Eagle tenon visual perception imitating unmanned aerial vehicle cluster'chasing-escaping 'tracking method

Abstract

The invention discloses an eagle-falcon-visual-perception-imitated unmanned aerial vehicle cluster'chasing-escaping 'tracking method. The method comprises the steps that 1, images of double fovea central receptive fields imitating eagle falcon vision are pre-compressed; 2, adaptively adjusting the contrast of the images imitating the eagle-tenon vision; 3, extracting image features imitating eagle and tenon vision; 4, self-adaptive adjustment of the space mask imitating the dynamic vision of the eagle falcon vision; 5, designing a discriminant correlation filtering tracker of the unmanned aerial vehicle target; and step 6, solving a filtering tracker. The method has the advantages that the feature extraction and target search efficiency is improved, the search range is guaranteed, and simplicity, high efficiency and high real-time performance are achieved; 2) designing an eagle-tenon-imitated visual contrast adaptation and color antagonism feature extraction pre-edge feature extraction method, and performing feature weight regulation and control according to positive and negative feedback effects of an eagle-tenon visual processing path; and 3) designing a self-adaptive space mask for the target correlation filtering tracker of the unmanned aerial vehicle, and ensuring that the tracker can better process target motion and reduce background information influence.

Description

Technical Field

[0001] The invention discloses a "pursuit-escape" tracking method for a drone cluster that imitates the visual perception of a falcon, and belongs to the field of computer vision tracking. Background Art

[0002] With the gradual development of modern unmanned technology, intelligent and autonomous equipment and systems have been widely used in various industries, especially drones have been used in many scenarios such as rescue and disaster relief, mapping, cargo transportation, etc., especially in the modern battlefield environment, drones have extremely high strategic value. Therefore, in anti-drone and drone swarm collaboration tasks, drone target tracking tasks are very necessary.

[0003] Target tracking technology is generally divided into three categories, including generative trackers, discriminative trackers, and deep learning trackers. The core idea of ​​the tracker is to establish a specific target model based on the existing target information, and determine the candidate position of the target in the subsequent image sequence based on the established model to achieve tracking. Generative trackers and deep learning trackers mainly achieve tracking by establishing an overall model of the target and finding similar areas in subsequent frames, which has a high computational cost; while discriminative trackers mostly establish a discriminant function corresponding to the extracted features as the target model, which reduces the model complexity and improves the tracking efficiency.

[0004] Drone targets are characterized by small size, fast movement speed, and complex movement scenes. In addition, in drone swarm combat scenarios, drone "chase-escape" tasks are important features of drone swarm game and combat. At this time, when the rapid movement of drones causes changes in their posture, the imaging of drone targets will not only change in position, but also in size or shape, and even in color when the ambient light changes. Existing drone target tracking algorithms do not perform well in the above situations. They often lose targets due to rapid drone movement or light changes, and the size of the tracked targets does not match. Therefore, these algorithms cannot adapt well to the multi-scene changes and drone target tracking tasks under rapid movement of drone swarm "chase-escape" tasks.

[0005] The hunting scenes of falcons in nature are similar to the target tracking missions of UAVs. As excellent aerial predators, falcons can not only find distant prey targets under complex backgrounds and large lighting differences, but also ensure the tracking of prey during the pursuit. The hunting characteristics of falcons are inseparable from their visual system. Analyzing the visual target tracking characteristics of falcons, establishing a unique image processing model, and mapping it to the target tracking mission of UAVs is a new method to solve the target tracking of UAVs. Summary of the invention

[0006] The present invention proposes a "pursuit-escape" tracking method for a drone cluster based on falcon visual perception. The method establishes a tracker front-end information processing model by imitating the falcon visual feature extraction framework, and optimizes the inter-frame tracker processing mode in combination with the falcon visual information processing mechanism. The method is used for drone target tracking problems under conditions of rapid motion and background changes. The method has the characteristics of high tracking accuracy, fast calculation speed, and good background adaptability, meets the actual drone target tracking needs, and has certain practical significance.

