CN117495918A - River water surface optical flow estimation method based on illumination self-adaptive ORB operator - Google Patents

Abstract

The invention discloses a method for estimating optical flow on the river surface based on the illumination adaptive ORB operator. The adaptive threshold ORB operator is used to extract feature points. The illumination adaptive ORB threshold is passed through the global adaptive threshold based on KSW entropy and the method based on KSW entropy. The local adaptive thresholds of image contrast are obtained by multiplying the weight coefficients respectively. The method first inputs two adjacent video frame RGB images, and then intercepts the ROI and constructs an image pyramid after grayscale processing. Then the illumination adaptive ORB threshold is calculated layer by layer and pixel by pixel in the first frame of the image pyramid, and the ORB feature points are determined and extracted using this threshold according to the FAST rapid optimization plan. Finally, the Harris scoring algorithm is used to sort the feature points by quality, and a certain number of high-quality feature points are selected. The sparse LK optical flow method is used to calculate the coordinate values of the feature points moving to the second frame of the image, and finally the distance between the two adjacent feature points is calculated. Optical flow motion vectors on frame images. The invention is suitable for observing the flow field of the river surface under complex and changeable illumination conditions.

Description

Optical flow estimation method on river water surface based on illumination adaptive ORB operator

Technical field

The invention relates to the technical field of image method flow field measurement, and in particular to a method for estimating optical flow on the river surface based on illumination adaptive ORB operator.

Background technique

Prototype observations of river flow fields provide important basic data for monitoring and early warning of flood and drought disasters, safe operation of reservoirs and dams, prevention and control of embankment and slope engineering diseases, verification and calibration of river dynamics models, etc., and are directly related to scientific decision-making on flood prevention and rescue, and the risk of water conservancy project accidents. Prevention and control effectiveness. The flow field measurement technology based on the principle of particle image velocimetry (PIV) has the characteristics of non-contact instantaneous full-field flow velocity measurement, and has been applied in laboratory flumes, river engineering models and river prototype observations. It obtains the magnitude, direction, characteristics and distribution of local fluid motion displacement and velocity through the analysis and calculation of particle image sequences, which greatly improves the measurement capabilities of various complex flows in laboratory environments. However, estimating the motion vector of a fluid based on particle images is the core and difficult point of PIV technology. The implementation method depends not only on the hardware system but also on the characteristics of the fluid to be measured. The flow field measurement technology based on Large Scale Particle Image Velocimetry (LSPIV) does not rely on artificial tracer particles and is positioned on natural floating objects in the river to calculate the global surface flow field. It estimates the movement of tracers in the flow field and uses grayscale correlation matching technology as the principle to establish a time-averaged flow field model in the form of vector averaging. However, this method requires a large amount of spatial cross-correlation calculations and the resolution of the velocity field is not high. The flow field measurement technology based on Large Scale Particle Tracking Velocimetry (LSPTV) performs well in the case of sparse particles. It obtains the space of particles in the image by locating tracer particles with lower density and larger particles. distribution, and further obtain the sub-pixel centroid position of the particles, and estimate the flow field by matching the particle centroids in different time frames. However, it cannot handle the situation of high particle concentration in the flow field, and is sensitive to some parameters that need to be manually set. The flow field measurement technology based on Space-time Image Velocimetry (STIV) has the characteristics of high spatial resolution and strong real-time performance. It synthesizes space-time images and detects the texture principal of space-time image detection based on Fourier transform. The direction calculates the one-dimensional time-averaged flow field. However, its temporal resolution is not high and it is sensitive to complex lighting.

The optical flow method was originally proposed in the field of computer vision. It is mainly used to estimate obvious rigid motion from image sequences. Because it can obtain dense velocity vector fields from image pairs, it is also used to estimate river motion scenes. In the field of fluid motion estimation, Horn and Schunck proposed a classical variational optical flow estimation model based on the brightness conservation assumption. However, this constant brightness constraint is too ideal to meet the actual situation. Therefore, the traditional H-S optical flow method is very sensitive to illumination changes, and the robustness of optical flow is very poor.

