CN111523478B - Pedestrian image detection method acting on target detection system - Google Patents
Pedestrian image detection method acting on target detection system Download PDFInfo
- Publication number
- CN111523478B CN111523478B CN202010331352.0A CN202010331352A CN111523478B CN 111523478 B CN111523478 B CN 111523478B CN 202010331352 A CN202010331352 A CN 202010331352A CN 111523478 B CN111523478 B CN 111523478B
- Authority
- CN
- China
- Prior art keywords
- pedestrian image
- camouflage pattern
- detection system
- pedestrian
- camouflage
- 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
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
-
- 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
-
- 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)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Human Computer Interaction (AREA)
- Traffic Control Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a pedestrian image detection method acting on a target detection system, which comprises the following steps: constructing a pedestrian image detection model in a target detection system, collecting a pedestrian image to be detected, inputting the pedestrian image to be detected into the pedestrian image detection model, and outputting a pedestrian image detection result by the pedestrian image detection model; by taking the pedestrian image with the camouflage pattern and the pedestrian image without the camouflage pattern as training data of the pedestrian image detection model, the pedestrian image detection model can have the capability of identifying the camouflage pattern, and meanwhile, by constructing the pedestrian image detection model in the target detection system, the target detection system can improve the defending capability and the detection capability of the target detection system and avoid the condition of missed detection or false detection by taking corresponding measures on the detection result of the pedestrian image detection model.
Description
Technical Field
The invention relates to the technical field of pedestrian tracking, in particular to a pedestrian image detection method acting on a target detection system.
Background
Pedestrian tracking, i.e. estimating the position and motion parameters of a pedestrian given the initial position of the pedestrian in the first frame of image in an image sequence, is widely used in many fields in real life, such as the video monitoring field, the intelligent robot field, the automobile auxiliary driving field and the automatic driving field, in which the pedestrian tracking is an indispensable technology, and in the social approach to the intelligent road, the pedestrian tracking plays an increasing role.
Most pedestrian tracking and image recognition methods are deployed based on a deep learning target detection model, however, the conventional recognition method is easy to attack by others by adopting a universal camouflage attack means due to insufficient training data, and an attacker slightly falsifies the content of an image or an image, so that specific image content cannot be detected and positioned by an artificial intelligent system, and the condition of missed detection or false detection occurs.
Disclosure of Invention
The invention aims to provide a pedestrian image detection method acting on a target detection system, which solves the problem of missing error detection caused by insufficient training data in the prior art.
The invention is realized by the following technical scheme:
a pedestrian image detection method acting on a target detection system, comprising the steps of:
step S1, constructing a pedestrian image detection model in a target detection system, wherein the input end of the pedestrian image detection model is a pedestrian image, and the output end of the pedestrian image detection model is a pedestrian image detection result, and the pedestrian image comprises a pedestrian image containing a camouflage pattern and a pedestrian image not containing the camouflage pattern;
s2, collecting pedestrian images to be detected;
s3, inputting a pedestrian image to be detected into the pedestrian image detection model;
and S4, outputting a pedestrian image detection result by the pedestrian image detection model.
As a further alternative to the pedestrian image detection method acting on the object detection system, the generation of the pedestrian image containing the camouflage pattern is based on the object detection system, comprising the steps of:
step S11, generating a random noise matrix based on standard front-end distribution, and processing the random noise matrix to obtain an original camouflage pattern;
step S12, performing physical simulation operation on the original camouflage pattern to obtain a simulated data set;
step S13, inputting the simulated data set into a target detection system for detection to obtain a detection result;
step S14, adjusting the simulated data set according to the detection result to obtain an initial pedestrian image containing a camouflage pattern;
step S15, optimizing the initial pedestrian image containing the camouflage pattern to generate the final pedestrian image containing the camouflage pattern.
As a further alternative of the pedestrian image detection method acting on the target detection system, the performing the physical simulation operation on the original camouflage pattern in step S12 includes the steps of:
step S121, performing physical simulation on the intrinsic characteristics of the original camouflage pattern;
step S122, performing physical simulation on the external environment where the original camouflage pattern is located.
As a further alternative to the pedestrian image detection method acting on the target detection system, the step S121 includes the steps of:
step S1211, simulating a stretched state of the original camouflage pattern in the case of a non-rigid/non-planar object;
step S1212, simulating state images of the original camouflage image in different shielding degrees;
step S1213 simulates a state in which the original camouflage image is at a different position on the object.
