A method for detecting targets in narrow underground spaces based on fly-like vision model
Through the object detection method based on the fly-like vision model, combined with motion information and color information, and using Bayesian model to detect moving targets, the problems of poor adaptability and insufficient reliability in complex environments are solved, and efficient and accurate target detection of small underground spaces is achieved.
Patent Information
- Application Number
- CN202411630491.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-15
- Publication Date
- 2025-04-04
AI Technical Summary
The existing underground narrow space object detection methods have poor adaptability in complex environments, difficult to quickly respond to light changes, obstacle occlusion and complex backgrounds, and lack reliability in extreme environments, making them prone to failures or false alarms.
The object detection method based on the fly-like vision model is adopted, and by constructing the motion information processing path and the color information processing path, the visual processing mechanism of fruit fly is simulated, the motion information and color information are integrated, and the motion target is detected using the Bayesian model.
It realizes efficient, accurate and stable object detection in a small underground space, adapts to complex environments, and improves detection accuracy and system reliability.
Smart Images
Figure CN119762742A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of digital image processing, and in particular relates to a method for detecting targets in underground narrow spaces based on a fly-like vision model. Background Art
[0002] With the development of science and technology and the advancement of urban modernization in my country, the utilization of urban underground space, such as subway construction and underground pipelines, is developing rapidly. For the utilization of underground space, due to insufficient light, limited space, many obstacles, etc., the underground space images obtained often have great limitations, the image brightness is low, and the conditions such as shooting target angle are harsh. The design, construction and monitoring of underground space all require accurate target recognition support. Through target detection, objects in underground space can be accurately identified, the accuracy of underground space detection can be improved, which is conducive to more sophisticated space design and construction, and enhances underground risk monitoring and maintenance, emergency response capabilities, etc. Therefore, underground narrow space target detection can not only accurately detect target information, but also has important significance for the safety, accuracy and intelligent management of underground projects.
[0003] Due to the narrow environment and weak light in underground spaces, target detection poses great challenges. The compound eye structure of fruit flies gives them a wide field of view and extremely high dynamic sensitivity, allowing them to capture fast-moving objects over a large area. The fly-like vision model attempts to imitate this characteristic to achieve fast and efficient target detection and tracking.
[0004] In the previous underground space target detection methods, there are methods based on traditional computer vision, deep learning, and multiple sensor fusion methods. These methods perform well in static or simple dynamic environments, but they have poor adaptability to underground lighting changes, obstacle occlusion, and complex backgrounds. The installation and operation of their hardware equipment are limited in the small underground space. The targets and environments in the underground space, such as mobile devices and personnel activities, may change rapidly, and traditional algorithms are difficult to adapt and respond quickly. These detection methods may fail in complex dynamic environments. At the same time, due to the limitations of the small underground space, it is difficult to achieve complex hardware configuration and regular maintenance of high-precision target detection systems. The reliability of traditional target detection systems in extreme environments is relatively insufficient, and they are prone to failures or false alarms, affecting the effect of safety monitoring. Compared with these methods, the target detection method based on the bionics simulation of the fruit fly visual system has a wide field of view, high dynamic sensitivity, and is more adaptable to complex environments. However, in previous motion detection models based on fruit fly vision, such as the LGMD model, only the motion information processing process is often considered, which leads to unclear edge features of moving targets and difficulty in distinguishing foreground and background in some complex scenes. Therefore, inventing a target detection and recognition method with strong environmental adaptability and high detection accuracy is of great significance for target detection in narrow underground spaces. Summary of the invention
[0005] In order to solve the above problems, the present invention provides a method for detecting targets in narrow underground spaces based on a fly-like vision model, which can perform target detection in narrow underground spaces more efficiently, accurately and stably.
[0006] The present invention is specifically a method for detecting targets in underground narrow spaces based on a fly-like vision model, and the method for detecting targets in underground narrow spaces comprises the following steps:
[0007] S1. Construct motion information processing pathway, simulate the Drosophila motion information processing mechanism to construct retinal layer, lamina layer, medulla layer and lobule layer, use scale invariance algorithm to extract image feature points, construct lateral inhibition matrix, generate specific lateral inhibition direction and integrate to get the approximate target area;
[0008] S2, construct a color information processing pathway, simulate the retinal layer to build mutually opposing photoreceptors, pre-process the input information through Gaussian filtering, and imitate the receptive field of medullary layer neurons to extract the edge information of the moving target;
[0009] S3. Construct a central complex computational model to integrate motion information and color information, and detect moving targets through a Bayesian model.
[0010] S1 specifically includes:
[0011] S11. Construct a retina layer structure model, simulate the characteristics of fruit fly vision that is sensitive to the red spectrum, select a single channel (R channel) for the RGB image sequence or video, and combine the retina layer scale-invariant feature transform SIFT with the inter-frame difference. Preliminary motion information is obtained through calculation, and the motion information P t The calculation formula for (x,y) is as follows:
[0012]
[0013] Where L t (x,y) represents the image input stream of time step t, x and y represent the spatial position information, is the convolution operation, and H is the transformation matrix calculated based on the scale invariant feature transformation:
[0014]
[0015] where λ i is the homogeneous coordinate of the scale invariant feature transformation, a is the scale factor, A 2×2 , T 2×1 is the translation vector.
[0016] S12, build a thin plate Lamina layer structure model, and calculate the lateral inhibition signal I of the four directions of the previous time step to the current area according to the inhibition signals in the left (L), right (E), upper (U), and lower (D) directions respectively t , and its calculation formula is as follows:
[0017]
[0018] Where q and s determine the size of the lateral inhibition matrix template and q and s are not equal, ω L ,ω E ,ω U and ω D is a template matrix that plays a lateral inhibitory role, such as ω L The left half of the elements in the left inhibition matrix are 1, and the rest are 0. The same is true for the inhibition template matrices in the other three directions.
[0019] S13. Construct a medulla layer structure model. In the medulla layer, four basic directions will be generated: from left to right, from right to left, from top to bottom, and from bottom to top. Direction selection is achieved by modifying the angle direction of T cell lateral inhibition in the medulla layer. The four basic directions T t The calculation formula is as follows:
[0020]
[0021] Where W is the global suppression weight, [] * The half-wave rectification operation is represented as max(0,x).
[0022] S14. Construct a computational model of the lobula layer, integrate the motion characteristics of the four basic directions and divide them into horizontal direction (HS) and vertical direction (VS) according to their preferred directions to realize the perception of the moving target. The lobula layer integrates the direction signal transmitted from the medulla layer to calculate the position signal S t (x,y) is as follows:
[0023]
[0024] S2 specifically includes:
[0025] S21. Construct a retina layer structure model, divide the input stream image at the current moment into four channels: red (R), green (G), blue (B) and yellow (Y), where Y = (R + G) / 2, and then pre-process the four channel information by simulating the receptive field of the retina layer through Gaussian filtering. The calculation formula is as follows:
[0026]
[0027] Among them, R(x,y), G(x,y), B(x,y) and Y(x,y) are the red, green, blue and yellow channel information at the position (x,y). and are the pre-processed information of the red, green, blue, and yellow channels of the retinal layer at position (x, y), is the convolution operation, and g(x,y) is the Gaussian filter.
[0028] The color confrontation mechanism is simulated through the red-green channel (RG) and the blue-yellow channel (BY) to obtain the output information Q1(x, y), Q2(x, y). The calculation formula is as follows:
[0029]
[0030] Among them, ω1 and ω2 are inhibition weights, and ω1 and ω2 are always opposite in positive and negative, playing a role of mutual inhibition.
[0031] S22, construct the medulla layer structure model, simulate the color-sensitive neuron receptive field that is opposite in color and space, Q1(x, y), Q2(x, y) obtains the color sensitivity information O(x, y) by adversarial neuron receptive field, θ i The response of the medulla at direction (x,y) is O(x,y,θ i ) is calculated as follows:
[0032]
[0033] in, The value of i ranges from 1 to a, and a is usually 16. M and N constrain the receptive field area of the medulla, F(τ,κ,θ i ) is the color receptive field, and its calculation formula is as follows:
[0034]
[0035] in, is the spatial receptive field, and its calculation formula is as follows:
[0036]
[0037] Among them, kδ is the standard deviation of Gaussian filtering, γ is the aspect ratio of Gaussian field space, is the result of τ and κ after the rotation matrix transformation, and the calculation formula is as follows:
[0038]
[0039] Finally, the color sensitivity information O(x, y) extracted by the color information processing channel at the position (x, y) is calculated as follows:
[0040]
[0041] S3 specifically includes:
[0042] S31. Construct a central complex calculation model, and use the texture and edge information obtained by the color information processing pathway to calculate the prior information based on the contour edge.
[0043] S32. Predict the size of the potential structure based on the motion path detection result, binarize the motion detection result to capture the rough potential area and size, and calculate the conditional probability through the ratio of the background and foreground after binarization.
[0044] S33, the possible position of the potential structure estimated by the motion information processing pathway and the prior information based on the contour edge are input into the Bayesian model, and the posterior probability p(s|x) of the pixel point x belonging to the significant structure is calculated based on the Bayesian inference, and the calculation formula is as follows:
[0045]
[0046] Where p(b) = 1-p(s) and p(s) are the prior probabilities that the pixel x belongs to the background and the salient structure, and p(x|s) and p(x|b) are conditional probabilities. The edge-based prior probability is used as the initial prior probability p(s), the salient structure is iteratively strengthened, and the obtained posterior probability p(x|s) is reinitialized and assigned to the prior probability p(s).
[0047] S31 specifically includes:
[0048] S311, the color channel has processed and obtained the edge information of the image.
[0049] S312. For each edge pixel, there are two half disks, "left" and "right", and the average edge response of the left and right half disks of the edge pixel is calculated. The disk is defined by the direction of each edge pixel, and its radius is d. Considering that the radius d has a strong correlation with the image size, viewing distance, etc., d is selected as a relative value of the image size. Set d = min(V, H) / 3, where V, H are the width and height of the given image.
[0050] S313, all pixels in the half disk with stronger edge response between the two half disks are voted to 1, and the pixels in the other half disk are voted to 0. The rough spatial weight of the significance of each pixel is determined by voting of all edge pixels, and S e represents the rough spatial weight of significance, S c represents the central bias weight.
[0051] S314, to S e and S cNormalize and get the final prior information. The calculation formula is as follows:
[0052] p=S e +S c (14)
[0053] Compared with the prior art, the beneficial effects are:
[0054] (1) The detection method of the present invention is different from traditional computer vision, deep learning and other methods. Instead, it detects targets in images and videos based on a fly-like visual model, extracts and integrates the target's motion information and color information, and then predicts the target's structure size and position information. This method has a small and complex visual unit, can capture fast-moving objects in a large range, is better adapted to the narrow space environment underground, is not easily affected by noise, and improves the accuracy of target detection.
[0055] (2) The detection method of the present invention extracts the target motion information through the motion information processing pathway to predict the size of its potential structure, and extracts the target color sensitivity information through the color information processing pathway to obtain the target edge information. This makes up for the defects that the edge features of the moving target are not obvious and the foreground and background are difficult to distinguish in some complex scenes, and lays the foundation for the subsequent central complex integration calculation.
[0056] (3) The detection method of the present invention fuses the moving target area and image edge information through the central complex calculation model, uses the output of the Bayesian model as new prior information, and refines the saliency structure through an iterative process, thereby obtaining accurate target detection information and having high applicability and detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0057] Figure 1 This is a flow chart of a method for detecting targets in a narrow underground space based on a fly-like visual model according to the present invention;
[0058] Figure 2 This is the structure diagram of the La layer of the thin plate layer proposed by the present invention;
[0059] Figure 3 This is a calculation structure diagram of the medulla layer and lobule layer proposed in the present invention. DETAILED DESCRIPTION
[0060] The specific implementation of the method for detecting targets in a narrow underground space based on a fly-like vision model of the present invention is described in detail below with reference to the accompanying drawings.
[0061] like Figure 1 As shown, a method for detecting targets in underground narrow spaces based on a fly-like visual model of the present invention comprises the following steps:
[0062] S1. Construct motion information processing pathway, simulate the fruit fly motion information processing mechanism to construct retinal layer, lamina layer, medulla layer and lobule layer, use scale invariance algorithm to extract image feature points, construct lateral inhibition matrix, generate specific lateral inhibition direction and integrate to get the approximate target area.
[0063] S2. Construct a color information processing pathway, simulate the retinal layer to build opposing photoreceptors, pre-process the input information through Gaussian filtering, and imitate the receptive field of medullary layer neurons to extract the edge information of the moving target.
[0064] S3. Construct a central complex computational model to integrate motion information and color information, and detect moving targets through a Bayesian model.
[0065] In S1, a motion information processing pathway is constructed to predict the size of potential target structures, which mainly includes the following steps:
[0066] S11. Construct a retina layer structure model, simulate the characteristics of fruit fly vision that is sensitive to the red spectrum, select a single channel (R channel) for the RGB image sequence or video, and combine the retina layer scale-invariant feature transform SIFT with the inter-frame difference. Preliminary motion information is obtained through calculation, and the motion information P t The calculation formula for (x,y) is as follows:
[0067]
[0068] Where L t (x,y) represents the image input stream of time step t, x and y represent the spatial position information, is the convolution operation, and H is the transformation matrix calculated based on the scale invariant feature transformation:
[0069]
[0070] where λ i is the homogeneous coordinate of the scale invariant feature transformation, a is the scale factor, A 2×2 , T 2×1 is the translation vector.
[0071] S12. Combination Figure 2 , construct the thin plate Lamina layer structure model, and according to the inhibition signals in the left (L), right (E), upper (U), and lower (D) directions, the lateral inhibition signals I of the four directions in the previous time step to the current area can be calculated respectively t , and its calculation formula is as follows:
[0072]
[0073] Where q and s determine the size of the lateral inhibition matrix template and q and s are not equal, ω L ,ω E ,ω U and ω D is a template matrix that plays a lateral inhibitory role, such as ω L The left half of the elements in the left inhibition matrix are 1, and the rest are 0. The same is true for the inhibition template matrices in the other three directions.
[0074] S13, Combination Figure 3 , construct the medulla layer structure model, in the medulla layer will produce four basic directions from left to right, from right to left, up to down and down to up, by modifying the angle direction of T cell lateral inhibition in the medulla layer to achieve direction selection, the four basic directions T t The calculation formula is as follows:
[0075]
[0076] Where W is the global suppression weight, [] * The half-wave rectification operation is represented as max(0,x).
[0077] S14. Construct a computational model of the lobula layer, integrate the motion characteristics of the four basic directions and divide them into horizontal direction (HS) and vertical direction (VS) according to their preferred directions to realize the perception of the moving target. The lobula layer integrates the direction signal transmitted from the medulla layer to calculate the position signal S t (x,y) is as follows:
[0078]
[0079] In S2, a color information processing pathway is constructed to extract color sensitivity information and obtain target edge information, which mainly includes the following steps:
[0080] S21. Construct a retina layer structure model, divide the input stream image at the current moment into four channels: red (R), green (G), blue (B) and yellow (Y), where Y = (R + G) / 2, and then pre-process the four channel information by simulating the receptive field of the retina layer through Gaussian filtering. The calculation formula is as follows:
[0081]
[0082] Among them, R(x,y), G(x,y), B(x,y) and Y(x,y) are the red, green, blue and yellow channel information at the position (x,y). and are the pre-processed information of the red, green, blue, and yellow channels of the retinal layer at position (x, y), is the convolution operation, and g(x,y) is the Gaussian filter.
[0083] The color confrontation mechanism is simulated through the red-green channel (RG) and the blue-yellow channel (BY) to obtain the output information Q1(x, y), Q2(x, y). The calculation formula is as follows:
[0084]
[0085] Among them, ω1 and ω2 are inhibition weights, and ω1 and ω2 are always opposite in positive and negative, playing a role of mutual inhibition.
[0086] S22, construct the medulla layer structure model, simulate the color-sensitive neuron receptive field that is opposite in color and space, Q1(x, y), Q2(x, y) obtains the color sensitivity information O(x, y) by adversarial neuron receptive field, θ i The response of the medulla at direction (x,y) is O(x,y,θ i ) is calculated as follows:
[0087]
[0088] in, The value of i ranges from 1 to a, and a is usually 16. M and N constrain the receptive field area of the medulla, F(τ,κ,θ i ) is the color receptive field, and its calculation formula is as follows:
[0089]
[0090] in, is the spatial receptive field, and its calculation formula is as follows:
[0091]
[0092] Among them, kδ is the standard deviation of Gaussian filtering, γ is the aspect ratio of Gaussian field space, is the result of τ and κ after the rotation matrix transformation, and the calculation formula is as follows:
[0093]
[0094] Finally, the color sensitivity information O(x, y) extracted by the color information processing channel at the position (x, y) is calculated as follows:
[0095]
[0096] In S3, the central complex calculation model mainly includes the following steps:
[0097] S31. Construct a central complex calculation model, and use the texture and edge information obtained by the color information processing pathway to calculate the prior information based on the contour edge.
[0098] S32. Predict the size of the potential structure based on the motion path detection result, binarize the motion detection result to capture the rough potential area and size, and calculate the conditional probability through the ratio of the background and foreground after binarization.
[0099] S33, the possible position of the potential structure estimated by the motion information processing pathway and the prior information based on the contour edge are input into the Bayesian model, and the posterior probability p(s|x) of the pixel point x belonging to the significant structure is calculated based on the Bayesian inference, and the calculation formula is as follows:
[0100]
[0101] Where p(b) = 1-p(s) and p(s) are the prior probabilities that the pixel x belongs to the background and the salient structure, and p(x|s) and p(x|b) are conditional probabilities. The edge-based prior probability is used as the initial prior probability p(s), the salient structure is iteratively strengthened, and the obtained posterior probability p(x|s) is reinitialized and assigned to the prior probability p(s).
[0102] In step S31, edge-based prior information is calculated, which mainly includes the following steps:
[0103] S311, the color channel has processed and obtained the edge information of the image.
[0104] S312. For each edge pixel, there are two half disks, "left" and "right", and the average edge response of the left and right half disks of the edge pixel is calculated. The disk is defined by the direction of each edge pixel, and its radius is d. Considering that the radius d has a strong correlation with the image size, viewing distance, etc., d is selected as a relative value of the image size. Set d = min(V, H) / 3, where V, H are the width and height of the given image.
[0105] S313, all pixels in the half disk with stronger edge response between the two half disks are voted to 1, and the pixels in the other half disk are voted to 0. The rough spatial weight of the significance of each pixel is determined by voting of all edge pixels, and S e represents the rough spatial weight of significance, S c represents the central bias weight.
[0106] S314, to S e and S c Normalize and get the final prior information. The calculation formula is as follows:
[0107] p=S e +S c(14)
[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and are not intended to limit the present invention. A person skilled in the art should understand that the specific implementation of the present invention can be modified or replaced by equivalents, but these modifications or changes are within the scope of protection of the pending claims.
Claims
1. A method for detecting targets in underground narrow spaces based on a fly-like visual model, characterized in that: The following steps are involved: S1. Construct motion information processing pathway, simulate the Drosophila motion information processing mechanism to construct retinal layer, lamina layer, medulla layer and lobule layer, use scale invariance algorithm to extract image feature points, construct lateral inhibition matrix, generate specific lateral inhibition direction and integrate to get the approximate target area; S2, construct a color information processing pathway, simulate the retinal layer to build mutually opposing photoreceptors, pre-process the input information through Gaussian filtering, and imitate the receptive field of medullary layer neurons to extract the edge information of the moving target; S3. Construct a central complex computational model to integrate motion information and color information, and detect moving targets through a Bayesian model.
2. The method for detecting targets in underground narrow spaces based on a fly-like visual model according to claim 1, characterized in that: S1 specifically includes: S11. Construct a retina layer structure model, simulate the characteristics of fruit fly vision that is sensitive to the red spectrum, select a single channel R channel for the RGB image sequence or video, and combine the retina layer scale-invariant feature transform SIFT with the inter-frame difference; obtain preliminary motion information through calculation, motion information P t The calculation formula for (x,y) is as follows: Where L t (x,y) represents the image input stream of time step t, x and y represent the spatial position information, is the convolution operation, and H is the transformation matrix calculated based on the scale invariant feature transformation: where λ i is the homogeneous coordinate of the scale invariant feature transformation, a is the scale factor, A 2×2 , T 2×1 is the translation vector; S12, build a thin plate Lamina layer structure model, and calculate the lateral inhibition signal I of the four directions of the previous time step to the current area according to the inhibition signals in the four directions of left (L), right (E), top (U), and bottom (D) t , and its calculation formula is as follows: Where q and s determine the size of the side inhibition matrix template, and q and s are not equal, ω L ,ω E ,ω U and ω D is the template matrix that plays a lateral inhibitory role; S13. Construct a medulla layer structure model. In the medulla layer, four basic directions will be generated: from left to right, from right to left, from top to bottom, and from bottom to top. Direction selection is achieved by modifying the angle direction of T cell lateral inhibition in the medulla layer. The calculation formulas for the four basic directions are as follows: Where W is the global suppression weight, [] * The half-wave rectification operation is represented by max(0,x); S14. Construct a computational model of the lobula layer, integrate the motion characteristics of the four basic directions and divide them into horizontal direction (HS) and vertical direction (VS) according to their preferred directions to realize the perception of the moving target. The lobula layer integrates the direction signal transmitted from the medulla layer to calculate the position signal S t (x,y) is as follows:
3. The method for detecting targets in underground narrow spaces based on a fly-like visual model according to claim 2, characterized in that: S2 specifically includes: S21. Construct a retina layer structure model, divide the input stream image at the current moment into four channels: red (R), green (G), blue (B) and yellow (Y), where Y = (R+G) / 2, and then pre-process the four channel information by simulating the receptive field of the retina layer through Gaussian filtering, and simulate the color confrontation mechanism through the red-green channel (RG) and the blue-yellow channel (BY) to obtain the output information Q1 (x, y), Q2 (x, y); S22. Construct a medulla layer structure model to simulate the receptive fields of color-sensitive neurons that are opposed in color and space. Q1(x, y) and Q2(x, y) obtain color sensitivity information O(x, y) by adversarial neuron receptive fields. The calculation formula is as follows:
4. The method for detecting targets in underground narrow spaces based on a fly-like visual model according to claim 3, characterized in that: S3 specifically includes: S31, constructing a central complex calculation model, and calculating the prior information based on the contour edge using the texture and edge information obtained by the color information processing pathway; S32, predicting the size of the potential structure according to the motion path detection result, binarizing the motion detection result to capture the rough potential area and size, and calculating the conditional probability through the ratio of the background and foreground after binarization; S33, the possible position of the potential structure estimated by the motion information processing pathway and the prior information based on the contour edge are input into the Bayesian model, and the posterior probability p(s|x) of the pixel point x belonging to the significant structure is calculated based on the Bayesian inference, and the calculation formula is as follows: Where p(b) = 1-p(s) and p(s) are the prior probabilities that the pixel x belongs to the background and the salient structure, and p(x|s) and p(x|b) are conditional probabilities. The edge-based prior probability is used as the initial prior probability p(s), the salient structure is iteratively strengthened, and the obtained posterior probability p(x|s) is reinitialized and assigned to the prior probability p(s).