 at time n, its feedback input F _ij , the linear connection input L _ij , the internal activity item U _ij , the membrane potential dynamic threshold θ _ij , the pulse output Y _ij of the pulse generator, and the extracted feature Q _ij of the present invention can be described as:

F _ij (n) = S _ij (1)

${\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>& \underset{\mathrm{<span\; class="notranslate">\; kl</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; kl</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; 0</span>}\\ \mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; otherwise</span>}\end{array}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>\right.$

U _ij (n)＝F _ij (n)·(1+βL _ij (n)) (3)

${\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>& {\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; otherwise</span>}\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}}{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

Q _ij (n)=Q _ij (n-1)+(iter-n)Y _ij (n) (6)

Among them, iter=16 represents the total number of iterative perception of the pulse-coupled neural network; S _ij is set to the pixel gray value of any band image perceived by the neuron located at (i,j); α _θ =0.5 represents the dynamic threshold θ The decay time constant of _ij ; V _θ = 100 is the threshold amplification factor, which determines the degree of improvement of the threshold at the moment of neuron firing, and plays an important role in regulating the firing cycle of neurons, so it usually takes a larger value; β = 0.2 is the connection strength , which determines the contribution of the linear connection input L _ij to the internal activity item U _ij , when β≠0, there is a coupling connection between the neurons of the PCNN, and the ignition of a neuron will contribute to other neurons within its connection range , the connection strength is usually chosen empirically and set as a constant;

(1-2) Establish the global feature histogram of each pixel: Use the pulse-coupled neural network to perceive the R, G, and B band images of the color image separately, and obtain the feature Q of the three bands after iter times ^R , Q ^G , Q ^B , and for each pixel, statistically obtain the feature histograms ^HR , H ^G , H ^B of the three bands in its neighborhood, and concatenate these histograms as the global feature histogram of the corresponding pixel H. The global feature histogram in the statistical area is used as the background model;

(1-3) Background model initialization: For each pixel, use steps (1-1) and (1-2) to extract global feature histograms using the previous K=4 frame images, and obtain K global feature histograms {H ₁ ,...,H _K }, and assign a weight w _k =1K to each histogram;

(1-4) Moving target detection: For each pixel, if the similarity between the global feature histogram H _c of the current frame image and any histogram corresponding to the background in the background model is greater than the threshold T _S , then the The pixel is the background; otherwise, the pixel is the foreground, that is, the moving target;

(1-5) Background model update: For each pixel, use steps (1-1) and (1-2) to extract the global feature histogram H _c of the current frame image to be detected; use the histogram intersection to calculate the current frame histogram Similarity between graph H _c and K histograms H _i (i=1,...,K) in the background model:

$\mathrm{<span\; class="notranslate">\; Sim</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; c</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; min</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; cn</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; in</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 7</span>}<span\; class="notranslate">\; )</span>$

Among them, N represents the number of histogram bins; if the similarity measure threshold is T _S =0.5, if the global feature histogram H _c of the current frame image and the K histograms in {H ₁ ,…,H _K } The similarities of all are lower than the threshold T _S , it is considered that the current histogram cannot match the background model, and the current histogram H _c is used to replace the histogram with the smallest weight in the background model; if the global feature histogram H of the current frame image The similarity between _c and some histograms in the background model {H ₁ ,…,H _K } is higher than the threshold T _S , then select the background model histogram with the greatest similarity (such as H _m ) as the best matching histogram , and update the best matching histogram H _m and its corresponding weight w _m by the current histogram H _c , namely

H _m =α _h H _c +(1-α _h )H _m (8)

w _m ＝w _m +α _w (1-w _m ) (9)

Among them, α _h =0.01 is the histogram learning factor, and α _w =0.01 is the weight learning factor. At the same time, the weight w _i (where i≠m) of other histograms in the background model should be adjusted accordingly, namely:

w _i =(1-α _w )w _i (10)

After completing the updating of the background model histograms and corresponding weights, the K histograms in the background model are reordered according to the weights of the histograms, and the first B=2 histograms are selected to correspond to the background.

Figure 2 is the moving target detection result obtained through the above steps without any post-processing method. Compared with other existing detection methods, after using the detection method of the present invention, the detection results contain fewer false targets, and the moving target The detection accuracy is higher.