As a further alternative of the pedestrian image detection method acting on the object detection system, the object detection system includes a region extraction module, a classification module, and a positioning module.
As a further alternative to the pedestrian image detection method acting on the target detection system, the step S13 includes the steps of:
step S131, inputting the simulated data set into a region extraction module to obtain a foreground object and a background object;
step S132, inputting the foreground object and the background object into a classification module, and classifying the composition of the foreground object and the composition of the background object;
step S133, inputting the composition of the foreground object and the composition of the background object into a positioning module to obtain the positions of the compositions of the foreground object and the background object.
As a further alternative to the pedestrian image detection method acting on the target detection system, the optimizing of the initial pedestrian image containing the camouflage pattern in step S15 includes the steps of:
step S151, processing an initial pedestrian image containing a camouflage pattern based on semantic constraint, and inputting the processed pedestrian image into a target detection system for detection;
in step S152, if the target detection system does not detect the camouflage pattern, the original pedestrian image including the camouflage pattern input to the target detection system is used as the final pedestrian image including the camouflage pattern, and if the target detection system detects the camouflage pattern, the camouflage pattern in the original pedestrian image including the camouflage pattern input to the target detection system is used as the original camouflage pattern to be processed in step S12.
As a further alternative to the pedestrian image detection method acting on the object detection system, the processing of the initial pedestrian image containing the camouflage pattern based on the semantic constraint in step S151 includes the steps of:
step S1511, selecting an initial pedestrian image containing a camouflage pattern, wherein the initial pedestrian image containing the camouflage pattern comprises the camouflage pattern and a natural image;
in step S1512, the camouflage pattern is projected into an infinite norm space of the natural image, and a numerical range of the camouflage image is generated in a constraint manner.
As a further alternative of the pedestrian image detection method acting on the target detection system, the pedestrian image detection model includes a full-connection parameter layer, a normalization layer, a nonlinear activation function layer, and a pedestrian image detection result output layer.
The invention has the beneficial effects that:
by using the method, the pedestrian image with the camouflage pattern and the pedestrian image without the camouflage pattern are used as training data of the pedestrian image detection model, so that the pedestrian image detection model can have the capability of identifying the camouflage pattern, meanwhile, the target detection system can improve the defending capability and the detection capability of the target detection system and avoid the condition of missed detection or false detection by constructing the pedestrian image detection model in the target detection system and taking corresponding measures for the detection result of the pedestrian image detection model.
Drawings
FIG. 1 is a flow chart of a pedestrian image detection method for an object detection system according to the present invention;
fig. 2 is a flow chart of a process of generating a pedestrian image including a camouflage pattern in a pedestrian image detection method acting on an object detection system of the present invention.
Detailed Description
The present invention will now be described in detail with reference to the drawings and the detailed description thereof, wherein the invention is illustrated by the schematic drawings and the detailed description thereof, which are included to illustrate and not to limit the invention.
As shown in fig. 1-2, a pedestrian image detection method acting on an object detection system includes the steps of:
step S1, constructing a pedestrian image detection model in a target detection system, wherein the input end of the pedestrian image detection model is a pedestrian image, and the output end of the pedestrian image detection model is a pedestrian image detection result, and the pedestrian image comprises a pedestrian image containing a camouflage pattern and a pedestrian image not containing the camouflage pattern;
s2, collecting pedestrian images to be detected;
s3, inputting a pedestrian image to be detected into the pedestrian image detection model;
and S4, outputting a pedestrian image detection result by the pedestrian image detection model.
In this embodiment, by taking the pedestrian image including the camouflage pattern and the pedestrian image not including the camouflage pattern as training data of the pedestrian image detection model, the pedestrian image detection model can have the capability of identifying the camouflage pattern, and meanwhile, by constructing the pedestrian image detection model in the target detection system, the target detection system can improve the defending capability and the detection capability of the target detection system by making corresponding measures for the detection result of the pedestrian image detection model, thereby avoiding the condition of missed detection or false detection.