In order to improve the robustness to illumination changes, many improvements have been made in the field of computer vision based on the classic variational optical flow method. For example, the structure-texture decomposition method is used for preprocessing to decompose the image into structural parts and contain fine-scale details. The texture part is then used instead of the original grayscale image for subsequent optical flow calculations. Also using the idea of structure-texture decomposition, the feature optical flow method decomposes the image information into target feature points and background, and estimates the motion of the feature points through the position changes of the same feature point in the time domain in adjacent frames of the image sequence, and then Estimating river movement. The main process of the feature optical flow method includes feature extraction and tracking. Commonly used feature extraction algorithms include SIFT (Scale-Invariant Feature Transform), SURF (Speed-Up Robust Feature), FAST (Features from Accelerated Segment Test) and ORB (Oriented FAST and Rotated BinaryRobustIndependent Elementary Feature) etc. SIFT and SURF construct a scale space and calculate the main direction of features. The algorithm has scale invariance and rotation invariance, but is more sensitive to illumination changes and has a slow calculation speed. The FAST algorithm uses the grayscale structure information of detected pixels and surrounding pixels to determine Feature points, the algorithm is simple and the calculation speed is fast. The ORB algorithm is further improved on the basis of FAST and adds an image pyramid, which makes the algorithm scale-invariant and further improves the calculation speed. Both the FAST and ORB algorithms use a single fixed threshold to detect feature points. This threshold relies on manual experience value setting. When the lighting conditions on the river surface change and uneven illumination, local glare, etc. occur on the river surface, the threshold parameters need to be manually adjusted. To ensure that a sufficient number of feature points are extracted for optical flow calculation.

Contents of the invention

Purpose of the invention: The present invention aims to provide a method for estimating optical flow on the river surface based on the illumination-adaptive ORB operator. The global KSW entropy and local gray contrast are used to obtain the global adaptive threshold and the local adaptive threshold respectively to achieve illumination adaptation. , which can stably extract river water surface feature points for water surface optical flow estimation under complex lighting conditions in the wild.

Technical solution: The river surface optical flow estimation method based on illumination adaptive ORB operator according to the present invention includes the following steps:

(1) Grayscale the RGB images of two adjacent river water surface video frames, and intercept the ROI image in the river water surface area;

(2) Downsample the two frames of ROI images according to the scale factor s, and construct an image pyramid to make the image scale invariant.

(3) For the first frame image pyramid, calculate the illumination adaptive ORB threshold T layer by layer and pixel by pixel, and extract M feature points from the first frame image through the ORB feature extraction method;

(4) Use the Harris scoring algorithm to sort the quality of M feature points, remove feature points with Harris corner response values less than 0, and obtain N high-quality feature points;

(5) Estimating the optical flow motion vector through the sparse LK optical flow method.

Further, in step (3), the illumination adaptive ORB threshold T is

T＝α*T ₁ +β*T ₂

Among them, α and β are weight coefficients, T ₁ is the global adaptive threshold calculated layer by layer for the first frame image pyramid, and T ₂ is the local adaptive threshold calculated pixel by pixel.

Further, the global adaptive threshold T ₁ is calculated through the KSW entropy method, as follows:

Suppose the gray level interval of the gray level distribution histogram of the pyramid image is [0, L], normalize the frequency of occurrence of all gray levels, and divide the gray level interval into L ₁ = [0, t] and L ₂ =[t+1,L], the sum of entropies S corresponding to L ₁ and L ₂ :

S＝H(L ₁ )+H(L ₂ )

Then the global adaptive threshold T ₁ of the current image pyramid is:

T ₁ =|t _max +t _min |

Among them, t _max is the gray value corresponding to the maximum sum of entropy S, and t _min is the gray value corresponding to the minimum sum of entropy.

Furthermore, the entropy values corresponding to L ₁ and L ₂ are H(L ₁ ) and H(L ₂ ) respectively:

Among them, p _i is the frequency of occurrence of each gray level in the gray level interval, is the sum of the frequencies of all gray levels in L ₁ ,/> is the sum of the frequencies of all gray levels in L ₂ .

Further, the local adaptive threshold T ₂ is calculated as follows:

After removing edge pixels with an odd number of pixels in width from each layer of the pyramid image, draw a circle with the current pixel point P as the center and m pixels as the radius, obtain the grayscale values of n pixels on the circumference, and remove the maximum and minimum values in order to Exclude outliers and average the remaining pixels to obtain the local adaptive threshold T ₂ of the current pixel:

Among them, I _i is the gray value of the pixel numbered i on the circle, I _max and I _min are respectively the maximum gray value and the minimum gray value among n pixels, m and n are positive integers.