[0007] The purpose of the present invention is to provide a UAV cluster "pursuit-escape" tracking method based on falcon visual perception, which aims to deal with the problems of low tracking accuracy and slow tracking speed caused by background interference and rapid movement in the process of UAV target tracking. By establishing the perception information method of falcon vision in the hunting process, establishing a multi-feature extraction and processing model, and combining the brain circuit processing method of visual information, establishing an inter-frame information correction model, and further mapping it to the discriminant filter UAV target tracking scene, a new method is proposed for the UAV target tracking problem.

[0008] The present invention proposes a UAV cluster "pursuit-escape" tracking method based on falcon visual perception, and the specific implementation steps are as follows:

[0009] Step 1: Image pre-compression simulating the bifoveal receptive field of hawk vision

[0010] For the initial image, in order to ensure the efficiency of the tracker in the process of feature extraction and candidate area search, an image pre-compression method was designed based on the dual foveal receptive field structure of falcon vision. Models simulating the deep fovea and shallow fovea of ​​falcon vision were established respectively to perform preliminary compression on the image, which reduced the computational cost of subsequent feature extraction and calculation while ensuring the effective features of the image.

[0011] S11. Image pre-compression of a falcon vision shallow fovea model

[0012] The shallow fovea in the falcon's vision has a large field of view and a central-peripheral receptive field structure. This process is simulated by Gaussian filtering. The filter kernel of this process can be expressed as:

[0013]

[0014] Among them, (x d ,y d ) is the distance between the pixel and the center of the shallow fovea, σ 1 is the standard deviation of the two-dimensional Gaussian filter kernel that simulates the receptive field size of ganglion cells. The pre-compression process can be expressed as:

[0015] I c,s (x,y)=I o (x,y)*G1 (x d ,y d ,σ 1 ) (2)

[0016] Among them, I o (x, y) is the original image, * is the convolution operation, I c,s (x,y) is the pre-compression result after processing by the shallow fovea filter kernel.

[0017] S12, Image pre-compression of the deep foveal model of falcon vision

[0018] Compared with the falcon vision shallow fovea model, the falcon vision deep fovea model has a higher density of visual cells in the central area, and the difference between central and peripheral attention is more obvious. Therefore, the deep central model has a stronger ability to pre-compress images and forms an attention mechanism. This mechanism takes an area of ​​interest as the core, and the information sensitivity of the surrounding areas decays exponentially. For this reason, a pre-compression model of deep fovea image imitating falcon vision is established. Figure 1 As shown. The region of interest is set to have a side length of ρ 0 Pixels outside this area are Figure 1 The method shown performs selective compression, and the candidate region radius ρ n It can be expressed as:

[0019] ρ n =ρ 0 α n ,n=1,2,…,N ρ (3)

[0020] Among them, ρ n is the side length of the rectangular pixel band of the candidate area centered on the candidate area in the nth layer of the original image, N ρ is the total number of pixel layers to be compressed, and α is the attenuation coefficient, which is defined as:

[0021]

[0022] Among them, ρ m is the side length of the maximum candidate rectangular pixel strip. At the same time, in order to ensure the integrity of the compressed image, the pixel interval of each candidate rectangular pixel strip during compression is defined as:

[0023]

[0024] Among them, Δρ n is the pixel sampling interval of the nth pixel band. n is the number of pixel band layers.

[0025] The model establishes a falcon-like deep foveal vision model by retaining the central receptive field and compressing the peripheral receptive field in an exponential form, so that a larger receptive field is included in a limited pixel area, and the shallow fovea pre-compresses the image I c,s Based on this operation, the pre-compressed image I is obtained. c,d . Feature extraction in pre-compressed image I c,d conduct.

[0026] Step 2: Image contrast adaptation to simulate falcon vision

[0027] Compared with other biological visual systems, falcon vision has higher contrast sensitivity. The regional contrast sensitivity function is used to measure the contrast information in the image and is defined as:

[0028]

[0029] Among them, I in (x, y) is the average brightness of the image in the pixel area, I out (x, y) is the average brightness of the image outside the pixel, and CS(x, y) is the regional contrast. The inner and outer regions are determined by specifying the inner and outer region radii, such as Figure 2 shown.