Preferably, the generation of the pedestrian image containing the camouflage pattern is based on an object detection system, comprising the steps of:
step S11, generating a random noise matrix based on standard front-end distribution, and processing the random noise matrix to obtain an original camouflage pattern;
step S12, performing physical simulation operation on the original camouflage pattern to obtain a simulated data set;
step S13, inputting the simulated data set into a target detection system for detection to obtain a detection result;
step S14, adjusting the simulated data set according to the detection result to obtain an initial pedestrian image containing a camouflage pattern;
step S15, optimizing the initial pedestrian image containing the camouflage pattern to generate the final pedestrian image containing the camouflage pattern.
In the embodiment, the camouflage capability of the camouflage pattern can be stronger and more variation of the camouflage pattern can be ensured by generating and adjusting the simulated data set according to the detection result, and the effectiveness and diversity of training data are ensured;
it should be noted that the pseudo-data set includes a camouflage image and an environment in which the camouflage image is located.
Preferably, the performing the physical simulation operation on the original camouflage pattern in the step S12 includes the following steps:
step S121, performing physical simulation on the intrinsic characteristics of the original camouflage pattern;
step S122, performing physical simulation on the external environment where the original camouflage pattern is located.
In this embodiment, the physical simulation is performed on the intrinsic characteristics of the original camouflage pattern, including simulating the image deformation state of the non-rigid body/non-planar object under the conditions of stretching, shielding, displacement, and the like, so that the original camouflage pattern can be more closely attached to the attack target, the camouflage capability of the camouflage pattern is enhanced, and meanwhile, the physical simulation is performed on the external environment where the camouflage pattern is located, including the conditions of different illumination conditions, imaging angles, distances, and the like in the environment where the object is located, the camouflage effect of the camouflage pattern under different environments can be simulated, so that the camouflage pattern can be better applied to the real physical world.
Preferably, the step S121 includes the steps of:
step S1211, simulating a stretched state of the original camouflage pattern in the case of a non-rigid/non-planar object;
step S1212, simulating state images with different shielding degrees of the camouflage image;
in step S1213, the state in which the camouflage image is at different positions on the object is simulated.
In this embodiment, affine and scale operations are performed on the camouflage image I by generating a homography matrix H and a scaling coefficient s through random parameters, and projecting the camouflage image I onto a plane at random angles, and simulating a stretching state I 'under the condition of a non-rigid body/non-planar object, wherein a calculation formula of any pixel (x', y ') in I' is as follows:
then carrying out random cutting and filling operation on the stretching image I', and simulating the state image I with different shielding degrees of the camouflage image c Finally, for the occlusion image I c Performing translation operation to simulate state I of camouflage image at different positions on object t Wherein I t Any one of the pixels (x t ,y t ) From the filling parameter p and the translation parameter (r x ,r y ) The formula is calculated as:
preferably, the target detection system comprises a region extraction module, a classification module and a positioning module.
Preferably, the step S13 includes the steps of:
step S131, inputting the simulated data set into a region extraction module to obtain a foreground object and a background object;
step S132, inputting the foreground object and the background object into a classification module, and classifying the composition of the foreground object and the composition of the background object;
step S133, inputting the composition of the foreground object and the composition of the background object into a positioning module to obtain the positions of the compositions of the foreground object and the background object.
In this embodiment, by obtaining the specific position of each component of the foreground object and the specific position of each component of the background object according to the region extraction module, the classification module and the positioning module of the target detection system, and adjusting the pseudo-data according to the obtained specific positions, an initialized pedestrian image containing the camouflage pattern can be obtained, the search function of the region extraction module in the target detection system can be weakened by adjusting the obtained initialized pedestrian image containing the camouflage pattern, the interference on the foreground candidate frame set P obtained by searching the target detection system is performed by using the camouflage pattern, the confidence level is reduced, so that the foreground object is classified as the background object, the search of the avoidance system is performed, and meanwhile, the judgment result of the classification module in the target detection system, namely, the foreground candidate frame set P containing the object t, can be misled * Outputting any candidate frame content in the set as a wrong result through the camouflage pattern misleading classification module
In addition, the positioning result in the target detection system can be distorted, and for the coordinate vector of the target candidate frame, the camouflage pattern shifts the center coordinate and the dimension of the positioning bounding box, so that the accuracy of the positioning result is reduced, and the wrong coordinate vector is output.
Preferably, the optimizing the initial pedestrian image including the camouflage pattern in the step S15 includes the steps of:
step S151, processing an initial pedestrian image containing a camouflage pattern based on semantic constraint, and inputting the processed pedestrian image into a target detection system for detection;
in step S152, if the target detection system does not detect the camouflage pattern, the original pedestrian image including the camouflage pattern input to the target detection system is used as the final pedestrian image including the camouflage pattern, and if the target detection system detects the camouflage pattern, the camouflage pattern in the original pedestrian image including the camouflage pattern input to the target detection system is used as the original camouflage pattern to be processed in step S12.
