CN112613427B - Road obstacle detection method based on visual information flow partition projection coding model - Google Patents
Road obstacle detection method based on visual information flow partition projection coding model Download PDFInfo
- Publication number
- CN112613427B CN112613427B CN202011578651.0A CN202011578651A CN112613427B CN 112613427 B CN112613427 B CN 112613427B CN 202011578651 A CN202011578651 A CN 202011578651A CN 112613427 B CN112613427 B CN 112613427B
- Authority
- CN
- China
- Prior art keywords
- response
- map
- contour
- formula
- azimuth
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 230000000007 visual effect Effects 0.000 title claims abstract description 39
- 238000005192 partition Methods 0.000 title claims abstract description 17
- 238000001514 detection method Methods 0.000 title claims abstract description 11
- 230000004044 response Effects 0.000 claims abstract description 57
- 239000011159 matrix material Substances 0.000 claims abstract description 19
- 238000011176 pooling Methods 0.000 claims abstract description 14
- 230000007246 mechanism Effects 0.000 claims abstract description 10
- 210000000857 visual cortex Anatomy 0.000 claims abstract description 8
- 230000008447 perception Effects 0.000 claims abstract description 7
- 238000000605 extraction Methods 0.000 claims abstract description 6
- 230000009123 feedback regulation Effects 0.000 claims abstract description 6
- 210000004556 brain Anatomy 0.000 claims abstract description 4
- 238000000034 method Methods 0.000 claims description 27
- 230000003042 antagnostic effect Effects 0.000 claims description 14
- 230000004888 barrier function Effects 0.000 claims description 8
- 230000003044 adaptive effect Effects 0.000 claims description 6
- 230000000295 complement effect Effects 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 6
- 238000010606 normalization Methods 0.000 claims description 6
- 230000008485 antagonism Effects 0.000 claims description 5
- 238000010586 diagram Methods 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 5
- 230000009471 action Effects 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 4
- 230000008569 process Effects 0.000 claims description 4
- 238000007670 refining Methods 0.000 claims description 4
- 230000002708 enhancing effect Effects 0.000 claims description 3
- 230000004927 fusion Effects 0.000 claims description 3
- 230000001629 suppression Effects 0.000 claims description 3
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 claims description 2
- 239000000284 extract Substances 0.000 claims description 2
- 239000002346 layers by function Substances 0.000 claims description 2
- 238000013178 mathematical model Methods 0.000 claims description 2
- 239000010410 layer Substances 0.000 description 7
- 210000004027 cell Anatomy 0.000 description 3
- 230000035945 sensitivity Effects 0.000 description 3
- 208000003098 Ganglion Cysts Diseases 0.000 description 2
- 208000005400 Synovial Cyst Diseases 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 241000287196 Asthenes Species 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000004069 differentiation Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000005764 inhibitory process Effects 0.000 description 1
- 238000000691 measurement method Methods 0.000 description 1
- 210000005036 nerve Anatomy 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 239000011664 nicotinic acid Substances 0.000 description 1
- 210000000977 primary visual cortex Anatomy 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
- G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/56—Extraction of image or video features relating to colour
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/60—Extraction of image or video features relating to illumination properties, e.g. using a reflectance or lighting model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
- G06V20/54—Surveillance or monitoring of activities, e.g. for recognising suspicious objects of traffic, e.g. cars on the road, trains or boats
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/513—Sparse representations
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to a road obstacle detection method based on a visual information flow partition projection coding model. Firstly, constructing a parallel vision path model in a V1 topological projection profile global perception unit, and extracting the brightness edge and the color edge of a road traffic map in parallel to obtain a topological projection profile map representing the overall characteristics of an obstacle and an optimal azimuth index matrix; then, a visual information difference enhancement model is constructed in a V4 sparse coding fine feature extraction unit, and the position response difference and the azimuth response difference are utilized to carry out contrast enhancement on the topological projection profile to obtain a refined profile; thirdly, providing a self-adaptive size sparse coding model, realizing intelligent focusing on the contour features of the obstacle according to the sparseness of the thinned contour map, and obtaining a pooling map representing the local features of the obstacle; and finally, simulating feedback regulation and control mechanisms of different brain regions of the visual cortex, correcting the topological projection contour map of the cross-visual area by using the pooling map, and fusing to obtain a final obstacle contour appearance result.