[0030] In order to reduce the influence of complex background interference on UAV feature extraction and highlight the UAV features in the complex background, a falcon-like visual contrast adaptive image contrast feature extraction model is designed based on the enhancement effect of falcon vision on contrast information:

[0031]

[0032] Among them, P in is the neighboring pixel in the area of ​​the pixel at position (x, y), I g (x, y) is the pixel grayscale value, mid(·) is the median calculation function, T cs,1 ,T cs,1 is the regional contrast threshold, I cs (x,y) is the result of adaptive feature extraction based on hawk-like visual contrast.

[0033] Step 3: Extract image features based on hawk vision

[0034] S31, Color and edge feature extraction imitating falcon vision

[0035] Falcon vision has a unique image feature processing mechanism. The color processing in the falcon's visual cells is mainly achieved by the combination of single cone cells of different wavelengths. The combination is mainly achieved through the difference of different color pathways. This process is called color antagonism. Color antagonism ensures the constancy of falcon vision in identifying colors under different light. Based on this process, a multi-antagonistic combination model based on red-green and yellow-blue color channels was established as follows:

[0036] I CO =H CO [I R I G I B I Y ] T (8)

[0037] Among them, I R I G I B I Y is the component of the red, green, blue and yellow channels of the image, and the yellow channel component I Y Obtained by combining the other three color channels, I CO is the 8-layer feature vector after color antagonism, H CO is an 8×4 weight matrix, expressed as:

[0038]

[0039] Among them, 0 is a 0 matrix of suitable dimension, H CO,1 is a 4×2 weight matrix, defined as:

[0040]

[0041] The dual antagonism in falcon vision is mainly used to extract edge information in images. In addition, the falcon thalamus nucleus has an orientation matrix structure that is sensitive to orientation perception. The falcon vision extractor model is used to simulate the edge information extraction model. The method first divides the image into blocks, calculates the gradient value and direction of each pixel in each block, and divides 360° into 9 directional grids. According to the gradient value and direction of each pixel, a 9-dimensional gradient histogram is established for each block, thereby obtaining a 9-dimensional HOG vector I for each block. HOG (x,y).

[0042] The normalized truncation method is used to calculate the interaction between blocks and combine them into a 36-dimensional vector. The process can be expressed as:

[0043]

[0044] Among them, N α,β (x, y) is the action vector of the surrounding blocks, I nHOGTo normalize the truncated vector, the vector dimension is 4×9, and then I nHOG Perform dimensionality reduction and use row summation and column summation to convert the 4×9 vector into a 1×13 dimensional vector I fHOG,1 Consider that each pixel gradient has a direction, that is, 360° is divided into 18 directional grids. According to the HOG feature extraction method, the HOG feature is obtained as an 18-dimensional vector I' HOG (x, y), the normalized truncation method and row summation are also used to reduce the dimension, and the 4×18 vector I' is obtained respectively nHOG and 1×18 vector I fHOG,2 . Splicing vector I fHOG,1 and I fHOG,2 Vector gets fHOG vector I fHOG .

[0045] S32, Multi-feature regulation of falcon-like visual processing

[0046] After falcons extract image features through visual-related structures and cells, they process, integrate and transmit them through multiple optic nerve pathways. Studies have found that there is a gain regulation network in the midbrain visual pathway of falcons. This network is composed of the optic tectum and the isthmus. When information is acquired and processed from the retina, it will be transmitted to the optic tectum in the form of neural stimulation. The isthmus receives cognitive regulation. The large cells and small cells in the isthmus will respond differently after receiving specific stimuli. They will give feedback to the input of the optic tectum. The large cells will produce negative feedback and the small cells will produce positive feedback, thereby achieving selective attention to the target. Following this process, a multi-feature regulation of falcon-like visual processing is established to provide feedback weights for different features. First, a basic weight model is established based on the image information entropy:

[0047]