In this embodiment, although the initial pedestrian image including the camouflage pattern has a certain camouflage capability, the camouflage effect is not certain, so that the training data of the pedestrian image detection model is insufficient, so that the initial pedestrian image including the camouflage pattern is further optimized through semantic constraint, the camouflage capability of the camouflage pattern is stronger, if the initial pedestrian image including the camouflage pattern still does not have a good camouflage effect after the semantic constraint, the camouflage pattern is sent back to the step S12 for repeated operation until the stop condition is met, and the effectiveness of the training data is further ensured.
Preferably, the processing of the initial pedestrian image containing the camouflage pattern based on the semantic constraint in the step S151 includes the following steps:
step S1511, selecting an initial pedestrian image containing a camouflage pattern, wherein the initial pedestrian image containing the camouflage pattern comprises the camouflage pattern and a natural image;
in step S1512, the camouflage pattern is projected into an infinite norm space of the natural image, and a numerical range of the camouflage image is generated in a constraint manner.
Preferably, the pedestrian image detection model comprises a full-connection parameter layer, a normalization layer, a nonlinear activation function layer and a pedestrian image detection result output layer.
Examples:
step S1, firstly, constructing a pedestrian image detection model in a target detection system, then generating a pedestrian image containing a camouflage pattern and a pedestrian image without the camouflage pattern as training data of the pedestrian image detection model, and finally training the pedestrian image detection model by using the generated training data; the pedestrian image without the camouflage pattern can be directly obtained from the camera, and the pedestrian image with the camouflage pattern is generated according to the following method:
1. generating a random noise matrix z with the dimension of h multiplied by w multiplied by c based on standard N (0, 1), and initializing a camouflage image I, wherein h, w and c are the height, width and channel dimension of the training set image respectively;
2. firstly, carrying out physical simulation on the internal characteristics of a camouflage pattern, including simulating image deformation states of stretching, shielding, displacement and the like under the condition of a non-rigid body/non-planar object, and specifically:
firstly, generating a scaling coefficient s and a homography matrix H through random parameters, carrying out affine and scaling operation on a camouflage image I, projecting the camouflage image I onto a plane with a random angle, and simulating a stretching state I' under the condition of a non-rigid body/non-planar object;
the homography matrix H is calculated from 4 co-ordinates p1, p2, p3, p4 and 4 source coordinates q1, q2, q3, q 4. The source coordinates and the uniform coordinates respectively represent the corresponding coordinate positions of the same pixel before and after affine, the 4 groups of coordinates are solved by a least linear square method, and the calculation formula of the homography matrix H is as follows:
wherein, (x) i ,y i ) And (x' i ,y′ i ) Respectively represent Ji Zuobiao p i And source coordinates q i Position of source coordinates q i Comprising 4 vertices of the camouflage image, the coordinates being (0, 0), (h, 0), (0,w) and (h, w), and the coordinates p i Adding random offset vectors on the basis of source coordinate vertexes to obtain (r 1, r 2), (h+r3, r 4), (r 5, w+r6) and (h+r7, w+r8), wherein ri is a random number generated based on the normal Ethernet distribution, and the positions of 4 vertexes of the affine camouflage image are represented;
on the basis of acquiring the homography matrix H and the scaling coefficient s, calculating the value of the corresponding pixel coordinate (x ', y ') in the affine image I ' according to any pixel (x, y) in the initial camouflage image I, wherein the value is shown in the following formula:
then, carrying out random cutting operation on the stretched image I ', simulating an image with the dimension of h ' xw ' in a shielding state, and refilling the image to the dimension of h x w to obtain a camouflage image I c The formula is:
I c ＝Padding(Cropping(I′,h′×w′),h×w)
wherein, padding and Cropping are image cutting and filling operations respectively;
finally, for occlusion image I c Performing translation operation to simulate state I of camouflage image at different positions on object t Wherein I t Any one of the pixels (x t ,y t ) From translation parameters (r x ,r y ) The calculation is carried out, and then the image is restored to h multiplied by w dimensions through filling, and the formula is as follows:
3. performing physical simulation on the external environment where the camouflage pattern is located to obtain a simulated data set, wherein the simulated data set comprises different conditions of illumination conditions, imaging angles, distances and the like in the environment where the object is located, and specifically comprises the following steps:
first, the camouflage image I subjected to intrinsic characteristic simulation t Added to the data set T to which the attacked target T belongs, an initial data set T' is generated, expressed as:
wherein M is a mask set during image superposition, and is used for ensuring that the value range of two different images after superposition operation is within the value range [0,255 ];
then, all training images in the initial data set T' are subjected to random transformation, different environments where the target object is located, including illumination conditions, imaging angles, distances and the like, are simulated, and a simulated data set is obtainedT * The formula is:
T * ＝{t * |t * ＝Transform(t′),t′～T′}
the Transform is a random image transformation operation, including linear affine, contrast correction, rotation, bilinear interpolation sampling, and the like.