Description
Technical Field
The invention belongs to the field of machine vision, and particularly relates to a road obstacle detection method based on a visual information flow partition projection coding model.
Background
Road obstacle detection tasks are one of the subtasks of intelligent traffic systems, and for vehicles traveling at high speeds, obstacles can be classified into various forms such as pedestrians, road traffic signs, mountain falling stones, vehicle spills, and the like. Road obstacle detection is of great importance for safe driving of vehicles and efficient management of traffic. Contours are an efficient representation of the target subject as low-dimensional spatial features of the image. The obstacle outline acquisition in the road traffic map is beneficial to improving the rapidity and accuracy of tasks such as subsequent obstacle classification, recognition and the like.
The contours detected by the traditional image processing technology adopting the mathematical differentiation operator often contain a large amount of texture information, and road traffic target barriers cannot be well distinguished from complex backgrounds. Along with the great emergence of research results in the field of biological vision and the rapid development of nerve computation, a target perception method based on a bionic mechanism is widely focused. The sensitivity characteristics of the classical receptive field and the edge line segments of the primary visual cortex are simulated by a two-dimensional Gabor energy model, the physiological structure of the non-classical receptive field is simulated by a Gaussian differential model, an isotropy and anisotropy inhibition method is provided, and texture information is effectively inhibited; in addition, the modeling of the color antagonism mechanism in the video path is studied, so that the effective extraction of the contour information of the color image is realized. And then further suppressing texture information by using an improved sparsity measurement method on the basis. The contour detection method models the processing process of the information flow in the visual path by utilizing different biological vision mechanisms, but simplifies the information flow projection process in the visual path, simplifies the processing effect of the advanced visual cortex on the visual information, performs black box simulation on the physiological characteristics of neuron information coding or feedback regulation and the like only on the basis of single visual path modeling, and is unfavorable for effectively extracting the obstacle target because the target usually does not have a structured contour shape for the road obstacle target under a complex background.
Disclosure of Invention
The invention provides a road obstacle detection method based on a visual information flow partition projection coding model.
The model provided by the invention is composed of a V1 topological projection contour global perception unit and a V4 sparse coding fine feature extraction unit, and respectively simulates the front-stage characteristic of a V1 layer and the coding characteristic of the V4 layer. Firstly, constructing a parallel view path model in a V1 topological projection profile global perception unit, extracting the brightness edge and the color edge of a road traffic map in parallel, and acquiring a topological projection profile E (x, y) representing the overall characteristics of an obstacle and an optimal azimuth index matrix Θ (x, y) in a V1 layer; then constructing a visual information difference enhancement model in a V4 sparse coding fine feature extraction unit, and respectively performing texture suppression and contour enhancement on the topological projection contour map E (x, y) by using the position response difference and the azimuth response difference to obtain a refined contour map E t (x, y); then put forward aSeed adaptive size sparse coding model according to refined contour map E t The sparseness degree of (x, y) realizes intelligent focusing on the outline features of the obstacle, and a pooling graph E for representing the local features of the obstacle is obtained at a V4 layer s (x, y); finally, simulating feedback regulation and control mechanisms of different brain regions of visual cortex, and utilizing pooling graph E s And (x, y) correcting the topological projection profile map E (x, y) of the cross-view area, and fusing to obtain a final barrier profile result.
Compared with the prior art, the invention has the following effects:
the invention constructs a parallel vision path model, and simulates the information flow partition projection characteristic to primarily sense the outline of the road obstacle. In consideration of the detail sensitivity difference of M-type ganglion cells and P-type ganglion cells to brightness information flow and color information flow and the partition projection characteristic of two types of visual information flow in a primary visual path, a parallel visual path model is constructed at the front stage of a V1 layer, the brightness edge and the color edge of a road traffic map are respectively extracted, primary contour response is obtained by fusion, and the model ensures the integrity of obstacle contour information perception in the road traffic map.
The invention provides a visual information difference enhancement model, which utilizes the physiological characteristics of a visual receptive field to carry out contrast enhancement on primary contour response. Considering the azimuth sensitivity characteristic of a classical receptive field, calculating azimuth response difference by utilizing the optimal azimuth information and orthogonal azimuth information corresponding to the primary profile response to enhance the profile pixel point; and calculating the position response difference to restrain the texture pixel point by using the response difference between the classical receptive field and the non-classical receptive field. The model refines the primary profile response by enhancing contrast.
