WO2015161490A1

WO2015161490A1 - Target motion detection method for water surface polarization imaging based on compound eyes simulation

Info

Publication number: WO2015161490A1
Application number: PCT/CN2014/076146
Authority: WO
Inventors: 陈哲; 徐立中; 王鑫; 石爱业; 王慧斌; 严锡君; 范超; 孔成
Original assignee: 陈哲; 徐立中; 王鑫; 石爱业; 王慧斌; 严锡君; 范超; 孔成
Priority date: 2014-04-24
Filing date: 2014-04-24
Publication date: 2015-10-29

Abstract

A target motion detection method for water surface polarization imaging based on compound eyes simulation is divided into three steps of polarization image collection calculation, bionic target detection and bionic motion target matching. A bionics method of a compound eyes simulation vision mechanism is adopted to realize the detection and tracking of a water surface target, and in the process of calculation, a scenario and target information are converted to a "compressive sensing" feature represented in the form of an impulse sequence. Subsequent target detection and target matching and tracking are both conducted based on the impulse sequence, and a response time sequence of an impulse and an impulse sequence mode are utilized to detect, match and track multi-targets in a scenario, so as to finally realize the detection of the target motion and estimation of motion vectors thereof. The method can be stably and reliably applied to water surface target motion detection under a complex water surface optical environment, and the computational efficiency is relatively high.

Description

Target motion detection method for water surface polarization imaging based on compound eye simulation

The invention relates to a target motion detecting method, in particular to a target motion detecting method based on compound eye simulation for water surface polarization imaging.

Background technique

Optical imaging detection and motion parameter estimation for surface targets such as oceans, lakes, and rivers have been widely used in many fields. Optical imaging studies currently used for surface target detection focus on light intensity or spectral information. However, due to the changing natural and climatic conditions and the variable optical properties of the target, it is not considered that the surface target detection based on the optical imaging method is universal and optimal. Therefore, some studies have gradually turned to other optical properties, the most typical of which is optical polarization. Since the researchers first introduced optical polarization characteristics into oil level detection in 1992 and proved that optical polarization is beneficial to improve the accuracy of surface target detection, the optical polarization difference is gradually regarded as a new water surface target detection method.

Due to the extremely complex water surface optical environment, even with a relatively sophisticated front-end water optical information acquisition unit, the information obtained is extremely unstable and contains a large amount of noise. Therefore, in order to finalize the inspection, the system must also rely on the support of the back-end detection algorithm. The algorithms currently used for surface target detection mainly include background modeling and non-background modeling. Among them, the target detection method based on background modeling has higher detection rate and higher anti-noise ability, so its robustness is better. However, the high complexity of the background model calculation and the variability of the natural background seriously affect the generalization of the method.

In contrast, the target detection algorithm for non-background modeling does not require a cumbersome modeling process. Target detection based on the information reflected by the surface image itself sometimes results in greater success, but without the support of the background model. The distinction between background information and target information is the key to this approach. In response to this critical point, the present invention solves the biomimetic solution to the hydrophilic insect-plutonium.

Summary of the invention

It is an object of the present invention to provide a target motion detection method for water surface polarization imaging based on compound eye simulation, which solves the technical problem of detecting and tracking targets in a complex water surface optical environment.

In order to solve the above technical problem, the present invention provides a target motion detection method for water surface polarization imaging based on compound eye simulation, which includes the following steps:

Step 1: Obtain a three-channel polarization image of the water surface scene through three sets of polarized water surface imaging systems, The three-pass polarization image is registered, and the three-pass polarization image is calculated by the Stokes model to obtain a polarization degree image of the optical information between the water surface target and the water surface background.

Step 2: Overlapping the polarization degree image by computer simulation of the compound vision to obtain the polarization degree information of the polarization degree image, and then constructing a large scene and a small scene channel by using the simulated vision, the large scene and the small scene channel respectively And compressing the water surface background and the water surface moving target in the polarization degree image according to the polarization degree information to obtain a sensitivity characteristic of the water surface moving target relative to the water surface background.