4. Inputting the simulated data set into a target detection system for detection to obtain a detection result, wherein the detection result comprises the following specific steps:
step S131, inputting the simulated data set into a region extraction module to obtain a foreground object and a background object;
step S132, inputting the foreground object and the background object into a classification module, and classifying the composition of the foreground object and the composition of the background object;
step S133, inputting the composition of the foreground object and the composition of the background object into a positioning module to obtain the positions of the compositions of the foreground object and the background object.
5. Adjusting the simulated data set according to the detection result to obtain an initial pedestrian image containing a camouflage pattern, wherein the initial pedestrian image comprises the following specific steps of:
firstly, weakening the searching function of a region extraction module in a target detection system, utilizing a camouflage pattern to interfere a foreground candidate frame set P obtained by searching the target detection system, and reducing the confidence coefficient of the foreground candidate frame set P, so that the foreground object is classified as a background object, and avoiding the searching of the system, wherein an optimization formula is expressed as follows:
wherein L is Euclidean metric distance, R is region extraction module, R (t * ) For the module pair image t * The search result of the (a) is a candidate frame containing a foreground object, b is a confidence vector of the candidate frame, the background is set to be 1.0, the foreground is set to be 0.0, and the accuracy of the search result of the area extraction module can be reduced by minimizing the formula, so that the target detection system cannot recognize the foreground object;
then misleading target detectionThe judgment result of the classification module in the system, namely, the foreground candidate frame set P containing the object t * Outputting any candidate frame content in the set as a wrong result through the camouflage pattern misleading classification module
The optimization formula is as follows:
wherein L is the cross entropy distance, C is the classification module, C (t * ) For the module pair image t * Minimizing the above formula can increase the content identified as erroneous by the classification module while reducing the confidence that the image content is identified as the target t
Probability of (2);
finally, the positioning result in the target detection system is distorted, for the coordinate vector of the target candidate frame, the camouflage pattern shifts the center coordinate and the dimension of the positioning bounding box, the accuracy of the positioning result is reduced, the wrong coordinate vector is output, and the formula is defined as follows:
wherein L is Euclidean distance, D is positioning module, D (t * ) For the module pair image t * The positioning result of the content in (a), namely the coordinate vector of the identification object candidate frame,
is an offset vector.
6. Optimizing the initial pedestrian image containing the camouflage pattern to generate the final pedestrian image containing the camouflage pattern, specifically:
selecting a piece of information to be attributed to the error knotFruit category t * Projecting a camouflage image into a small-range infinite norm space of the natural image, and restricting the numerical range of the camouflage image to be generated, wherein the numerical range is expressed as follows:
wherein, the project is a Projection function,
for identifying error content after attack>
And the corresponding natural image epsilon is a smaller fixed value and is used for restraining the visual similarity degree of the camouflage image and the natural image.
Step S2, acquiring pedestrian images to be detected through a monitoring camera;
s3, inputting a pedestrian image to be detected into the pedestrian image detection model;
and S4, outputting a pedestrian image detection result by the pedestrian image detection model.
The foregoing has described in detail the technical solutions provided by the embodiments of the present invention, and specific examples have been applied to illustrate the principles and implementations of the embodiments of the present invention, where the above description of the embodiments is only suitable for helping to understand the principles of the embodiments of the present invention; meanwhile, as for those skilled in the art, according to the embodiments of the present invention, there are variations in the specific embodiments and the application scope, and the present description should not be construed as limiting the present invention.