The invention provides a novel sparse coding method for adaptively selecting sparse kernel sizes. In consideration of rapidity and accuracy of the visual nerve center in sensing the spatial position of the target main body, a sparse coding model with adaptively selected sparse kernel size is constructed, the adaptively selected sparse kernel size is carried out according to the sparse degree of pixel distribution in the refined contour map, and information redundancy is removed while intelligent focusing is completed on the contour features of the obstacle. Compared with the traditional method for sparse coding by adopting a fixed-size sparse kernel, the self-adaptive size sparse coding method provided by the invention is more in line with the dynamic focusing characteristic of a biological vision system on the target contour characteristic.
A road obstacle detection method based on an information flow partition projection coding model is provided. Firstly, simulating partition projection characteristics of different information flows, constructing a parallel vision path model, respectively extracting brightness edges and color edges of a road traffic map, and fusing in a V1 layer to obtain a primary contour response; then, constructing a visual information difference enhancement model by utilizing the physiological characteristics of the visual receptive field, and refining the primary contour response by using the position response difference and the azimuth response difference; then simulating dynamic focusing characteristics of the visual nerve center on the target contour features, constructing a self-adaptive size sparse coding model, and obtaining a pooling graph; and finally, simulating a feedback regulation mechanism between visual layers, correcting the primary contour response of the cross-visual area by using the pooling graph, and fusing to obtain a final obstacle contour result.
Drawings
Fig. 1 is a flowchart of a road obstacle detection method according to the present invention.
FIG. 2 is a schematic diagram of the optimal orientation and orthogonal orientation of the receptive field.
The specific embodiment is as follows:
note that: taking E (x, y) as an example, where (x, y) represents the two-dimensional coordinate position of the pixel in the image E, E (x, y) represents the pixel value at the coordinate position (x, y) in the image E, and the steps will not be described.
A specific embodiment of the present invention will be described with reference to fig. 1 and 2.
And (1) constructing a parallel vision path model, and extracting a topological projection profile E (x, y) and an optimal azimuth index matrix theta (x, y). For the road traffic diagram to be detected, the brightness component I (x, y) and the red, green and blue color components R (x, y), G (x, y) and B (x, y) are decomposed, and the number of rows and columns of each component are respectively m and n. The parallel view path model is composed of a brightness path and a color path, and extracts the brightness edge and the color edge of the road traffic map respectively.
Step 1.1 the mathematical model of the luminance path is shown in formula (1):
wherein sigma and gamma represent the size and ellipticity of classical receptive fields, and the default values are set to 2 and 0.5, respectively.
Representing selective orientations, defaulting to 8 orientations equally spaced, i.e. θ i = {0 °,45 °,90 °,135 °,180 °,225 °,270 °,315 ° }, which represents convolution operation, e (x, y; θ i ) Representative pixel coordinates are (x, y), selective orientation is θ i The corresponding luminance component edge response.
For each pixel, selecting the maximum value of the brightness component edge response corresponding to all the selective orientations, linearly normalizing the maximum value to be used as output, and simultaneously recording the corresponding optimal orientation index to obtain the contour response E of the brightness path L (x, y) and an optimal azimuth index matrix Θ L (x, y) as shown in formula (2):
where N (-) represents the linear normalization operation.