Preferably, m is 3 and n is 16.

Preferably, in step (3), in order to speed up detection efficiency, the FAST rapid optimization scheme is used in the determination process of feature points.

Further, the FAST rapid optimization plan is as follows:

Calculate the absolute value of the difference between the current pixel point P and the gray value of the four pixels at the horizontal and vertical lines on the circle. If three absolute values are greater than the illumination adaptive ORB threshold T, continue to judge other pixels on the circle. Otherwise, the pixel is discarded directly, and the mark value P _l of the current pixel P is

Among them, if P _l is equal to 1, then retain the pixel, record it as a feature point, and mark it as 1; otherwise, discard the pixel and mark it as 0; I ₀ is the gray value of the current pixel point P, and I _i is the circle The gray value of the i-th pixel on .

Further, in step (5), the process of estimating the optical flow motion vector by the sparse LK optical flow method:

The sparse LK optical flow method is used to track the high-quality feature points on the first frame of the image, determine their movement to the coordinates of the second frame of the image, and calculate the optical flow motion vector of the high-quality feature points on the two frames of images.

Further, the optical flow motion vector of high-quality feature points is as follows:

_{Among them} , _v The horizontal pixel coordinates on the frame image, y ₁ is the longitudinal pixel coordinate of the high-quality feature point on the first frame image, x ₂ is the horizontal pixel coordinate of the high-quality feature point on the second frame image, y ₂ is the high-quality feature point on the second frame image. Vertical pixel coordinates on the two frames of images.

Beneficial effects: Compared with the existing technology, the significant advantages of the present invention are: the river surface optical flow estimation method of the illumination adaptive ORB algorithm in the present invention does not require manual parameter adjustment, and can calculate the ROI according to the global KSW entropy and local gray contrast. Each pixel of the image calculates its adaptive threshold, which can stably detect and extract a sufficient number of river surface feature points when the illumination of the river surface changes non-uniformly for further sparse LK optical flow estimation, and the error of the optical flow estimation result remains within Within a certain range; it solves the problem that under complex outdoor river water surface lighting conditions, illumination changes cause uneven illumination on the river water surface, and the ORB feature point extraction algorithm with a single fixed threshold requires manual adjustment of parameters to obtain the feature points of the ROI image.

Description of drawings

Figure 1 is a flow chart of the present invention;

Figure 2 is the current pixel and neighborhood ring pixel information map;

Figure 3 shows the first frame of video image among two adjacent frames of video images;

Figure 4 is the second frame of video image among two adjacent frames of video images;

Figure 5 is a schematic diagram of the ROI area of the first frame of video image;

Figure 6 shows the feature point detection results of the first frame of ROI image by the illumination adaptive ORB algorithm;

Figure 7 shows the feature point detection results of the first frame ROI image by the fixed threshold ORB algorithm;

Figure 8 shows the feature point detection results of the first frame of ROI image by the FAST algorithm;

Figure 9 shows the optical flow diagram calculated by the lighting adaptive ORB algorithm;

Figure 10 shows the optical flow diagram calculated by the fixed threshold ORB algorithm;

Figure 11 shows the optical flow diagram calculated by the FAST algorithm;

Figure 12 shows the coordinate position map of the manually marked feature points on the first frame of the image;

Figure 13 is a marker map of the position of the feature points calculated by the LK optical flow method in the second frame of the image;

Figure 14 is a coordinate position diagram of manually marked feature points moving to the second frame image.

Detailed ways

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

As shown in Figure 1, the specific implementation of the present invention is as follows:

Step1: Take two adjacent frames of RGB images from the monitoring video of the wild river scene, as shown in Figures 3 and 4. Convert the two frames of RGB images into grayscale images, and intercept the ROI image in the river water surface area. The lighting conditions in the ROI area are relatively complex, as shown in Figure 5. Area 1 is the ROI area used for calculation of the river water surface. Area 1 is all in Among the shadows formed by the trees on the shore on the water, Area 2 included in Area 1 creates a glare on the water due to the sun.