The method for superimposing the polarization degree image by computer simulation of the compound vision includes: constructing a virtual complex eye group through a plurality of partial image windows, wherein the virtual eye group is suitable for simulating the eye through sliding scanning In the structure, the scanning degree sampling is performed on the polarization degree image by using a plurality of small eyes to obtain the polarization degree information of the image. The sliding scanning mode includes: the small eyelets distributed around the virtual eye group in the small eye group respectively The small eye at the center position is slid to achieve a convergent sliding scan, or the small eyes that have been concentrated at the center position are respectively slid around to achieve a split sliding scan.

Step 3: Using a pseudo cell model to retrieve a compressed image of a large scene and a small scene channel, and transforming the compressed image through a threshold filter to generate a continuous pulse sequence with a sensitive feature of the water moving target, and according to the sensitivity feature , to achieve target motion detection in the water surface scene.

Further, the method for calculating the polarization image by using the polarization image registration technology and the Stokes model in the first step includes: using a feature point matching algorithm based on the SURF corner point to three-channel polarization image of the water surface scene taken at the same time After pixel-level registration, the Stokes equation is used to calculate the degree of polarization information corresponding to each pixel in the image to obtain a frame of polarization image.

Further, in the step 3, the method for detecting target motion in the water surface scene according to the sensitivity characteristic of the continuous pulse sequence comprises: analyzing the response and response timing of the continuous pulse sequence by using the sensitivity characteristics of the continuous pulse sequence; The target pulse sequence of the degree image is matched, that is, the pulse sequence generated by the current frame polarization degree image is combined with the pulse sequence of the next frame polarization degree image, and the response timings of the same pulse are corresponding in the pulse sequence features of different frames. The same target achieves the matching of the sensitive features of the water moving target in the continuous pulse sequence, and then completes the motion detection of the water moving target.

The above technical solution of the present invention has the following advantages compared with the prior art: (1) The technical solution of the present invention has the characteristics of strong anti-interference ability, polarization imaging of a specific spectral segment and polarization image fusion, without background modeling Under the premise of any prior knowledge, it can effectively suppress the complex surface, underwater and atmospheric optical noise in the scene, enhance the contrast between the target and the background, and improve the motion vector estimation of the water flow tracer. The accuracy. In addition, this polarization information is also useful for extracting the invariant features of the optical surface scene and suppressing dynamic changes in the scene, such as illumination changes. For residual noise, the special characteristics of the impulse response characteristics are utilized in the image processing stage, and the noise information and useful information are effectively distinguished to further implement noise filtering; (2) The target motion detection accuracy is high. In the target detection method, the bionic technique of the compound eye vision is used, and the extracted pulse information can highlight the features such as texture and edge, and the sensitivity to the target is high. Compared with the common target detection method, the method is more suitable for multi-target. A water surface scene with uneven spatial and temporal distribution and greater environmental impact. In the motion detection, the multi-frame based pulse sequence is based on the pulse response timing to achieve the target matching. Compared with the method based on statistical decision, the matching accuracy of the method is higher and the matching speed is faster; (3) The complexity of the algorithm is significantly reduced. In the process of bionic-based water surface motion detection method, complex, high-dimensional scene and target information are compressed into one-dimensional pulse sequences without losing key information. In the process of detection and matching, only the mode, amplitude, frequency, timing, etc. of the one-dimensional pulse sequence need to be analyzed and calculated to achieve target detection and matching. Compared with the general target detection and tracking method for the processing of high-dimensional and complex information, the complexity of the algorithm is significantly reduced, and the computational efficiency is significantly improved. Meet the system's requirements for real-time performance.

DRAWINGS

In order to make the content of the present invention easier to understand, the present invention will be further described in detail below with reference to the accompanying drawings

1 is a flow chart of a target motion detecting method according to an embodiment of the present invention;

2 is a flowchart of a method for detecting a moving target of a compound eye in a pseudo eye; FIG. 3 is a process diagram of generating a feature of a SURF corner feature in the embodiment of the present invention;

4 is a schematic diagram of a principle of a nearest neighbor algorithm in an embodiment of the present invention;

FIG. 5 is a schematic view showing the sliding of the virtual complex eye group of the present invention.