Step 1.2 the color channels are modeled using a color antagonistic mechanism, and are divided into four types of antagonistic channels, R-on/G-off, G-on/R-off, B-on/Y-off, Y-on/B-off, wherein the yellow component Y (x, Y) = (R (x, Y) +G (x, Y))/2. The R-on/G-off type antagonistic channel is exemplified. Firstly, simulating the action of cone cells, processing R (x, y) component and G (x, y) component by a Gaussian filter (default variance is set to be 1), and marking the result asThen calculating the edge response of the single antagonistic receptive field, and marking the result as S RG (x, y) as shown in formula (3):
then simulating the action of the double antagonistic receptive fields, calculating the position of pixel coordinates (x, y) and the selective azimuth is theta i Corresponding color component edge response d RG (x,y；θ i ) As shown in formula (4):
for each pixel, selecting the maximum value of the edge response of the color component corresponding to all the selective orientations, linearly normalizing the maximum value to be used as output, and simultaneously recording the corresponding optimal orientation index to obtain the edge response D of the R-on/G-off type antagonism channel RG (x, y) and an optimal azimuth index matrix Θ RG (x, y) as shown in formula (5):
similar to the R-on/G-off type antagonistic channels, the edge response D of the other three types of antagonistic channels is calculated GR 、D BY 、D YB And an optimal azimuth index matrix Θ GR 、Θ BY 、Θ YB Or by calculation. For each pixel coordinate position, taking the maximum value of edge responses in four antagonistic channels as output, and simultaneously recording the corresponding optimal azimuth index to obtain the contour response E of the color channel C (x, y) and an optimal azimuth index matrix Θ C (x, y) as shown in formula (6):
and 1.3, respectively fusing the contour response of the brightness path and the color path and the optimal azimuth index matrix, and simulating the front-stage characteristic of the visual cortex V1 functional layer to obtain a topological projection contour map E (x, y) and the optimal azimuth index matrix theta (x, y). As shown in formula (7):
step (2) constructing a visual information difference enhancement model, and carrying out contrast enhancement on the topological projection profile E (x, y) by utilizing the position response difference and the azimuth response difference to obtain a refined profile E t (x, y). First, for the topology projection profile E (x, y) obtained in step (1), a Gaussian function G (x, y; σ) and a Gaussian difference function DoG are used, respectively + (x, y; sigma) convolving with E (x, y) to obtain classical receptive field visual input L C (x, y; sigma) and non-classical receptive field visual input L N (x, y; sigma) as shown in formula (8):
wherein,
the default value of sigma in the formula (8) is the same as that in the formula (1), and L in the formula C (x, y; sigma) and L N (x, y; sigma) performing a difference operation to obtain a position response difference DeltaL (x, y; sigma) as shown in formula (10):
ΔL(x,y；σ)＝max{L C (x,y；σ)-L N (x,y；σ),0} (10)
then, an orthogonal azimuth index matrix Θ of the topology projection profile E (x, y) is calculated + (x, y) as shown in formula (11):
Θ + (x,y)＝(Θ(x,y)+2)mod 8 (11)
where mod represents the remainder taking operation; the theta (x, y) obtained in the step (1) stores the optimal azimuth information at the pixel coordinates (x, y) in the form of indexes, wherein the value range is {0,1,2,3,4,5,6,7}, {0,4} represents the horizontal optimal azimuth, {2,6 represents a vertical optimum orientation, {1,5} represents a forward optimum orientation, {3,7} represents a reverse optimum orientation; theta (theta) + (x, y) represents an orthogonal orientation complementary to the optimal orientation Θ (x, y), wherein the horizontal orientation is complementary to the vertical orientation and the left diagonal orientation is complementary to the right diagonal orientation. A schematic relationship of the optimal orientation to the orthogonal orientation is shown in fig. 2.
For each pixel coordinate (x, y) in the topology projection profile E, the sum of the two neighborhood pixel values is used as response information. Calculating optimal azimuth response O respectively Θ (x, y) and orthogonal azimuth response O Θ+ (x, y), and performing a difference operation on the two to obtain an azimuth response difference Δo (x, y), as shown in formula (12):
finally, contrast enhancement is carried out on E (x, y) to obtain a refined contour map E t (x, y) as shown in formula (13):
E t (x,y)＝E(x,y)+E enha (x,y)-E inhi (x,y) (13)
wherein E is enha (x, y) represents the result of linear normalization of the azimuth response difference Δo (x, y) for enhancing the contour pixel point. E (E) inhi (x, y) represents the result of exponential normalization of the position response difference DeltaL (x, y; sigma) for suppressing background pixels.
Step (3) constructing an adaptive size sparse coding model, and refining a contour map E t (x, y) performing sparse coding to obtain a pooling graph E s (x, y). For the refined profile E obtained in step (2) t And (x, y), calculating an adaptive threshold value thresh according to a maximum inter-class variance method, binarizing and counting the proportion of the contour pixels, wherein the proportion is represented by a formula (14):
wherein,representing the binarized result of the refined contour map, the count (& gt) operation is used for counting binary imagesThe number of pixel values 1.