Claims (5)
1. A pedestrian image detection method acting on a target detection system, characterized by: the method comprises the following steps:
step S1, constructing a pedestrian image detection model in a target detection system, wherein the input end of the pedestrian image detection model is a pedestrian image, and the output end of the pedestrian image detection model is a pedestrian image detection result, and the pedestrian image comprises a pedestrian image containing a camouflage pattern and a pedestrian image not containing the camouflage pattern;
s2, collecting pedestrian images to be detected;
s3, inputting a pedestrian image to be detected into the pedestrian image detection model;
step S4, the pedestrian image detection model outputs a pedestrian image detection result;
wherein the generation of the pedestrian image containing the camouflage pattern is based on the target detection system, comprising the following steps:
step S11, generating a random noise matrix based on standard front-end distribution, and processing the random noise matrix to obtain an original camouflage pattern;
step S12, performing physical simulation operation on the original camouflage pattern to obtain a simulated data set;
step S13, inputting the simulated data set into a target detection system for detection to obtain a detection result;
step S14, adjusting the simulated data set according to the detection result to obtain an initial pedestrian image containing a camouflage pattern;
step S15, optimizing the initial pedestrian image containing the camouflage pattern to generate a final pedestrian image containing the camouflage pattern;
the performing the physical simulation operation on the original camouflage pattern in the step S12 includes the following steps:
step S121, performing physical simulation on the intrinsic characteristics of the original camouflage pattern;
step S122, performing physical simulation on the external environment where the original camouflage pattern is located;
the step S121 includes the steps of:
step S1211, simulating a stretched state of the original camouflage pattern in the case of a non-rigid/non-planar object;
step S1212, simulating state images of the original camouflage image in different shielding degrees;
step S1213, simulating the state that the original camouflage image is at different positions on the object;
the optimizing of the original pedestrian image including the camouflage pattern in the step S15 includes the steps of:
step S151, processing an initial pedestrian image containing a camouflage pattern based on semantic constraint, and inputting the processed pedestrian image into a target detection system for detection;
in step S152, if the target detection system does not detect the camouflage pattern, the original pedestrian image including the camouflage pattern input to the target detection system is used as the final pedestrian image including the camouflage pattern, and if the target detection system detects the camouflage pattern, the camouflage pattern in the original pedestrian image including the camouflage pattern input to the target detection system is used as the original camouflage pattern to be processed in step S12.
2. A pedestrian image detection method acting on an object detection system according to claim 1, characterized in that: the target detection system comprises a region extraction module, a classification module and a positioning module.
3. A pedestrian image detection method acting on an object detection system according to claim 2, characterized in that: the step S13 includes the steps of:
step S131, inputting the simulated data set into a region extraction module to obtain a foreground object and a background object;
step S132, inputting the foreground object and the background object into a classification module, and classifying the composition of the foreground object and the composition of the background object;
step S133, inputting the composition of the foreground object and the composition of the background object into a positioning module to obtain the positions of the compositions of the foreground object and the background object.
4. A pedestrian image detection method acting on an object detection system as claimed in claim 3, wherein: the processing of the initial pedestrian image containing the camouflage pattern based on the semantic constraint in the step S151 includes the steps of:
step S1511, selecting an initial pedestrian image containing a camouflage pattern, wherein the initial pedestrian image containing the camouflage pattern comprises the camouflage pattern and a natural image;
in step S1512, the camouflage pattern is projected into an infinite norm space of the natural image, and a numerical range of the camouflage image is generated in a constraint manner.