Claims (6)

1. the moving target detection method based on Pulse Coupled Neural Network, is characterized in that, comprises the following steps:

A. utilize Pulse Coupled Neural Network perception sequence of video images, extract the global characteristics of video image;

B. set up the global characteristics histogram of each pixel;

C. for each pixel, the global characteristics histogram that before utilizing, K frame is corresponding is as the initial back-ground model of this pixel;

D. for each pixel, the histogrammic similarity of global characteristics accordingly in the global characteristics histogram that calculates current frame image and background model, whether detect this pixel is moving-target;

E. for each pixel, utilize the global characteristics histogram of current frame image to upgrade its background model.

2. the moving target detection method based on Pulse Coupled Neural Network according to claim 1, is characterized in that, step a specifically comprises: a pixel in the corresponding video image of a neuron in Pulse Coupled Neural Network; For the neuron that is positioned at (i, j) in Pulse Coupled Neural Network, it is subject to environmental stimuli information S at moment n _ijwith other neuron n-1 moment pulse information { Y in adjacent k * l neighborhood _klimpact after, feed back input F _ij, the linear input L that connects _ij, internal activity item U _ij, film potential dynamic threshold θ _ij, pulse producer pulse output Y _ijand the global characteristics Q of the video image extracting _ijbe respectively:

Fi _j(n)＝Si _j；

$L_{\mathrm{ij}}\left(n\right)=\left\{\begin{array}{cc}1,& \underset{\mathrm{kl}}{\mathrm{\&Sigma;}}{Y}_{\mathrm{kl}}(n-1)>0\\ 0,& \mathrm{otherwise}\end{array};\right.$

U _ij(n)＝F _ij(n)·(1+βL _ij(n))；

${\mathrm{\&theta;}}_{\mathrm{ij}}\left(n\right)=e^{-{\mathrm{\&alpha;}}_{\mathrm{\&theta;}}}{\mathrm{\&theta;}}_{\mathrm{ij}}(n-1)+V_{\mathrm{\&theta;}}{Y}_{\mathrm{ij}}\left(n\right);$

$Y_{\mathrm{ij}}\left(n\right)=\left\{\begin{array}{cc}1,& U_{\mathrm{ij}}\left(n\right)>{\mathrm{\&theta;}}_{\mathrm{ij}}(n-1)\\ 0,& \mathrm{otherwise}\end{array}\right.;$

Q _ij(n)＝Q _ij(n-1)+(iter-n)Y _ij(n)；

Wherein, iter is the total degree of Pulse Coupled Neural Network iteration perception; S _ijfor being positioned at the grey scale pixel value of arbitrary band image of the neuron institute perception of (i, j); α _θfor dynamic threshold θ _ijdamping time constant; V _θfor threshold value amplification coefficient; β is strength of joint.

3. the moving target detection method based on Pulse Coupled Neural Network according to claim 2, it is characterized in that, step b comprises: according to the feature of coloured image R, G, tri-band images of B, use Pulse Coupled Neural Network to carry out respectively perception to the R of coloured image, G, tri-band images of B, after iteration iter time, obtain respectively the feature Q of three wave bands ^r, Q ^g, Q ^b; For each pixel, in its neighborhood, statistics obtains the feature histogram H of three wave bands ^r, H ^g, H ^b, the global characteristics histogram H using these histogram series connection as respective pixel.

4. according to the arbitrary described moving target detection method based on Pulse Coupled Neural Network of claim 1～3, it is characterized in that, in step c, the foundation of initial back-ground model specifically comprises: for each pixel, before utilizing, K two field picture is extracted and obtained K global characteristics histogram { H by step a and b ₁..., H _k; To each global characteristics histogram, give a weight w _k, and

5. the moving target detection method based on Pulse Coupled Neural Network according to claim 4, is characterized in that, steps d specifically comprises: for each pixel, if the global characteristics histogram H of current frame image _cbe greater than threshold value T with the individual arbitrary histogrammic similarity corresponding to background of front B in background model _s, this pixel is background; Otherwise this pixel is prospect, i.e. moving-target.

6. the moving target detection method based on Pulse Coupled Neural Network according to claim 5, is characterized in that, background model is upgraded specifically and being comprised described in step e:

E1. for each pixel, to current frame image to be detected, use step a and b to extract and obtain its global characteristics histogram H _c; Utilize histogram intersection to calculate the global characteristics histogram H of current frame image _cwith K in background model histogram H _ibetween similarity:

wherein, N represents the number of histogram bin; I=1,2 ..., K;

E2. establishing similarity measurement threshold value is T _sif, the global characteristics histogram H of current frame image _cwith { H ₁..., H _kin similarity between K histogram all lower than threshold value T _s, current histogram cannot mate with background model, now uses the global characteristics histogram H of current frame image _creplace the histogram of weight minimum in background model; If the global characteristics histogram H of current frame image _cwith background model { H ₁..., H _kin similarity between certain several histogram higher than threshold value T _s, select the background model histogram H of similarity maximum _mas optimum matching histogram, and by the histogram H of current frame image _cto optimum matching histogram H _mand corresponding weight w _mupgrade, that is:

H _m＝α _hH _c+(1-α _h)H _m；

w _m＝w _m+α _w(1-w _m)；

Wherein, α _hfor the histogram study factor, α _wfor the weight study factor; Simultaneously to other histogrammic weight w in background model _ialso adjust accordingly, that is:

W _i=(1-α _w) w _i, i ≠ m wherein;

E3. descending to the K in background model histogram rearrangement according to each histogrammic weight, and select front B histogram corresponding to background.

Images