Step2: Downsample the two frames of ROI images proportionally. Each frame of the original high-resolution ROI image is used as the first layer. In order to reduce information loss as much as possible, downsampling is performed according to the scale factor s=1.2 to construct an 8-layer pyramid image to make the image scale invariant;

Step3: Traverse each layer of the pyramid image, traverse each pixel, calculate the ORB adaptive threshold T for different pixels, and use the ORB feature extraction method to extract feature points based on the threshold T. The maximum number of feature points is set to M=3000. Among them, in the process of determining feature points, in order to speed up detection efficiency, the FAST rapid optimization scheme is used.

Among them, the ORB adaptive threshold T is the sum of the global adaptive threshold T ₁ and the local adaptive threshold T ₂ multiplied by the weight coefficient respectively, as shown in Equation (1);

T＝α*T ₁ +β*T ₂ (1)

Among them, α and β are weight coefficients. Through a large number of experiments, it has been demonstrated that the feature point extraction effect is optimal when α = 0.2 and β = 0.2.

Step3.1: Calculate a global adaptive threshold T ₁ for each layer of the ROI image pyramid in the first frame. The global adaptive threshold T ₁ is calculated based on the KSW entropy of the current pyramid image. The specific implementation method is as follows:

First calculate the one-dimensional grayscale distribution histogram of the current pyramid image, which is divided into a total of L+1 grayscale levels from 0 to L, and the frequency of occurrence of all grayscale levels is normalized. Then, a grayscale value t is preset. At this time, t divides the grayscale range into two parts: L ₁ =[0,t] and L ₂ =[t+1,L]. Each grayscale level in L ₁ appears. The frequency is {p ₁ , p ₂ , p ₃ ,…p _t }, and the frequency of each gray level in L ₂ is {p _t+1 , p _t+2 , p _t+3 ,…p _L }. Sum the frequencies of all gray levels in L ₁ and write it as Then the entropy values corresponding to the gray scale ranges of L ₁ and L ₂ are H(L ₁ ) and H(L ₂ ) respectively. The specific calculation formulas are as shown in equations (2) and (3).

Sum the entropy values of the two parts and record the sum of entropies as S. The calculation formula of S is as shown in Equation (4).

S＝H(L ₁ )+H(L ₂ ) (4)

Then, traverse each gray value t in the gray level range 0 to L, calculate the corresponding sum of entropy S according to the above formula, record the corresponding gray value when the sum of entropy is maximum as t _max , and the sum of entropy is the maximum The gray value corresponding to hour is t _min .

Then the final global adaptive threshold T ₁ is equal to the absolute value of the difference between the two, and the formula is shown in Equation (5).

T ₁ =|t _max +t _min | (5)

Step3.2: After calculating the global adaptive threshold T ₁ of each layer for the first frame image pyramid, remove edge pixels 7 pixels wide from each layer of the pyramid, traverse all the remaining pixels, and Calculate a local adaptive threshold T ₂ for each point. Each time a local adaptive threshold T ₂ is obtained, whether the current pixel point is a feature point is determined based on the improved ORB adaptive threshold T. The specific calculation method of the local adaptive threshold T ₂ is as follows:

As shown in Figure 2, with the current pixel as the center of the circle and 3 pixels as the radius, obtain the grayscale values of 16 pixels on the circle, remove the maximum and minimum values in order to exclude outliers, and average the remaining 14 pixels to obtain The local adaptive threshold T ₂ of the current pixel. The specific calculation formula is shown in Equation (6).

Among them, I _i is the gray value of the i-th pixel on the circumference of the current pixel point, I _max and I _min are respectively the maximum and minimum gray values among the 16 pixels on the circumference, and n is a fixed value of 16.

Step3.3: Since ORB is improved on the basis of FAST, the specific feature point determination rules are still consistent with FAST. Therefore, in order to speed up the calculation efficiency, when judging whether the current pixel point is a feature point, use the FAST fast optimization scheme. The specific implementation rule is to first test whether the absolute value of the difference between the gray value of the pixel points 1, 5, 9, and 13 on the current pixel circle shown in Figure 1 and the gray value of the current pixel point is greater than the previously calculated value. Adaptive threshold T, if at least 3 pixels among the 4 pixels meet the condition, then the remaining pixels on the circle will be further judged, otherwise, the pixel will be discarded. The specific formula is shown in Equation (7).