Then, calculating the size of a sparse kernel of the sparse coding model, selecting the shape of a sparse kernel window as a square window for simplifying calculation, and adaptively selecting the window size according to the outline pixel proportion description size, wherein the formula (15) is as follows:
wherein w is 1 、w 2 、w 3 Representing sparse kernel window sizes of different sizes. The invention is provided with w 1 ＝3、w 2 ＝5、w 3 =7. thresh1 and thresh2 are used to measure the refined profile E t Threshold parameters of the sparseness degree default to the value thresh1=0.2, thresh2=0.1. According to E t The relative sizes of the middle outline pixel proportion project and the threshold parameter are equal to the sparse kernel window size w s And performing self-adaptive selection. Adopting mirror symmetry method to refine contour map E before sparse coding t And filling boundary pixels. For E t Calculating an average value of all pixels in a window taking the average value as the center as sparse coding output of the pixel point, taking the window size as window moving step length, and realizing sparse expression of visual information in space, wherein the sparse expression is shown in a formula (16):
wherein w and h are respectively the lateral offset and the longitudinal offset, and w s The size of the sparse nuclear window is also the window moving step length; floor (·) represents a rounding down function, E s (x, y) is a pooled graph after the thinning process.
Step (4)And calculating an obstacle outline result T (x, y) after feedback correction. Pooling the FIG. E obtained in step (3) s And (x, y) feeding back to the V1 visual cortex in a cross-visual area mode, correcting the topological projection profile map E (x, y) obtained in the step (1) in a form of an adjustment coefficient, and obtaining a final barrier profile result T (x, y) through pixel-by-pixel multiplication and fusion, wherein the final barrier profile result T (x, y) is shown as a formula (17):
wherein, the size (·) represents a bilinear interpolation operation,the size of the sparse precise sketch enlarged by bilinear interpolation operation is the same as that of the topological projection outline.
Claims (8)
1. The road obstacle detection method based on the visual information flow partition projection coding model is characterized by comprising the following steps of:
step one: at V 1 Constructing a parallel vision path model in a topology projection profile global perception unit, and extracting the brightness edge and the color edge of a road traffic map in parallel so as to obtain a topology projection profile E (x, y) representing the overall characteristics of the obstacle and an optimal azimuth index matrix theta (x, y);
step two: at V 4 Constructing a visual information difference enhancement model in the sparse coding fine feature extraction unit, and respectively performing texture suppression and contour enhancement on the topological projection contour map E (x, y) by using the position response difference and the azimuth response difference to obtain a refined contour map E t (x,y)；
Step three: an adaptive size sparse coding model is provided for refining the contour map E t (x, y) carrying out pooling operation, realizing intelligent focusing on the outline features of the obstacle according to the image characteristics, and obtaining a pooling graph E representing the local features of the obstacle s (x,y)；
Step four: feedback regulation mechanism for simulating different brain regions of visual cortexUsing pooled graphs E s (x, y) correcting the topological projection profile map E (x, y) of the cross-view area, and fusing to obtain a final barrier profile result T (x, y);
at V 1 Constructing a parallel vision path model in a topology projection profile global perception unit, and extracting the brightness edge and the color edge of a road traffic map in parallel so as to obtain a topology projection profile E (x, y) representing the overall characteristics of the obstacle and an optimal azimuth index matrix theta (x, y); the method comprises the following steps:
for a road traffic diagram to be detected, decomposing a brightness component I (x, y) and red, green and blue color components R (x, y), G (x, y) and B (x, y); the number of rows and columns of each component are m and n respectively; the parallel vision path model consists of a brightness path and a color path, and extracts the brightness edge and the color edge of the road traffic map respectively;
step 1.1 the mathematical model of the luminance path is shown in formula (1):
wherein sigma and gamma represent the size and ellipticity of classical receptive field,representing selective orientations, taking equally spaced 8 orientations, i.e. θ i = {0 °,45 °,90 °,135 °,180 °,225 °,270 °,315 ° }, which represents convolution operation, e (x, y; θ i ) Representative pixel coordinates are (x, y), selective orientation is θ i The corresponding luminance component edge response;
for each pixel, selecting the maximum value of the brightness component edge response corresponding to all the selective orientations, linearly normalizing the maximum value to be used as output, and simultaneously recording the corresponding optimal orientation index to obtain the contour response E of the brightness path L (x, y) and an optimal azimuth index matrix Θ L (x, y) as shown in formula (2):
wherein, N (·) represents a linear normalization operation;