[0048] Among them, w b (I k ) is the extracted feature vector I k The calculated basic weight, p i Indicates the proportion of significant pixels with gray value i. This basic weight represents the effectiveness of the information contained in the image. Inspired by the positive and negative feedback effects of falcon visual processing, the multi-feature control weight is designed as:

[0049]

[0050] Among them, w I (I k ) is the feature vector I k The multi-feature control weights of eagle-like visual processing, w b,t (I k ) is the basic feature weight of the feature vector at time t, w b,t-1 (Ik ) is the basic feature weight of the feature vector at time t-1, N c is the sum of all feature dimensions, including I cs ,I co ,I fHOG , norm is the normalization function, which standardizes the calculation result scale to ensure that the sum of all feature weights is 1. In the initial frame calculation process, w b,0 (I k ) are the same.

[0051] Step 4: Adaptive adjustment of spatial mask based on hawk-like dynamic vision

[0052] Falcon vision has sensitive dynamic vision, and the falcon's visual pathway is also subject to cognitive regulation, especially the role of spatial information. Therefore, a mask model that can adapt to the volume is designed based on the falcon's dynamic vision and spatial regulation, so as to achieve adaptive regulation of the size changes during drone tracking. Define the mask response variable k β1,k for:

[0053] k β1,k :=||f k-1 *β k-1 x c,k (w I (I c ),I c )-y k || (14)

[0054] Among them, f k-1 is the filter model of the k-1th frame discriminant filter tracker, which acts on the k-1th frame mask β k-1 and the k-th frame feature vector x c,k As the filtering result, y k is the prediction result, and its difference norm k β1,k Measure the mask response of the kth frame, mask β k-1 As a filter coefficient matrix, it is used to perform spatial priori screening for multiple eigenvectors, and the matrix elements are between 0 and 1. In order to describe the impact of target dynamics on mask response, the mask response change rate is defined as:

[0055] k β2,k :=k β1,k -k β1,k-1 (15)

[0056] Among them, k β1,k-1 is the mask response of the k-1th frame. The mask update model is designed as:

[0057]

[0058] Among them, r1 ,r 2 ,r 3 ,r 4 is the mask coefficient, ensuring β k The elements are between 0 and 1.

[0059] Step 5: Design of discriminant correlation filter tracker for UAV targets

[0060] Through the calculation of steps 1 to 4, combined with the extraction of feature vectors, the adjustment of weights and the adaptive mask, the discriminant correlation filter tracker imitating the falcon vision is designed as follows:

[0061]

[0062] Among them, λ 1 ,λ 2 is the spatial regularization coefficient to prevent the filter from overfitting, λ 3 is the time regularization coefficient, f t-1 is the previous frame filter, ||·|| 2 is the 2-norm, ||·|| F is the Frobenius norm, c is the characteristic channel, I c Contains I cs ,I co ,I fHOG , x c (w I (I c ),I c ) is the image feature after multi-feature regulation weights are applied.

[0063] Step 6: Filter Tracker Solution

[0064] Construct the auxiliary subsystem as follows:

[0065] f=h (18)

[0066] Using the augmented Lagrangian method (17-18), the unified function is obtained:

[0067]

[0068] The filter optimization problem defined in equation (19) is decomposed into three subsystems:

[0069]

[0070] Among them, η is the Lagrangian operator, μ is the scale factor of the Lagrangian operator, and the alternating direction multiplier method is used to solve the above problem. First, the optimization problems of the above two subsystems are solved alternately, and the Lagrangian operator is updated with the third subsystem, and the filter f is obtained by iterative solution.