Among them, P is the current pixel of the image, and P _l is the mark value of the current pixel. If P is equal to 1, the pixel is retained and recorded as a feature point; otherwise, the pixel is discarded. I ₀ is the gray value of the current pixel, I _i is the gray value of the i-th pixel on the circle, and T is the improved ORB adaptive threshold calculated above.

Step 4: After detecting M = 3000 feature points, use the Harris scoring algorithm to calculate the Harris corner response function value of each feature point, and rank all feature points in quality order according to the size of the corner response of the feature point. The response function of the corner point is a large positive number, the response function of the edge is a large negative number, and the response function of the flat area is a small value. Therefore, the feature point with a larger Harris corner response function value is regarded as a high-quality feature point. First, the feature points with Harris corner response value less than 0 are recorded as wrong feature points and removed, and then N=500 high-quality feature points with the largest Harris corner response function value are retained. The distribution diagram is shown in Figure 6. Among them, the circle marks the location of the feature point.

Comparison algorithm one - the feature point distribution diagram extracted by the fixed threshold ORB feature point detection algorithm is shown in Figure 7. The circles marked are the positions of the feature points, and the boxes marked are those that were not detected by the method but were successfully detected by the method of the present invention. extracted feature points.

Comparison algorithm two - the extracted feature point distribution diagram of the FAST feature point detection algorithm is shown in Figure 8. The circles marked are the positions of the feature points, and the boxes marked are the ones that were not detected by the method but successfully extracted by the method of the present invention. Feature points.

It can be seen from Figure 5 that the algorithm of the present invention has completely extracted the floating objects with continuous movement that can be clearly observed in the ROI area, while the comparison algorithm has omissions, that is, the feature points marked by the box are all obvious floating objects on the water surface. , should be extracted as feature points for optical flow estimation, but both comparison algorithms have omissions and were not successfully extracted.

In order to verify the stability of the algorithm, three methods were used to detect feature points in a total of 750 frames of the 30s video including the video frame in Figure 3, and then LK optical flow tracking was performed. The number of feature points extracted from 750 frames of images, the number of error feature points and the number of successfully tracked feature points were averaged. The results are shown in Table 1.

Table 1 Performance comparison of three detection algorithms

It can be seen from Table 1 that the ORB feature point detection method based on illumination adaptation of the present invention has extracted a sufficient number of feature points that can be used for calculation. There are 7 error feature points extracted by the FAST method. The feature points of the three methods all have good successful tracking rates.

Step5: After obtaining N=500 high-quality feature points on the first frame of the image, input the two frames of pyramid images and the high-quality feature points on the first frame of the image, and use the sparse LK optical flow method to identify the high-quality feature points on the first frame of the image. Track it and estimate its movement to the second image. For the tracked feature point, according to its pixel coordinate _Pi (x ₁ , y ₁ ) on the first frame image and its movement to the pixel coordinate _Pi (x ₂ , y ₂ ) on the second frame image, Calculate its optical flow motion vector on two adjacent images. Among them, i represents the label of the feature point, x ₁ represents the horizontal pixel coordinate of the feature point on the first frame image, y ₁ represents the longitudinal pixel coordinate of the feature point on the first frame image, x ₂ represents the feature point in the second frame The horizontal pixel coordinate on the image, y ₂ represents the vertical pixel coordinate of the feature point on the second frame image. The specific calculation formulas for calculating the optical flow motion vector of the feature point based on the corresponding image coordinates of the same feature point on the two frames of images are as shown in Equations (8) and (9).

In Equations (8 ₎ and (9), _v Pick

Figure 6 is an optical flow result diagram of the LK calculation of the feature points extracted by adaptive ORB, fixed threshold ORB and FAST. The optical flow result diagram calculated by the method of the present invention is shown in Figure 9. The optical flow calculated by the fixed threshold ORB algorithm The result graph is shown in Figure 10, and the optical flow result graph calculated by the FAST algorithm is shown in Figure 11. The dots represent the end point of the optical flow vector, and the lines represent the motion trajectories of feature points on two adjacent frames of images. To facilitate visualization, the size of the optical flow in the image is expanded to 11 times its original size. It can be seen from Figure 9, Figure 10 and Figure 11 that all three algorithms calculate a certain number of optical flows for the feature points extracted in the feature point extraction stage. According to the principle that the optical flow sizes of the same y pixel coordinates should be basically consistent, the present invention The optical flow calculation results calculated by this method are relatively consistent with the actual situation.

Select 5 feature points from 500 feature points to calculate the optical flow vector for performance evaluation. Feature points a, b, c, and d are 4 high-quality feature points among the 500 feature points. Feature point e is the success of the method of the present invention. Feature points extracted but not detected by the two comparison algorithms. Since the motion information of the feature points of the river water surface image has no real value, according to the principle that the neighborhood information of the feature points is basically consistent on the two adjacent images, the coordinate positions of the feature points on the two adjacent images are manually marked. Manually labeled values are used as ground truth values for performance evaluation. The manually marked position map of the five feature points on the first frame of the image is shown in Figure 12. The movement of the manually marked feature points to the true coordinates of the second frame of the image is shown in Figure 13 according to the above consistent motion principle. Move the feature points calculated by the LK optical flow method to the coordinate position of the second frame image for marking, as shown in Figure 14. Among them, the obviously black marked pixel on the left in area 3 of Figure 12, Figure 13 and Figure 14 is the coordinate position of feature point a, the obviously black marked pixel on the right is the coordinate position of feature point b, and the obviously black marked pixel in area 4 is the coordinate position of feature point a. The marked pixel point is the coordinate position of the feature point c, the obviously black marked pixel point in area 5 is the position of the feature point d, and the obvious white marked pixel point in area 6 is the position of the feature point e. The coordinates of the five feature points on the ROI image of the first frame calculated by the three algorithms, the coordinates of the manually estimated feature points on the ROI of the first frame image, and the motion of the feature points calculated by the LK optical flow method based on the coordinates of the feature points of the first frame. The coordinates on the ROI of the second frame image and the artificially estimated feature point movement to the coordinates on the ROI of the second frame image are shown in Table 2.

Table 2 Coordinate calculation results of experimental feature points on two frames of images in the wild river scene

It can be seen from Table 2 that the five feature points extracted by the adaptive ORB algorithm have a certain effectiveness in optical flow motion estimation. The one-way estimation errors of the position coordinates of the five experimental feature points are all within 1 pixel, with a maximum of 0.56 pixels.

According to equations (8) and (9), the optical flow motion vectors of the five experimental feature points extracted by the illumination adaptive ORB algorithm and the manually estimated optical flow motion vector are calculated, and the calculation error values of the two are obtained. The results are as shown in Table 3 shown.

Table 3 Optical flow calculation results of experimental feature points on two adjacent frames of images in the wild river scene

It can be seen from Table 3 that the _V Manually marked value, which is consistent with the rule that similar floating objects at the same starting point distance basically maintain the same flow rate. The experimental data fully proves the feasibility of the river surface optical flow estimation method based on the ORB illumination adaptive operator of the present invention.

Claims (10)

1. A river water surface optical flow estimation method based on an illumination self-adaptive ORB operator is characterized by comprising the following steps:

(1) Graying RGB images of two adjacent river water surface video frames, and intercepting ROI images in river water surface areas;

(2) Downsampling two frames of ROI images according to a scale factor s to construct an image pyramid;

(3) Aiming at the pyramid of the first frame image, calculating an illumination self-adaptive ORB threshold T layer by layer and pixel by pixel, and extracting M feature points from the first frame image by an ORB feature extraction method;

(4) Carrying out quality sequencing on M characteristic points through a Harris scoring algorithm, and removing characteristic points with Harris corner response values smaller than 0 to obtain N high-quality characteristic points;

(5) The optical flow motion vector is estimated by the sparse LK optical flow method.

2. The method for estimating the optical flow of the river water surface based on the illumination-adaptive ORB operator according to claim 1, wherein in the step (3), the illumination-adaptive ORB threshold T is

T＝α*T ₁ +β·T ₂

Wherein alpha and beta are weight coefficients, T ₁ For the global adaptive threshold value calculated layer by layer for the first frame image pyramid, T ₂ A locally adaptive threshold is calculated for each pixel.

3. The method for estimating the river water surface optical flow based on the illumination-adaptive ORB operator according to claim 2, wherein the global adaptive threshold T is calculated by a KSW entropy method ₁ The method is characterized by comprising the following steps:

let the gray level interval of the gray level distribution histogram of pyramid image be [0, L]Normalizing the frequency of occurrence of all gray levels to divide the gray level interval into L ₁ ＝[0,t]And L ₂ ＝[t+1,L]，L ₁ And L ₂ Sum of corresponding entropies S:

S＝H(L ₁ )+H(L ₂ )

then the global adaptive threshold T of the current image pyramid ₁ The method comprises the following steps:

T ₁ ＝|t _max +t _min |

wherein t is _max Ash corresponding to the sum of entropy S maximumDegree value, t _min The gray value corresponding to the minimum sum of the entropies.

4. A river water surface optical flow estimation method based on illumination-adaptive ORB operator according to claim 3, wherein L ₁ And L ₂ The corresponding entropy values are H (L ₁ ) And H (L) ₂ )：

Wherein p is _i For the frequency of occurrence of each gray level in the gray level interval, P _L1 Is L ₁ The sum of the frequencies of occurrence of all gray levels, P _L2 Is L ₂ The sum of the frequencies of occurrence of all gray levels in the picture.

5. The method for estimating the river water surface optical flow based on the illumination-adaptive ORB operator according to claim 2, wherein the locally adaptive threshold T ₂ The calculation method is as follows:

after removing edge pixel points with odd pixel width of each layer of pyramid image, drawing a circle by taking the current pixel point P as a circle center and m pixels as a radius, obtaining gray values of n pixel points on the circle, removing maximum and minimum values to remove abnormal values, and averaging the rest pixel points to obtain a local self-adaptive threshold T of the current pixel point ₂ ：

Wherein I is _i Is the gray value of the pixel point with the number I on the circumference, I _max And I _min The maximum gray value and the minimum gray value in n pixel points are respectively, and m and n are positive integers.

6. The method for estimating the river water surface optical flow based on the illumination-adaptive ORB operator according to claim 5, wherein m is 3 and n is 16.

7. The method for estimating the river water surface optical flow based on the illumination-adaptive ORB operator according to claim 6, wherein in the step (3), a FAST optimization scheme is adopted in the process of judging the feature points.

8. The method for estimating the river water surface optical flow based on the illumination self-adaptive ORB operator according to claim 7, wherein the FAST optimization scheme is specifically as follows:

respectively calculating absolute values of differences between gray values of the current pixel P and the 4 pixel points on the horizontal line and the vertical line on the circumference, if 3 absolute values are larger than the illumination self-adaptive ORB threshold T, continuously judging other pixel points on the circumference, otherwise, directly discarding the pixel point, and directly discarding the marking value P of the current pixel P _l Is that

Wherein if P _l If the pixel point is equal to 1, reserving the pixel point, marking the pixel point as a characteristic point, and marking the pixel point as 1; otherwise, discarding the pixel point, and marking as 0; i ₀ For the gray value of the current pixel point P, I _i Is the gray value of the pixel point at the i-th position on the circumference.

9. The method for estimating the river water surface optical flow based on the illumination-adaptive ORB operator according to claim 1, wherein in the step (5), the sparse LK optical flow method estimates the optical flow motion vector:

and tracking the high-quality feature points on the first frame image by a sparse LK optical flow method, determining coordinates of the high-quality feature points moving to the second frame image, and calculating optical flow motion vectors of the high-quality feature points on the two frame images.

10. The method for estimating the river water surface optical flow based on the illumination-adaptive ORB operator according to claim 9, wherein the optical flow motion vectors of the high-quality feature points are as follows:

wherein v is _x Transverse pixel optical flow motion vector v which is high-quality characteristic point _y Longitudinal pixel optical flow motion vector of high-quality feature point, delta t is time interval of two adjacent frames of images, and x ₁ For the transverse pixel coordinates, y, of the high-quality feature points on the first frame image ₁ For the longitudinal pixel coordinates, x, of the high-quality feature points on the first frame image ₂ For the transverse pixel coordinates, y, of the high-quality feature points on the second frame image ₂ Vertical pixel coordinates of the high-quality feature points on the second frame image.