5. A pedestrian image detection method acting on an object detection system as claimed in claim 4, wherein: the pedestrian image detection model comprises a full-connection parameter layer, a normalization layer, a nonlinear activation function layer and a pedestrian image detection result output layer.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010331352.0A CN111523478B (en) | 2020-04-24 | 2020-04-24 | Pedestrian image detection method acting on target detection system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010331352.0A CN111523478B (en) | 2020-04-24 | 2020-04-24 | Pedestrian image detection method acting on target detection system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111523478A CN111523478A (en) | 2020-08-11 |
| CN111523478B true CN111523478B (en) | 2023-04-28 |
Family
ID=71904518
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010331352.0A Active CN111523478B (en) | 2020-04-24 | 2020-04-24 | Pedestrian image detection method acting on target detection system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111523478B (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102982703A (en) * | 2012-12-12 | 2013-03-20 | 成都合纵连横数字科技有限公司 | Interactive method of automobile driving simulator and virtual traffic environment simulation system |
| CN107301376A (en) * | 2017-05-26 | 2017-10-27 | 浙江大学 | A kind of pedestrian detection method stimulated based on deep learning multilayer |
| CN109766790A (en) * | 2018-12-24 | 2019-05-17 | 重庆邮电大学 | A kind of pedestrian detection method based on self-adaptive features channel |
| CN109886157A (en) * | 2019-01-30 | 2019-06-14 | 杭州芯影科技有限公司 | A kind of face identification method and system based on millimeter-wave image |
| CN110569602A (en) * | 2019-09-10 | 2019-12-13 | 中国科学技术大学 | Data acquisition method and system for unmanned vehicle |
| CN110717532A (en) * | 2019-09-27 | 2020-01-21 | 广东工业大学 | Real-time detection method for robot target grabbing area based on SE-RetinaGrasp model |
-
2020
- 2020-04-24 CN CN202010331352.0A patent/CN111523478B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102982703A (en) * | 2012-12-12 | 2013-03-20 | 成都合纵连横数字科技有限公司 | Interactive method of automobile driving simulator and virtual traffic environment simulation system |
| CN107301376A (en) * | 2017-05-26 | 2017-10-27 | 浙江大学 | A kind of pedestrian detection method stimulated based on deep learning multilayer |
| CN109766790A (en) * | 2018-12-24 | 2019-05-17 | 重庆邮电大学 | A kind of pedestrian detection method based on self-adaptive features channel |
| CN109886157A (en) * | 2019-01-30 | 2019-06-14 | 杭州芯影科技有限公司 | A kind of face identification method and system based on millimeter-wave image |
| CN110569602A (en) * | 2019-09-10 | 2019-12-13 | 中国科学技术大学 | Data acquisition method and system for unmanned vehicle |
| CN110717532A (en) * | 2019-09-27 | 2020-01-21 | 广东工业大学 | Real-time detection method for robot target grabbing area based on SE-RetinaGrasp model |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111523478A (en) | 2020-08-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111429533B (en) | Camera lens distortion parameter estimation device and method | |
| CN110378837B (en) | Target detection method and device based on fish-eye camera and storage medium | |
| CN101860729A (en) | Target tracking method for omnidirectional vision | |
| JP5833507B2 (en) | Image processing device | |
| CN111797688A (en) | Visual SLAM method based on optical flow and semantic segmentation | |
| CN112652020B (en) | Visual SLAM method based on AdaLAM algorithm | |
| CN108537844A (en) | A kind of vision SLAM winding detection methods of fusion geological information | |
| CN114399675A (en) | Target detection method and device based on machine vision and laser radar fusion | |
| JP6016242B2 (en) | Viewpoint estimation apparatus and classifier learning method thereof | |
| CN111523478B (en) | Pedestrian image detection method acting on target detection system | |
| CN110322476B (en) | Target tracking method for improving STC and SURF feature joint optimization | |
| CN113012298B (en) | Curved MARK three-dimensional registration augmented reality method based on region detection | |
| CN113837243B (en) | RGB-D camera dynamic visual odometer method based on edge information | |
| CN113723432B (en) | Intelligent identification and positioning tracking method and system based on deep learning | |
| CN113159158A (en) | License plate correction and reconstruction method and system based on generation countermeasure network | |
| CN114972451A (en) | Rotation-invariant SuperGlue matching-based remote sensing image registration method | |
| Ueda et al. | Data Augmentation for Semantic Segmentation Using a Real Image Dataset Captured Around the Tsukuba City Hall | |
| CN110866952A (en) | Repositioning and loop detection method based on depth map coding | |
| Sang et al. | An improved checkerboard detection algorithm based on adaptive filters | |
| CN113361544B (en) | Image acquisition equipment, and external parameter correction method, device and storage medium thereof | |
| CN114863692B (en) | Vehicle pattern recognition fee evasion checking method based on local feature alignment | |
| CN116704487B (en) | License plate detection and identification method based on Yolov s network and CRNN | |
| CN118097566B (en) | Scene change detection method, device, medium and equipment based on deep learning | |
| CN117593650B (en) | Moving point filtering vision SLAM method based on 4D millimeter wave radar and SAM image segmentation | |
| CN114612798B (en) | Satellite image tampering detection method based on Flow model |
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 |