[0071] The present invention proposes a "pursuit-escape" tracking method for a drone cluster that simulates the visual perception of a falcon, and the advantages and effects thereof are: 1) a multi-level image pre-compression model that simulates the shallow fovea and deep fovea of ​​falcon vision is established, which improves the image processing efficiency, reduces the dimension of the feature extraction image, and ensures that the tracker search is not affected; 2) in combination with a variety of feature extraction mechanisms of falcon vision, adaptive contrast feature extraction, color antagonism feature extraction and edge feature extraction algorithms are designed for drone target contrast enhancement, and multi-feature regulation combination weights are calculated based on the pathways in the eagle brain visual processing, which ensures the effective use of multiple features and reduces the interference of interference and background information on drone targets; 3) based on the spatial regulation and dynamic visual characteristics of falcon vision, an adaptive spatial mask that simulates the dynamic vision of falcon vision is designed, and the tracking of fast-moving drones is realized by combining the discriminant correlation filter, which proves that the present invention has certain practical significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 Schematic diagram of the image pre-compression principle simulating the deep fovea of ​​hawk vision.

[0073] Figure 2 Schematic diagram of the inner and outer areas of the regional sensitivity function.

[0074] Figure 3 Flowchart for the implementation of the drone swarm “chase-escape” tracking method based on falcon visual perception.

[0075] Figure 4 Partial results of tracking test of the method of the present invention in the drone video of scene 1.

[0076] Figure 5 Partial results of tracking test of the method of the present invention in scene 2 UAV video.

[0077] Figure 6 Some results of the tracking test of the method of the present invention in the drone video under special lighting in scene 3. The definitions of the relevant variables in the text and the accompanying drawings are as follows:

[0078] σ 1 : Standard deviation of the two-dimensional Gaussian filter kernel that simulates the receptive field size of ganglion cells;

[0079] (x d ,y d ): distance of the pixel from the center of the shallow fovea;

[0080] (x,y): pixel coordinates in the image;

[0081] G 1 (x d ,y d ,σ 1 ): Gaussian filter;

[0082] I c,s (x,y): image pre-compression result;

[0083] I o (x,y): original image;

[0084] ρ 0 : The side length of the region of interest;

[0085] ρ n : The side length of the rectangular pixel band of the candidate area centered on the candidate area in the original image of the (n)th layer;

[0086] N: total number of pixel layers to be compressed;

[0087] ρ

[0088] α: attenuation coefficient;

[0089] ρ m : The maximum candidate area rectangular pixel strip side length;

[0090] Δρ n : Pixel interval;

[0091] CS(x,y): regional contrast;

[0092] I in (x, y): the average brightness of the image in the pixel area;

[0093] I out (x, y): average brightness of the image outside the pixel;

[0094] mid(): median calculation function;

[0095] T cs,1 ,T cs,1 : Regional contrast threshold;

[0096] I cs (x,y): Result of adaptive feature extraction of falcon-like visual contrast;

[0097] I g (x,y): pixel grayscale;

[0098] I R I G I B I Y : The components of the red, green, blue and yellow channels of the image;

[0099] I CO : Feature vector after color antagonism;

[0100] H CO : Color weight matrix;

[0101] I HOG (x,y): HOG feature vector;

[0102] I fHOG :fHOG feature vector;

[0103] w b (I k ):Basic weights for feature extraction;

[0104] w I (I k ): Imitation brain visual processing control weight;

[0105] k β1,k :mask response variable;

[0106] f: filter model;

[0107] β k-1 :Mask;

[0108] r 1 ,r 2 ,r 3 ,r 4 : mask coefficient;

[0109] λ 1 ,λ 2 ,λ 3 : Regularization term coefficient;

[0110] η: Lagrangian operator;

[0111] μ: scale factor of Lagrangian operator; DETAILED DESCRIPTION

[0112] The effectiveness of the drone cluster "pursuit-escape" tracking method based on falcon visual perception proposed in the present invention is verified by a specific example below. In this example, a group of drone flight video data is selected as a test sample. The first frame of each video is used to mark the exact position of the drone and serve as the initial input of the tracker. The video contains a four-rotor drone target. The video includes different scenes and lighting at different times. The video frequency is 30Hz and the resolution is 1280*720. The simulation environment of this example is configured with a Ryzen 6800H processor, Nvidia GeForce RTX 3060 image processing unit, 3.2Ghz main frequency, 32G memory, and the software is MATLAB 2020a version.

[0113] The UAV swarm "chase-escape" tracking method based on falcon visual perception is shown in the following flowchart: Figure 3 As shown, the specific practical steps of this example are as follows:

[0114] First, initialize the relevant parameters of the falcon-like vision algorithm

[0115] Initial frame f 0 By calculating the true value, set f 0 =h 0 ; η 0 =0; k β1,0 =0; k β2,0 =0;β 0 =I; w b,0 (I co )=w b,0 (I cs )=w b,0 (I fHOG )=1 / 3. Starting from the second frame, the video sequence to be tested is input into the algorithm frame by frame for calculation.

[0116] Step 1: Image pre-compression of the bifoveal receptive field of eagles and raptors

[0117] S11, Pre-compression of shallow foveal images simulating hawk vision

[0118] Set σ in the falcon-like shallow foveal filter model s =0.7. Starting from the second frame, the original image is substituted into formula (1) to calculate the filter kernel corresponding to each pixel, where the center of the shallow fovea is the center of the target area of ​​the previous frame, and then the image is pre-compressed by processing through formula (2) to obtain image I c,s .

[0119] S12, hawk-like deep foveal image pre-compression

[0120] A central-peripheral image compression method inspired by the deep foveal mechanism of hawk vision. c,s As input, set ρ in the foveal-like vision deep foveal model 0 =65,ρ m =100,N ρ = 10, the image search area is defined as a 100*100 square area at the center of the target area of ​​the previous frame, and the region of interest is a 65*65 square area. Figure 1 The compression is performed in the manner shown, and the elements outside the 65*65 area are selected according to formula (3-5), and they are closely arranged to obtain the 75*75 image compression result I c,d .

[0121] Step 2: Image contrast adaptation to simulate falcon vision

[0122] The image compression result I c,d Convert RGB image to grayscale image I gSet the radius of the inner pixel area to 1 and the radius of the outer pixel area to 3, and calculate the mean of the neighboring pixels whose Euclidean distance to each pixel is less than 1 as the I of the central pixel. in (x, y), the mean of the neighboring pixels with a Euclidean distance less than 3 is taken as the I of the central pixel out (x, y), for each pixel, use formula (6) to calculate the regional contrast sensitivity, and set the contrast sensitivity threshold to T cs,1 =1.8, T cs,2 = 1.2, and use formula (7) to adaptively adjust the grayscale features of each pixel to obtain the adaptive result I of the simulated falcon visual contrast cs The dimensions are 75*75*1

[0123] Step 3: Extract image features based on hawk vision

[0124] S31, color feature extraction imitating falcon vision

[0125] Based on the multi-channel color antagonism mechanism of falcon vision, the image color information is extracted. c,d Red I R , green I G , blue I B Three-channel monochrome image, and the yellow channel monochrome image is obtained by calculating the red and green channels 0.5*(I R +I G )=I y . Use formula (8-10) to calculate the color antagonism feature result and obtain the color feature vector I with a dimension of 75*75*8 co .

[0126] S32, edge feature extraction imitating falcon vision

[0127] Considering the rich edge-sensitive cells of falcons, high-dimensional edge features are extracted from the image. c,d Convert RGB image to grayscale image I g Use the x and y direction gradient operators to obtain the x and y direction gradients corresponding to each pixel, and further calculate the gradient direction and gradient magnitude. Divide the image into blocks, each block contains 5*5 pixels, and use the non-directional segmentation method to divide 360° into 9 regions. Create a 9-dimensional gradient histogram for each and integrate it into the HOG vector I HOG , and use directional segmentation to generate 18 regions, which are integrated into the HOG vector I' HOG On this basis, the normalized truncation formula (11) is used to obtain the fHOG component I corresponding to the two HOG vectors fHOG,1 , I fHOG,2 , combined into the overall fHOG vector I fHOGIn order to ensure consistency, sampling is used to ensure that the vector dimension is 75*75*31.

[0128] S33, Multi-feature Control of Falcon Vision

[0129] On the basis of obtaining multiple eigenvectors, the weight value w of each eigenvector is calculated using formula (12-13) in imitation of the positive and negative feedback mechanism of the hawk's visual processing pathway: I (I co ), w I (I cs ), w I (I fHOG ), these weight values ​​regulate the three sets of feature vectors w I (I co )·I co , w I (I cs )·I cs , w I (I cs )·I cs And concatenate the three sets of regulated feature vectors to obtain a feature vector with a dimension of 75*75*40.

[0130] Step 4: Adaptive spatial mask calculation based on hawk-like dynamic vision

[0131] Falcon vision will spatially screen information based on cognition. To this end, a mask based on prior information is calculated. At the same time, the target tracking mechanism of falcon vision is simulated, and a dynamic adjustment item is added to the mask. The mask response variable is calculated by formula (14) and the mask change rate variable is calculated by formula (15) based on the filter result and prediction result of the previous frame, and the mask is further updated by formula (16).

[0132] Step 5: Design and solve the discriminant correlation table filter of UAV target

[0133] Setting Lambda 1 =λ 2 =2.2,λ 3 =15, μ=15. The filter optimization problem is determined according to formula (17). In order to solve the problem, formula (17-18) is processed by augmented Lagrangian function to obtain formula (19). In order to further reduce the complexity of the solution, the ADMM method is used to divide the solution problem into three subsystems containing two convex optimization problems, namely formula (20). Two of the optimization problems are solved alternately, and the third subsystem establishes the connection between the two optimization problems. When solving the filter of each frame image, the two optimization problems in formula (20) are iterated alternately for 3 times to obtain the final filter result. The filter is applied to the image, and the area with the maximum response is selected. The target position is greater than the screening threshold.

[0134] Step 6: Output of multi-scenario drone target tracking results

[0135] Combined with the tracking algorithms provided in steps 1 to 5, the videos containing the moving targets of the drone under different scenes and lighting conditions are used as the algorithm input, and the target position and size information of each frame image calculated by the tracking algorithm is output. Figure 4-5 The present invention is applied to the video of two groups of drones landing from the air to the ground. The video has the characteristics of the size change of the drones caused by scene switching and drone movement. The algorithm can ensure that the drones can be tracked normally in different scenes, and when the size of the drones changes from far to near, the tracking algorithm can still adapt to the size change of the drones to track the target; Figure 6 The UAV motion tracking effect of the present invention under special lighting conditions is shown. The present invention realizes UAV tracking when the target is relatively blurred relative to the background, which proves the robustness of the present invention.

Claims

1. A drone cluster "chase-escape" tracking method that simulates the visual perception of a falcon, characterized by: The steps of this method are as follows: Step 1: Pre-compression of images imitating the bifoveal receptive field of eagles and raptors; Combining the imaging mechanism of the deep and shallow double fovea structure of raptor vision, an image compression model was established; two pre-compression methods for original images were designed; S11, image pre-compression of the falcon vision shallow fovea model; S12, image pre-compression of the deep foveal model of hawk vision; Step 2: Image contrast adaptation and adjustment imitating the visual perception of a falcon; An effective model describing image contrast information is established, and based on the high contrast characteristics of falcon vision, it is mapped to the UAV target contrast adaptation adjustment model; Step 3: Extract image features by simulating the visual perception of hawks; Falcons have unique image processing mechanisms, including color antagonism, edge sensitivity, and positive and negative feedback of visual information loops. Combined with the feature extraction requirements of UAV target tracking tasks, a falcon-like perception feature extraction method and multi-feature control method are designed; S31, color and edge feature extraction imitating falcon vision; S32, Multi-feature regulation of falcon-like visual processing; Step 4: Adaptive adjustment of spatial mask of "chase-escape" dynamic vision simulating the visual perception of hawks; Step 5: Design of discriminant correlation filter tracker for UAV targets; Step 6: Filter tracker solution; According to the tracking method designed in steps one to six, after inputting the image sequence containing the UAV moving target and the real position of the target in the first frame, the position and size of the UAV target in the continuous image sequence can be estimated in multiple scenes to achieve tracking.

2. The method for rapid tracking of unmanned aerial vehicle targets based on falcon visual perception according to claim 1 is characterized in that: The specific process of step one is as follows: S11, image pre-compression of the falcon vision shallow fovea model; The shallow fovea in falcon vision has a large visual field and a central-peripheral receptive field structure. The image processing of this process is simulated by a Gaussian filter. The filter kernel of this process is expressed as: Among them, (x d ,y d ) is the distance between the pixel and the center of the shallow fovea, σ1 is the standard deviation of the two-dimensional Gaussian filter kernel that simulates the receptive field size of the ganglion cell; the pre-compression process is expressed as: Yo c,s (x,y)=I o (x,y)*G1(x d ,and d ,σ1) (2) Among them, I o (x, y) is the original image, * is the convolution operation, I c,s (x, y) is the pre-compression result after shallow foveal filter kernel processing; S12, image pre-compression of the deep foveal model of hawk vision; Based on the shallow fovea processing results, a core area with a side length of ρ0 is set, and the side length of the receptive field area is ρ m , pixels outside the core area in the receptive field are selectively compressed, and the radius of the candidate area is ρ n It is expressed as: r n =ρ0α n ,n=1,2,…,N ρ (3) Among them, ρ n is the side length of the rectangular pixel band of the candidate area centered on the candidate area in the nth layer of the original image, N ρ is the total number of pixel layers to be compressed, and α is the attenuation coefficient, which is defined as: At the same time, in order to ensure the integrity of the compressed image, the pixel sampling interval of each candidate rectangular pixel band during compression is defined as: Among them, Δρ n is the pixel sampling interval of the nth pixel band; n is the number of pixel band layers.

3. The method for rapid tracking of unmanned aerial vehicle targets based on falcon visual perception according to claim 1 is characterized in that: The specific process of step 2 is as follows: The regional contrast sensitivity function used to measure the contrast information in the image is defined as: Among them, I in (x, y) is the average brightness of the image in the pixel area, I out (x, y) is the average brightness of the image outside the pixel, and CS(x, y) is the regional contrast; The design of the falcon-like perception contrast adaptive image contrast feature extraction model is as follows: Among them, P in is the neighboring pixel in the area of ​​the pixel at position (x, y), I g (x, y) is the pixel grayscale value, mid(·) is the median calculation function, T cs,1 ,T cs,1 is the regional contrast threshold, I cs (x,y) is the result of adaptive feature extraction based on hawk-like perception contrast.

4. The method for rapid tracking of unmanned aerial vehicle targets based on falcon visual perception according to claim 1 is characterized in that: The specific process of S32 in step 3 is as follows: Establish a basic weight model based on image information entropy: Among them, w b (I k ) is the extracted feature vector I k The calculated basic weight, p i Indicates the proportion of significant pixels with gray value i; this basic weight represents the effectiveness of the information contained in the image; inspired by the positive and negative feedback effects of falcons chasing targets, the multi-feature control weight is designed as: Among them, w I (I k ) is the feature vector I k The multi-feature regulation weights of the visual processing of the simulated falcon chasing the fleeing target, w b,t (I k ) is the basic feature weight of the feature vector at time t, w b,t-1 (I k ) is the basic feature weight of the feature vector at time t-1, N c is the sum of all feature dimensions, including I cs ,I co ,I fHOG , norm is the normalization function.

5. The method for rapid tracking of unmanned aerial vehicle targets based on falcon visual perception according to claim 1 is characterized in that: The specific process of step 4 is as follows: Define the mask response variable k β1,k for: k β1,k :=||f k-1 *β k-1 x c,k (w I (I c ),I c )-y k || (10) Among them, f k-1 is the filter model of the k-1th frame discriminant filter tracker, which acts on the k-1th frame mask β k-1 and the k-th frame feature vector x c,k As the filtering result, y k To predict the results, define the mask response change rate: k β2,k :=k β1,k -k β1,k-1 (11) Among them, k β1,k-1 is the mask response of the k-1th frame; the mask update model is designed as: Among them, r1, r2, r3, r4 are mask coefficients to ensure β k The elements are between 0 and 1.

Images