Specific form

The present invention will be described in detail below with reference to the accompanying drawings and embodiments:

Example

As shown in Fig. 1 and Fig. 2, a target motion detection method based on compound eye simulation for water surface polarization imaging, three sets of polarized water surface imaging systems realize polarization optical imaging of three-channel specific spectral segments, and use image processing dedicated DSP chip as processing The central processing unit (host computer) is a general-purpose PC or a ruggedized terminal device. According to the process of information acquisition and processing, the workflow of the polarization imaging system can be divided into four steps: polarization image acquisition, polarization image registration and polarization image fusion. Polarized image acquisition for a particular spectral segment can be taken in front of the camera or Install the filter before the photosensitive element (determined according to the requirements of the imaging equipment and the production process). The three sets of polarized surface imaging systems use three sets of CMOS image sensors, and the CMOS image sensors are affixed with 0 °, 45 ° and 90 ° polarizers and spectral filters. It enables infrared polarization imaging to enhance the optical contrast between the water flow tracer and the water surface.

After the polarization image is acquired, three images are simultaneously input into the DSP chip through three information channels. The image registration procedure has been preloaded in the DSP chip, and the present invention employs the improved SURF algorithm for image registration.

Applying the SURF algorithm to the image registration process can be divided into the following steps: Calculate the gradients ^fx and ^fy of each pixel in the horizontal and vertical directions in the image, and the product of the two, to obtain the Hessian matrix H:

H :

(1) Perform Gaussian filtering on the image to obtain the filtered H. The discrete two-dimensional zero-mean Gaussian function is:

(x ² + y ² )

H gauss exp (- ) * Η

2σ ²

Where ( ^χ , ^y ) is the coordinates of the image pixel and ^σ is the variance coefficient of the Gaussian function.

(2) Calculate the interest value ^p u" of each pixel on the image _{: the} interest value determines the corner point, and the corner point may also be called a point of interest.

: [ f, X f - ( fxf _y r] -k[ f, + f ] where modulation parameter ^k (specifically, the difference between the modulation angle point in the X and y directions, used to describe the point and the adjacent pixel The difference is typically 0. 0 0. 06.

(3) Select a local extremum point (as a feature point). In the SURF algorithm, a feature point is a pixel point corresponding to a maximum interest value in a local range.

(4) Select an appropriate threshold T (the typical choice of the present invention is Τ=40 (Γ700), and select a certain number of corner points (the typical choice of the present invention is 15CT200 corner points). The threshold value is when the interest value R of the point is greater than a certain value. Select the point for the value of the value or discard the point.

The specific steps include:

First, construct a direction for the extracted corner points, and take the direction of the largest pixel point in the gradient direction of each pixel in the neighborhood of the feature point as the direction of the SURF corner point. Here we take the neighborhood size to be 3 X 3 .

Establishing a coordinate system as a feature point according to the configured SURF feature point direction; (1) Filtering the image to generate the scale space, and obtaining images of different scales, taking the SURF feature point 16 X 16 pixel neighborhood, and then dividing the field into 4 identical sub-regions, calculating the gradient direction of each sub-region, uniform Divided into 8 directions; 16*16 is empirical, divided into 4 regions, each region is 8*8 and the following discussion is also consistent

(2) Sort the eight direction gradients to obtain a 128-dimensional feature vector, which is the SURF feature description vector.

The 8 X 8 pixel feature point neighborhood is taken as an example to illustrate the generation process of the SURF feature descriptor, as shown in Figure 3: The neighborhood of the feature point pixel.

Through the above steps, the descriptor of the SURF corner point can be obtained by calculating the similarity between the left and right image feature point description vectors to determine whether the two points are matching points. The nearest neighbor (Nearest Neighbor) search algorithm is used to perform matching operations on SURF corner points by exhaustive search, where the nearest neighbor point is the feature point with the shortest Euclidean distance from the sample feature point, and the next nearest neighbor point is Refers to the feature point of Euclidean distance which is slightly longer than the nearest neighbor distance, and takes the ratio ^{d of the} Euclidean distance of the two feature points as a similarity measure. This ratio is also called NNDR (Nearest Neighbor Distance Ratio). ), As shown in Figure 4:

Where point p is any point in space, point q is its nearest neighbor, r is its next nearest neighbor, and the ratio of nearest neighbor to next nearest Euclidean distance is:

Among them, D _nearest is the nearest neighbor Euclidean distance, Dh. ― _nearest is the nearest neighbor Euclidean distance.

The matching steps of the SURF feature points are:

(1) The two images are freely defined in the matching process between the standard image and the image to be matched, respectively. Among them, the matching is the operation of two images, the standard image is one of the two images, and the standard image and the image to be matched are extracted from the respective SURF feature points.

(2) taking the feature points in the standard image in turn, is the corner point, searching for the nearest neighbor and the nearest neighbor of the feature point in another image, and calculating the ratio of the two;

(3) Compare the ratio between the two and the set threshold. If the ratio is less than the set threshold, it indicates that the feature point is the same point as the feature point in the image to be matched, and the match is successful. Otherwise search again.

Input the exact polarization degree image into the polarization degree information calculation fusion operation module, through Stokes The equation calculates the degree of polarization information. According to the Stokes equation ^{1 = a Sin[2} (^- ") ⁺ 3⁄4 ^{] + b} and

The target grayscale image in the «= 0°, 45°, 90° direction can be solved for the corresponding pixel parameters _a , _b , and then the degree of polarization = ^. Finally, the polarization degree information is normalized and can be characterized as a grayscale feature map, that is, a polarization degree image.

Constructing a virtual small eye group with five partial windows (3 X 3 or 5 X 5 ) bundled in an overlapping manner, the small eye group slidingly scanning and scanning several small eyes in the compound eye structure to overlap and sample the polarization degree image, and read Polarization information. The degree of polarization in the degree of polarization image is represented by a normalized pixel intensity value. Electrophysiological studies, the system function between the small eye group input and output can be described as a Gaussian function. The present invention does not consider the dynamic change of the response in the field of view, and can be simplified into a step function, ie

... , , , , ,

r (x) = E(AI (x) -thre)

Where X is the pixel position in the image, I(x) is the intensity of the point, thre is the response threshold, and Ε(·) is the step function. Through the above description, the input, output and system response functions in the small-eye system can be expressed.

So far, the polarization degree image information is converted into a single pulse discharge characteristic. Subsequently, using the imitation "Large Scene (LF)" and "Small Scene (SF)" systems to compress and characterize the scene, using the optimized "scheduling" model of the "pool cell" to highlight features such as texture and edges to form motion Sensitive pulse sequence characteristics of the target. In turn, two visual mechanisms are simulated, namely, visual 分 visual shunt, nonlinear adaptive suppression, and pool cell scheduling processing.

Specific methods include:

FIG. 5 shows a split sliding scan, and the movement of the convergent sliding scan is opposite to that shown in FIG. 5, specifically constructing a virtual image by overlapping five partial image windows (the window size is 3 X 3 or 5 X 5 ). The complex eye group (corresponding to the change of J), the virtual small eye group is scanned and scanned to simulate overlapping images of the fusion polarization image by a plurality of small eyes in the compound eye structure, and the polarization degree information is read.

Subsequently, using the large scene (LF) and small scene (SF) visual perception systems in the pseudo vision to compress and characterize the surface scene, a large scene (LF) and small scene (SF) channel is formed.

In the insect compound eye system, there are two sets of parallel information integration system large scene integration model and small scene integration model. These two integration models integrate the primary motion perception signals obtained by retinal cells in different ways. The large scene integration model mainly produces a strong response to the slowly changing background features in the scene; while the small scene integration model responds to the target of high-speed motion in the scene, Estimate the direction of motion. In the present invention, large scenes suppress complex background features; while small scenes have extremely high sensitivity to target features. ;

First calculate the correlation between the (i, ^j )th pixel and the adjacent position pixel, that is, calculate the ⁽ⁱ , j)th pixel and the surrounding four pixels in the top, bottom, left and right respectively. Gray correlation coefficient ^Vu(i , ^j ) in four directions _; v _d (i,j). V _r (i,j)

V _u (i,j) = I(i,j)xI(i + l,j)

(1)

V _d (i,j) = I(i,j)xI(il,j)

X(i,j) = I(i,j)xI(i,j-l)

V _r (i,j) = I(i,j)xI(i,j+l)

Similarly, the correlation coefficients of the other three directions can be obtained. Then integrate the formula according to the big scene:

(2)

(3)

Integrate the correlation outputs in the corresponding two directions. Equation (1) is for the local region of size ^M and ^N , and calculates the integration result of features in the horizontal direction. Equation (2) corresponds to the integration result in the vertical direction and is the shunt suppression coefficient. The denominator in formula (2) and formula (3) is the sum of the features in the local region, where n = 3 is used to simulate the saturation nonlinearity of the cell, which can enhance the weaker correlation signal, and the correlation is stronger. The signal is suppressed. ^{q = 5} simulates the nonlinear filtering function of cells. The large scene integration process is a regularization process on the features in the local area, eliminating the suppression of the background by the features. Since n = ³ and -1<_0<1, (D=l, r, u, d), |V _D | ⁿ <V _D , therefore, third-order nonlinear processing can be used. The background feature with a small amplitude caused by noise is suppressed, and when ^V D ⁽ⁱ ' ^{j) > 1} , l ^V D| ^{n > V} D , indicating a signal with a large amplitude abnormality for the gray feature, large scene Integration will enhance it. (is enhanced because the gray level changes greatly)

The same small scene integration model formula is:

X[AV ⁺ (i,j)-BV-(i,j)]

(4) The integration model of the small scene realizes the enhancement of the target feature. Where A and B are simulated nerve fiber integration parameters, and ^{v (t} ) is the positive and negative channels of the bionic target detector, and the brightness increasing signals representing different polarities are distinguished. Taking the horizontal direction as an example, Y ^+(t ) represents The brightness from left to right increases, and ^{v (t} ) represents the positive and negative channels of the detector from right to left. ( ¹ , 』') and ( ¹ , 』) respectively simulate the response of the left and right sides of the brain.

The LF and SF channels are scheduled by the cell model, and the polarization degree information obtained by the above virtual small eye group is transformed into a pulse sequence by threshold filter simulation to form a sensitivity feature of the water surface moving target, which is used as a motion detection for different targets in the water surface scene. Basis.

The information obtained by the small eye is divided into the LF and SF neurons according to the topology of the compound eye. The topology of the neurons is consistent with the topology of the small eye, so the spatial information corresponding to the input of the small eye can be completely preserved. Among them, LF neurons are sensitive to targets in a wide range; while SF is sensitive to changes in features in a small local range, and the dynamic adjustment of the target size is achieved by local central side suppression mechanism, while rapidly polarizing And the adaptive processing mechanism of slow depolarization can effectively suppress the background texture features of local regions. The abrupt signal with low frequency and large amplitude is enhanced; and the texture signal with high frequency and low variation is adaptively suppressed. Taking the horizontal direction as an example (the same in the vertical direction), for the input signal ^Qn(i ,j), the nonlinear adaptive mechanism is simulated as -d/dAs{on(i, j)} = (οη(ητ,η)- Ηη(ϊ, j))/f, where on(i, j) is the (i' j) pixel suppressed at the center side The characteristic intensity of the on channel in the mechanism is the Euclidean distance between the an(i,j) and αη η) pixels. ζ" is the response attenuation (enhancement) coefficient. If the signal strength at the ⁽ⁱ , ^j ) position is lower than the intensity ^Qn(I ^ ^η ) of the surrounding area signal, then =" = 10, achieving a slow decay of contrast; ^{= i} , a fast increase in contrast.

The response information of LF and SF is input into the bionic pool cell model, and interaction is realized under its scheduling.

[n] = e-^U^ [n -1] +V _F ∑ [n -l] + For LF input and feedback mechanism: w

U [n] = [ _n _l] +V _{L Wl} ,j, _k , [n -1]

Input and feedback mechanism for SF: ' w where, ^[Π] is the response of the LF neuron (i, J) at time n is corresponding to the response of the input ^Qn (i, j) input to LF at that time, For the LF neuron (i, j) (is a concept), the input at time n corresponds to the response of the input to the SF, which is the input to the SF, for the small-eye input, "for the attenuation coefficient, ^w is Integrate the weights, ^Vp , ^Vl are the gains of the input. Then the final integrated pulse output is -

Among them, ^τ is the connection strength between LF neurons and SF neurons under pool cell scheduling.

So far, the polarization degree image information is converted into a pulse sequence feature. Where the pulse discharge area represents the target area -

0 = ατ _β (3⁄4 [η] = 1) where ⁰ is the target area and arg is the inverse operator.

The timing of the pulses distinguishes the categories of the targets, enabling the detection and classification of multiple targets:

€= ατ _β (0 _Μ [η] = 1) where ^C is the target class, arg is the inverse operator, and n is the timing of the impulse response.

Finally, based on the pulse sequence, the information processing mode of the central medullary neuron is simulated. In the pulse sequence represented by the multi-frame image, it is considered that the detection targets corresponding to the same sequence in different pulse sequences are the same target, thereby simply implementing Match of goals. According to the pixel difference between t, t _i consecutive frames, the motion vector is estimated to realize the detection of the target motion:

Where MV is the target motion vector, which is the position of the target at ¹ time, ^ is the target at time ¹ The location, ^C is the target category, that is, the target category.

This completes the motion detection of the surface target.

It is apparent that the above-described embodiments are merely illustrative of the invention and are not intended to limit the embodiments of the invention. Other variations or modifications of the various forms may be made by those skilled in the art in light of the above description. There is no need and no way to exhaust all of the implementations. It is to be understood that the obvious changes or modifications which come within the spirit of the invention are still within the scope of the invention.

Claims

claims

1. A target motion detection method for water surface polarization imaging based on compound eye simulation, which is characterized by including - Step 1: Obtain three-channel polarization images of the water surface scene through three sets of polarization water surface imaging systems, and register the three-channel polarization images , and calculate the three-channel polarization image through the Stokes model to obtain a polarization image that contrasts and fuses the optical information between the water surface target and the water surface background;

Step 2: Perform overlapping sampling of the polarization degree image through computer simulation of dragonfly compound eye vision to obtain the polarization degree information of the polarization degree image, and then use simulated dragonfly vision to construct large scene and small scene channels. The large scene and small scene channels are respectively Image compression is performed on the water surface background and the water surface moving target in the polarization degree image according to the polarization degree information to obtain the sensitivity characteristics of the water surface moving target relative to the water surface background; wherein, the polarization degree image is processed by computer simulation of dragonfly compound eye vision. The method of overlapping sampling includes: constructing a virtual dragonfly compound eye ommatidium group through several local image windows. The virtual dragonfly compound eye ommatidium group is suitable for simulating the dragonfly compound eye structure through sliding scanning, using several ommatidia to slide the polarization image. Scan and sample to obtain the polarization information of the image; wherein, the sliding scanning method includes: each ommatidium distributed in the virtual dragonfly compound eye ommatidium group slides toward the ommatidium at the center to achieve convergent sliding Scan, or each ommatidium that has been gathered at the center slides around to achieve diffuse sliding scanning;

Step 3: Use the imitation pool cell model to retrieve the compressed images of large scene and small scene channels, convert the compressed image through a threshold filter to generate a continuous pulse sequence with sensitivity characteristics of water surface moving targets, and based on the sensitivity characteristics , to achieve target motion detection in water scenes.

2. The target motion detection method according to claim 1, characterized in that: the method of calculating the polarization image using polarization image registration technology and the Stokes model in step one includes: using feature points based on SURF corner points The matching algorithm performs pixel-level registration of the three-channel polarization images of the water scene taken at the same time, and then uses the Stokes equation to calculate the polarization information corresponding to each pixel in the image to obtain a frame of polarization image.

3. The target motion detection method according to claim 2, characterized in that: in step three, the method for detecting target motion in the water surface scene based on the sensitivity characteristics of the continuous pulse sequence includes: using the sensitivity of the continuous pulse sequence. Features, analyze the response and response timing of the continuous pulse sequence; match the target pulse sequence of the multi-frame polarization image, that is, merge the pulse sequence generated by the current frame of polarization image with the pulse sequence of the next frame of polarization image. , in the pulse sequence characteristics of different frames, the response timing of the same pulse corresponds to the same target, realizing the sensitivity characteristics of water surface moving targets in continuous pulse sequences. matching, and then complete the motion detection of the water surface moving target.