step 1.2, modeling a color channel by using a color antagonism mechanism, wherein the color channel is divided into four types of antagonism channels of R-on/G-off, G-on/R-off, B-on/Y-off and Y-on/B-off, and yellow component Y (x, Y) = (R (x, Y) +G (x, Y))/2; take the R-on/G-off type antagonistic channel as an example; firstly, simulating the action of cone cells, processing R (x, y) component and G (x, y) component by using Gaussian filter respectively, and recording the result as Then calculating the edge response of the single antagonistic receptive field, and marking the result as S RG (x, y) as shown in formula (3):
then simulating the action of the double antagonistic receptive fields, calculating the position of pixel coordinates (x, y) and the selective azimuth is theta i Corresponding color component edge response d RG (x,y；θ i ) As shown in formula (4):
for each pixel, selecting the maximum value of the edge response of the color component corresponding to all the selective orientations, linearly normalizing the maximum value to be used as output, and simultaneously recording the corresponding optimal orientation index to obtain the edge response D of the R-on/G-off type antagonism channel RG (x, y) and an optimal azimuth index matrix Θ RG (x, y) as shown in formula (5):
the same calculation mode as that of R-on/G-off type antagonistic channels is adopted to obtain the edge response D of other three antagonistic channels GR 、D BY 、D YB And an optimal azimuth index matrix Θ GR 、Θ BY 、Θ YB The method comprises the steps of carrying out a first treatment on the surface of the For each pixel coordinate position, taking the maximum value of edge responses in four antagonistic channels as output, and simultaneously recording the corresponding optimal azimuth index to obtain the contour response E of the color channel C (x, y) and an optimal azimuth index matrix Θ C (x, y) as shown in formula (6):
step 1.3, respectively fusing the contour response of the brightness path and the color path and an optimal azimuth index matrix, and simulating the front-stage characteristic of the visual cortex V1 functional layer to obtain a topological projection contour map E (x, y) and an optimal azimuth index matrix theta (x, y); as shown in formula (7):
2. the method for detecting the road obstacle based on the visual information flow partition projection coding model according to claim 1, wherein the method comprises the following steps: sigma, gamma are set to 2 and 0.5 in order.
3. The method for detecting the road obstacle based on the visual information flow partition projection coding model according to claim 1, wherein the method comprises the following steps: the second step is as follows: at V 4 Constructing a visual information difference enhancement model in the sparse coding fine feature extraction unit, and respectively performing texture suppression and contour enhancement on the topological projection contour map E (x, y) by using the position response difference and the azimuth response difference to obtain a refined contour map E t (x, y); the method comprises the following steps:
first, for the topology projection profile E (x, y) obtained in the first step, a Gaussian function G (x, y; sigma) and a Gaussian difference function DoG are used, respectively + (xY; sigma) is convolved with E (x, y) to obtain visual input L of classical receptive field C (x, y; sigma) and non-classical receptive field visual input L N (x, y; sigma) as shown in formula (8):
wherein,
in the pair (8), L C (x, y; sigma) and L N (x, y; sigma) performing a difference operation to obtain a position response difference DeltaL (x, y; sigma) as shown in formula (10):
△L(x,y；σ)＝max{L C (x,y；σ)-L N (x,y；σ),0} (10)
then, an orthogonal azimuth index matrix Θ of the topology projection profile E (x, y) is calculated + (x, y) as shown in formula (11):
Θ + (x,y)＝(Θ(x,y)+2)mod 8 (11)
where mod represents the remainder taking operation; the theta (x, y) obtained in the step (1) stores the optimal azimuth information at the pixel coordinates (x, y) in an index form, wherein the value range is {0,1,2,3,4,5,6,7}, {0,4} represents the horizontal optimal azimuth, {2,6} represents the vertical optimal azimuth, {1,5} represents the forward-inclined optimal azimuth, {3,7} represents the reverse-inclined optimal azimuth; theta (theta) + (x, y) represents an orthogonal orientation complementary to the optimal orientation Θ (x, y), wherein the horizontal orientation is complementary to the vertical orientation, and the left diagonal orientation is complementary to the right diagonal orientation;
for each pixel coordinate (x, y) in the topological projection profile diagram E, taking the sum of two neighborhood pixel values as response information; calculating optimal azimuth response O respectively Θ (x, y) and orthogonal azimuth responseAnd performing difference operation on the two to obtain azimuth response difference delta O (x, y), as shown in formula (12):
finally, contrast enhancement is carried out on E (x, y) to obtain a refined contour map E t (x, y) as shown in formula (13):
E t (x,y)＝E(x,y)+E enha (x,y)-E inhi (x,y) (13)
wherein E is enha (x, y) represents the result of linear normalization of the azimuth response difference delta O (x, y) and is used for enhancing the contour pixel points; e (E) inhi (x, y) represents the result of exponential normalization of the position response difference DeltaL (x, y; sigma) for suppressing background pixels.
4. The method for detecting road obstacle based on visual information flow partition projection coding model according to claim 3, wherein: sigma is set to 2.
5. The method for detecting road obstacle based on visual information flow partition projection coding model according to claim 3, wherein: and step three: an adaptive size sparse coding model is provided for refining the contour map E t (x, y) carrying out pooling operation, realizing intelligent focusing on the outline features of the obstacle according to the image characteristics, and obtaining a pooling graph E representing the local features of the obstacle s (x, y); the method comprises the following steps:
for the refined profile E obtained in the step two t And (x, y), calculating an adaptive threshold value thresh according to a maximum inter-class variance method, binarizing and counting the proportion of the contour pixels, wherein the proportion is represented by a formula (14):
wherein,binarization of representative refined profileThe count (·) operation is used to count binary imagesThe number of middle pixel values 1;
then, calculating the size of a sparse kernel of the sparse coding model, selecting the shape of a sparse kernel window as a square window for simplifying calculation, and adaptively selecting the window size according to the outline pixel proportion description size, wherein the formula (15) is as follows:
wherein w is 1 、w 2 、w 3 Representing sparse kernel window sizes of different sizes; thresh1 and thresh2 are used to measure the refined profile E t Threshold parameters of sparseness according to E t The relative sizes of the middle outline pixel proportion project and the threshold parameter are equal to the sparse kernel window size w s Performing self-adaptive selection; adopting mirror symmetry method to refine contour map E before sparse coding t Filling boundary pixels; for E t Calculating an average value of all pixels in a window taking the average value as the center as sparse coding output of the pixel point, taking the window size as window moving step length, and realizing sparse expression of visual information in space, wherein the sparse expression is shown in a formula (16):
wherein w and h are respectively the lateral offset and the longitudinal offset, and w s The size of the sparse nuclear window is also the window moving step length; floor (·) represents a rounding down function, E s (x, y) is a pooled graph after the thinning process.
6. The method for detecting road obstacle based on the visual information flow partition projection coding model according to claim 5, wherein: w (w) 1 ＝3，w 2 ＝5，w 3 ＝7。
7. The method for detecting road obstacle based on the visual information flow partition projection coding model according to claim 5, wherein: thresh1=0.2, thresh2=0.1.
8. The method for detecting road obstacle based on the visual information flow partition projection coding model according to claim 5, wherein: and step four: feedback regulation mechanism simulating different brain regions of visual cortex and using pooling diagram E s (x, y) correcting the topological projection profile map E (x, y) of the cross-view area, and fusing to obtain a final barrier profile result T (x, y); the method comprises the following steps:
pooling graph E obtained in step three s And (x, y) feeding back to the V1 visual cortex in a cross-visual area mode, correcting the topological projection profile map E (x, y) obtained in the step one in a form of an adjustment coefficient, and obtaining a final barrier profile shape result T (x, y) through pixel-by-pixel multiplication and fusion, wherein the final barrier profile shape result T (x, y) is shown as a formula (17):
wherein, the size (·) represents a bilinear interpolation operation,the dimension of the pooled map amplified by bilinear interpolation operation is the same as that of the topological projection profile map.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011578651.0A CN112613427B (en) | 2020-12-28 | 2020-12-28 | Road obstacle detection method based on visual information flow partition projection coding model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011578651.0A CN112613427B (en) | 2020-12-28 | 2020-12-28 | Road obstacle detection method based on visual information flow partition projection coding model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112613427A CN112613427A (en) | 2021-04-06 |
| CN112613427B true CN112613427B (en) | 2024-02-27 |
Family
ID=75248224
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011578651.0A Active CN112613427B (en) | 2020-12-28 | 2020-12-28 | Road obstacle detection method based on visual information flow partition projection coding model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112613427B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113435455B (en) * | 2021-05-12 | 2024-03-22 | 深圳灵图创新科技有限公司 | Image contour extraction method based on space-time pulse coding |
| CN113391368B (en) * | 2021-06-30 | 2022-10-21 | 山东国瑞新能源有限公司 | Road exploration method and equipment based on virtual imaging technology |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104484667A (en) * | 2014-12-30 | 2015-04-01 | 华中科技大学 | Contour extraction method based on brightness characteristic and contour integrity |
| CN109489576A (en) * | 2018-10-19 | 2019-03-19 | 杭州电子科技大学 | A kind of profile testing method based on primary vision access computation model |
| CN111222518A (en) * | 2020-01-16 | 2020-06-02 | 杭州电子科技大学 | Contour feature extraction method based on frequency division visual mechanism |
| CN111402285A (en) * | 2020-01-16 | 2020-07-10 | 杭州电子科技大学 | Contour detection method based on visual mechanism dark edge enhancement |
-
2020
- 2020-12-28 CN CN202011578651.0A patent/CN112613427B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104484667A (en) * | 2014-12-30 | 2015-04-01 | 华中科技大学 | Contour extraction method based on brightness characteristic and contour integrity |
| CN109489576A (en) * | 2018-10-19 | 2019-03-19 | 杭州电子科技大学 | A kind of profile testing method based on primary vision access computation model |
| CN111222518A (en) * | 2020-01-16 | 2020-06-02 | 杭州电子科技大学 | Contour feature extraction method based on frequency division visual mechanism |
| CN111402285A (en) * | 2020-01-16 | 2020-07-10 | 杭州电子科技大学 | Contour detection method based on visual mechanism dark edge enhancement |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112613427A (en) | 2021-04-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Han et al. | Underwater image processing and object detection based on deep CNN method | |
| CN111079685B (en) | 3D target detection method | |
| Fritsch et al. | Monocular road terrain detection by combining visual and spatial information | |
| CN103914699A (en) | Automatic lip gloss image enhancement method based on color space | |
| CN103996198B (en) | The detection method of area-of-interest under Complex Natural Environment | |
| CN106650640A (en) | Negative obstacle detection method based on local structure feature of laser radar point cloud | |
| CN108665463A (en) | A kind of cervical cell image partition method generating network based on confrontation type | |
| CN110619638A (en) | Multi-mode fusion significance detection method based on convolution block attention module | |
| CN104408711B (en) | Multi-scale region fusion-based salient region detection method | |
| CN107633513A (en) | The measure of 3D rendering quality based on deep learning | |
| CN112613427B (en) | Road obstacle detection method based on visual information flow partition projection coding model | |
| CN103699900B (en) | Building horizontal vector profile automatic batch extracting method in satellite image | |
| CN104766096B (en) | A kind of image classification method based on multiple dimensioned global characteristics and local feature | |
| CN111563447A (en) | Crowd density analysis and detection positioning method based on density map | |
| CN110414385B (en) | Lane line detection method and system based on homography transformation and characteristic window | |
| CN112488046B (en) | Lane line extraction method based on high-resolution images of unmanned aerial vehicle | |
| CN104616308A (en) | Multiscale level set image segmenting method based on kernel fuzzy clustering | |
| CN107315998A (en) | Vehicle class division method and system based on lane line | |
| CN103996185A (en) | Image segmentation method based on attention TD-BU mechanism | |
| CN102799646B (en) | A kind of semantic object segmentation method towards multi-view point video | |
| CN103679718A (en) | Fast scenario analysis method based on saliency | |
| CN111369617A (en) | 3D target detection method of monocular view based on convolutional neural network | |
| CN103955945A (en) | Self-adaption color image segmentation method based on binocular parallax and movable outline | |
| CN106529441B (en) | Depth motion figure Human bodys' response method based on smeared out boundary fragment | |
| Khan et al. | Lrdnet: lightweight lidar aided cascaded feature pools for free road space detection |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |