CN114842362A - Pull type unmanned aerial vehicle mobile airport and operation method thereof - Google Patents

Pull type unmanned aerial vehicle mobile airport and operation method thereof Download PDF

Info

Publication number
CN114842362A
CN114842362A CN202210557467.0A CN202210557467A CN114842362A CN 114842362 A CN114842362 A CN 114842362A CN 202210557467 A CN202210557467 A CN 202210557467A CN 114842362 A CN114842362 A CN 114842362A
Authority
CN
China
Prior art keywords
feature map
feature
stopped
neural network
point
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.)
Withdrawn
Application number
CN202210557467.0A
Other languages
Chinese (zh)
Inventor
华威
张鹏飞
邵安强
包菊芬
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Taichang Technology Hangzhou Co ltd
Original Assignee
Taichang Technology Hangzhou Co ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Taichang Technology Hangzhou Co ltd filed Critical Taichang Technology Hangzhou Co ltd
Priority to CN202210557467.0A priority Critical patent/CN114842362A/en
Publication of CN114842362A publication Critical patent/CN114842362A/en
Withdrawn legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/17—Terrestrial scenes taken from planes or by drones
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B60—VEHICLES IN GENERAL
    • B60P—VEHICLES ADAPTED FOR LOAD TRANSPORTATION OR TO TRANSPORT, TO CARRY, OR TO COMPRISE SPECIAL LOADS OR OBJECTS
    • B60P3/00—Vehicles adapted to transport, to carry or to comprise special loads or objects
    • B60P3/06—Vehicles adapted to transport, to carry or to comprise special loads or objects for carrying vehicles
    • B60P3/11—Vehicles adapted to transport, to carry or to comprise special loads or objects for carrying vehicles for carrying aircraft
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted 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
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • 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
    • Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation
    • Y02T90/10—Technologies relating to charging of electric vehicles
    • Y02T90/14—Plug-in electric vehicles

Abstract

The application relates to the field of mobile unmanned airports, and particularly discloses a mobile airport of a pulling type unmanned aerial vehicle and an operation method thereof. The pulling type unmanned aerial vehicle mobile airport is right through using a depth convolution neural network model to treat the image of the stopping point and treat that the panoramic image around the stopping point encodes in order to be used for categorised feature extraction, just so can be based on treat the stopping point with treat that the characteristic information around the stopping point comes to synthesize right treat whether the stopping point suits to park and accurately classify and judge, and then make the pulling type unmanned aerial vehicle mobile airport can intelligently select suitable stopping point when patrolling and examining automatically to guarantee unmanned aerial vehicle's normal work and security.

Description

Pull type unmanned aerial vehicle mobile airport and operation method thereof
Technical Field
The present invention relates to the field of mobile unmanned airports, and more particularly, to a mobile airport for a towed unmanned aerial vehicle and an operation method thereof.
Background
With the rapid development of economy, the demand of the whole society for social energy is larger and larger, and the ultrahigh-voltage large-capacity power line is greatly expanded, so that the power line needs to span various complex environments, such as mountainous areas, basins, reservoirs, lakes and the like, which bring great difficulty to the maintenance and overhaul of the high-voltage power line, particularly in forest areas, high-altitude areas, ice and snow covered areas, mountain landslide areas, areas where geological disasters are easy to occur and the like. And the application of the unmanned aerial vehicle makes the power inspection and the emergency repair of the areas easy.
Along with the more mature and stable unmanned aerial vehicle technique, also more and more are being used in the field of patrolling and examining. At present, unmanned aerial vehicle's patrolling and examining is controlled through patrolling and examining the platform by the staff to patrol and examine and overhaul overhead high tension transmission line through machine-mounted sensing equipment. Along with the improvement of the intelligent patrol requirement, the requirement on the field full-automatic patrol of the unmanned aerial vehicle is also increasingly improved, such as field automatic charging, automatic battery replacement, autonomous lifting and descending, emergency and police, patrol and the like of the unmanned aerial vehicle. Because of fixed unmanned aerial vehicle nest can not satisfy in the random item of patrolling and examining, remove the nest and need reequip on the pick up car fill to form pull type unmanned aerial vehicle and remove the airport.
For fixed unmanned aerial vehicle nest, the pull type unmanned aerial vehicle removes the nest and goes and select to stop in suitable place on the road surface and provide service for unmanned aerial vehicle including charging, trading electric etc.. When selecting a stopping point, not only the condition of the road surface (e.g., whether the flatness of the road surface meets the requirement, whether the road surface is tight, etc.) but also the environment around the point to be stopped (e.g., the surrounding is a complex traffic condition such as a highway, etc.) need to be considered, which is not suitable as the stopping point.
Therefore, a towed drone mobile airport that can intelligently select an appropriate stopping point when automatically patrolling is desired.
Disclosure of Invention
The present application is proposed to solve the above-mentioned technical problems. The embodiment of the application provides a pull type unmanned aerial vehicle mobile airport and operation method thereof, it is right through using the degree of depth convolution neural network model treat the image of stopping point with treat that the panoramic image around the stopping point encodes in order to be used for categorised feature extraction, just so can be based on treat the stopping point with treat that the characteristic information around the stopping point synthesizes right whether the stopping point is suitable for the parking and accurately categorised the judgement, and then make pull type unmanned aerial vehicle mobile airport can intelligently select suitable stopping point when patrolling and examining automatically to guarantee unmanned aerial vehicle's normal work and security.
According to an aspect of the present application, there is provided a towed drone mobile airport comprising:
the system comprises a to-be-stopped point image acquisition unit, a to-be-stopped point image acquisition unit and a to-be-stopped point image acquisition unit, wherein the to-be-stopped point image acquisition unit is used for acquiring an image of a to-be-stopped point through an overlooking camera deployed in a mobile airport of the towed unmanned aerial vehicle;
the panoramic image acquisition unit is used for acquiring panoramic images around the point to be stopped through a panoramic camera deployed in the towed unmanned aerial vehicle mobile airport;
the first convolution coding unit is used for enabling the image of the point to be stopped to pass through a first convolution neural network to obtain a first feature map, wherein the first convolution neural network comprises a deep-shallow feature fusion module;
a second convolution coding unit, configured to pass the panoramic image around the point to be stopped through a second convolution neural network using a spatial attention mechanism to obtain a second feature map;
a global normalization modification unit configured to perform global normalization of the first feature map with respect to the second feature map to generate a corrected second feature map, wherein the global normalization is performed based on a ratio between a logarithmic value of a sum of feature values at respective positions in the first feature map and a logarithmic value of a sum of feature values at all positions in the second feature map;
a feature map fusion unit, configured to fuse the first feature map and the corrected second feature map to generate a classification feature map; and
and the decision unit is used for enabling the classification characteristic graph to pass through a classifier to obtain a classification result, and the classification result is used for indicating whether the point to be stopped is suitable for parking.
In the above-mentioned trailing unmanned aerial vehicle mobile airport, the first convolution coding unit is configured to pass through a first convolution neural network the image of the point to be stopped to obtain a first feature map, where the first convolution neural network includes a deep-shallow feature fusion module, including: a shallow feature extraction subunit, configured to obtain a shallow feature map from an mth layer of the first convolutional neural network, where M is greater than or equal to 1 and less than or equal to 6; a deep feature map extraction subunit, configured to obtain a deep feature map from a last layer of the first convolutional neural network; a fusion subunit, configured to fuse the shallow feature map and the deep feature map through a deep-shallow feature fusion module of the first convolutional neural network to obtain the first feature map.
In above-mentioned pull type unmanned aerial vehicle removes airport, the subunit that fuses is further used for: fusing the shallow feature map and the deep feature map using the deep-shallow feature fusion module to obtain the first feature map with the following formula:
F 1 ＝αF s +βF d
wherein, F 1 Is the first characteristic diagram, F s As the shallow feature map, F d For the deep feature map, "+" indicates the addition of elements at corresponding positions of the shallow feature map and the deep feature map, and α and β are weighting parameters for controlling the balance between the shallow feature map and the deep feature map in the first feature map.
In the above towed unmanned aerial vehicle mobile airport, the second convolutional encoding unit is further configured to: each layer of the second convolutional neural network model performs, in forward pass of the layer, on input data: performing convolution processing based on a two-dimensional convolution kernel on the input data to generate a convolution characteristic diagram; pooling the convolved feature map to generate a pooled feature map; performing activation processing on the pooled feature map to generate an activated feature map; performing global average pooling along a channel dimension on the activation feature map to obtain a spatial feature matrix; performing convolution processing and activation processing on the spatial feature matrix to generate a weight vector; weighting each feature matrix of the activation feature map by using the weight value of each position in the weight vector to obtain a generated feature map; wherein the generated feature map output by the last layer of the second convolutional neural network model is the second feature map.
In the above mobile airport for towed unmanned aerial vehicle, the global normalization correction unit is further configured to: globally normalizing the first feature map relative to the second feature map to generate the corrected second feature map in the following formula;
wherein the formula is:
Figure BDA0003652728560000031
wherein
Figure BDA0003652728560000032
Respectively, the first characteristic diagram F 1 The second characteristic diagram F 2 The (i, j, k) th position of (a).
In the above-mentioned towed unmanned aerial vehicle mobile airport, the decision unit is further configured to: the classifier processes the classification feature map to generate a classification result according to the following formula, wherein the formula is as follows: softmax { (W) n ,B n ):…:(W 1 ,B 1 ) L project (F), where project (F) represents the projection of the classification feature map as a vector, W 1 To W n As a weight matrix for each fully connected layer, B 1 To B n A bias matrix representing the layers of the fully connected layer.
According to another aspect of the application, a method of operating a towed drone mobile airport, comprising:
the method comprises the steps of acquiring images of points to be stopped through an overlooking camera deployed at a towed unmanned aerial vehicle mobile airport;
acquiring panoramic images around the point to be stopped by a panoramic camera deployed at the mobile airport of the towed unmanned aerial vehicle;
the image of the point to be stopped is processed through a first convolution neural network to obtain a first feature map, wherein the first convolution neural network comprises a deep-shallow feature fusion module;
passing the panoramic image around the point to be stopped through a second convolutional neural network using a spatial attention mechanism to obtain a second feature map;
global normalization of the first feature map with respect to the second feature map to generate a corrected second feature map, wherein the global normalization is performed based on a ratio between a logarithmic value of a sum of feature values of respective positions in the first feature map and a logarithmic value of a sum of feature values of all positions in the second feature map;
fusing the first feature map and the corrected second feature map to generate a classification feature map; and
and passing the classification characteristic diagram through a classifier to obtain a classification result, wherein the classification result is used for indicating whether the point to be stopped is suitable for stopping.
In the above operation method of the towed unmanned aerial vehicle mobile airport, passing the image of the point to be stopped through a first convolutional neural network to obtain a first feature map, includes: obtaining a shallow feature map from an Mth layer of the first convolutional neural network, wherein M is greater than or equal to 1 and less than or equal to 6; obtaining a deep profile from the last layer of the first convolutional neural network; fusing, by a deep-shallow feature fusion module of the first convolutional neural network, the shallow feature map and the deep feature map to obtain the first feature map.
In the above operation method of the towed unmanned aerial vehicle mobile airport, fusing the shallow feature map and the deep feature map by a deep-shallow feature fusion module of the first convolutional neural network to obtain the first feature map, including: fusing the shallow feature map and the deep feature map using the deep-shallow feature fusion module to obtain the first feature map with the following formula:
F 1 ＝αF s +βF d
wherein, F 1 Is the first characteristic diagram, F s As the shallow feature map, F d For the deep feature map, "+" indicates the addition of elements at corresponding positions of the shallow feature map and the deep feature map, and α and β are weighting parameters for controlling the balance between the shallow feature map and the deep feature map in the first feature map.
In the above method for operating a towed unmanned mobile airport, passing a panoramic image of the surrounding of the point to be stopped through a second convolutional neural network using a spatial attention mechanism to obtain a second feature map, the method includes: each layer of the second convolutional neural network model performs, in forward pass of the layer, on input data: performing convolution processing based on a two-dimensional convolution kernel on the input data to generate a convolution characteristic diagram; pooling the convolved feature map to generate a pooled feature map; performing activation processing on the pooled feature map to generate an activated feature map; performing global average pooling along a channel dimension on the activation feature map to obtain a spatial feature matrix; performing convolution processing and activation processing on the spatial feature matrix to generate a weight vector; weighting each feature matrix of the activation feature map by using the weight value of each position in the weight vector to obtain a generated feature map; wherein the generated feature map output by the last layer of the second convolutional neural network model is the second feature map.
In the above method for operating a mobile airport using a towed drone, global normalization of the first feature map with respect to the second feature map to generate a corrected second feature map includes: globally normalizing the first feature map relative to the second feature map to generate the corrected second feature map in the following formula;
wherein the formula is:
Figure BDA0003652728560000051
wherein
Figure BDA0003652728560000052
Respectively, the first characteristic diagram F 1 The second characteristic diagram F 2 The (i, j, k) th position of (a).
In the above operation method of the towed unmanned aerial vehicle mobile airport, passing the classification feature map through a classifier to obtain a classification result, the method includes: the classifier processes the classification feature map to generate a classification result according to the following formula;
wherein the formula is: softmax { (W) n ,B n ):…:(W 1 ,B 1 ) L project (F), wherein project (F) represents projecting the classification feature map as a vector, W 1 To W n As a weight matrix for each fully connected layer, B 1 To B n A bias matrix representing the layers of the fully connected layer.
Compared with the prior art, the application provides a pull type unmanned aerial vehicle removes airport and operation method thereof, it is right through using the degree of depth convolution neural network model treat the stopping point the image with treat that the panoramic image around the stopping point encodes in order to be used for categorised feature extraction, just so can be based on treat the stopping point with treat that the characteristic information around the stopping point comes to synthesize right treat whether the stopping point is suitable to park and carry out accurate classification judgement, and then make pull type unmanned aerial vehicle remove airport can intelligently select suitable stopping point when patrolling and examining automatically to guarantee unmanned aerial vehicle's normal work and security.
Drawings
The above and other objects, features and advantages of the present application will become more apparent by describing in more detail embodiments of the present application with reference to the attached drawings. The accompanying drawings are included to provide a further understanding of the embodiments of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings, like reference numbers generally represent like parts or steps.
Fig. 1 is an application scenario diagram of a mobile airport for a towed drone according to an embodiment of the present application.
Fig. 2 is a block diagram of a towed drone mobile airport according to an embodiment of the present application.
Fig. 3 is a block diagram of a first volume-coded unit in a towed drone mobile airport according to an embodiment of the present application.
Fig. 4 is a flowchart of an operation method of a mobile airport for a towed drone according to an embodiment of the present application.
Fig. 5 is a schematic configuration diagram of an operation method of a mobile airport for a towed drone according to an embodiment of the present application.
Detailed Description
Hereinafter, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. It should be understood that the described embodiments are only some embodiments of the present application and not all embodiments of the present application, and that the present application is not limited by the example embodiments described herein.
Overview of a scene
As described above, with the rapid development of economy, the demand of the whole society for social energy is increasing, and the power line with ultrahigh voltage and large capacity is greatly expanded, so that the power line needs to span various complex environments, such as mountainous areas, basins, reservoirs, lakes, and the like, which all bring great difficulty to the maintenance and repair of the high-voltage power line, especially in forest areas, high-altitude areas, ice and snow covered areas, mountain landslide areas, areas where geological disasters are likely to occur, and the like. And the application of the unmanned aerial vehicle makes the power inspection and the emergency repair of the areas easy.
Along with the more mature and stable unmanned aerial vehicle technique, also more and more are being used in the field of patrolling and examining. At present, the inspection of the unmanned aerial vehicle is controlled by workers through an inspection platform, and the overhead high-voltage transmission line is inspected and overhauled through airborne sensing equipment. Along with the improvement of the intelligent patrol requirement, the requirement on the field full-automatic patrol of the unmanned aerial vehicle is also increasingly improved, such as field automatic charging, automatic battery replacement, autonomous lifting and descending, emergency and police, patrol and the like of the unmanned aerial vehicle. Because of fixed unmanned aerial vehicle nest can not satisfy in the random item of patrolling and examining, remove the nest and need reequip on the pick up car fill to form pull type unmanned aerial vehicle and remove the airport.
For fixed unmanned aerial vehicle nest, the pull type unmanned aerial vehicle removes the nest and goes and select to stop in suitable place on the road surface and provide service for unmanned aerial vehicle including charging, trading electric etc.. When selecting a stopping point, not only the condition of the road surface (e.g., whether the flatness of the road surface meets the requirement, whether the road surface is tight, etc.) but also the environment around the point to be stopped (e.g., the surrounding is a complex traffic condition such as a highway, etc.) need to be considered, which is not suitable as the stopping point.
Therefore, a towed drone mobile airport that can intelligently select an appropriate stopping point when automatically patrolling is desired.
It should be understood that the unmanned aerial vehicle nest is used for realizing functions of takeoff, landing, storage, automatic charging/battery replacement and the like of the unmanned aerial vehicle. The full-automatic nest assembly comprises: the system comprises a machine nest box body, a platform lifting and centering system, an electrical system, a battery replacing and charging system, a temperature control system, a UPS (optional), an industrial control system, a machine nest internal video monitoring system and the like.
In current unmanned aerial vehicle removes the nest, need put the car hopper with removing the nest, like this, through vehicle manufacture factory repacking back, can't pass through the requirement and the regulation that the car pipe was required, and then lead to the vehicle can't be listed or the difficulty of listing. And when the nest is installed, the nest not fixed on the vehicle is not convenient to carry and install, and is inconvenient to fix, and meanwhile, the nest has certain requirements on pickup trucks, so that the utilization rate of the vehicle is low, and the specified requirements on the size of a hopper behind the vehicle are met, and the safety of the unmanned aerial vehicle in the landing process is also reduced.
Based on this, the applicant of the present application chooses to separate the vehicle and the nest, presenting it in the form of a mobile nest of a towed drone. Can make car hopper itself need not reequip like this to avoid the difficulty of registering license, pull formula unmanned aerial vehicle removes the quick-witted nest and matches with the multi-vehicle type simultaneously, be fit for all kinds of vehicle and pull, its light in weight is not bulky, with the horizontal leveling of floorpan easily, does benefit to taking off and descending of unmanned aerial vehicle.
It should be understood that in selecting the stopping point, not only the case of the stopping point, for example, whether the road surface is smooth, whether the road surface is tight, etc., but also the case around the stopping point, need to be considered. This is essentially a classification problem, i.e. a classification decision is made based on the features of the acquisition stopping point and the surrounding stopping point: whether the point to be stopped is suitable for parking.
In recent years, deep learning and neural networks have been widely used in the fields of computer vision, natural language processing, text signal processing, and the like. In addition, deep learning and neural networks also exhibit a level close to or even exceeding that of humans in the fields of image classification, object detection, semantic segmentation, text translation, and the like. The development of deep learning and neural network provides solution thinking and scheme for the intelligent selection of the stopping point of the mobile airport of the towed unmanned aerial vehicle.
Specifically, in the technical solution of the present application, an image of a point to be stopped may be acquired by a camera (e.g., a common RGB camera) disposed in an overhead manner and deployed at a towed unmanned mobile airport, and a panoramic image of the surroundings of the point to be stopped may be acquired by a panoramic camera deployed at the towed unmanned mobile airport. In particular, here, the coupling and hook for the mobile nest of the towed unmanned aerial vehicle are connected with tractors such as off-road vehicles, SUVs, commercial vehicles and small trucks, and the synchronous travel and the continuity are finished by the synchronous brake and the tractors. And, the pull type unmanned aerial vehicle removes the aircraft nest outward appearance is square, regular, symmetrical, has cuboid, quasi-cuboid. Meanwhile, the mobile nest of the trailer type unmanned aerial vehicle is provided with one side expansion device, and the compartment body can slide outwards by up to one meter by only lightly pressing a button, so that the taking-off and landing safety of the unmanned aerial vehicle is improved.
In view of the fact that the image of the point to be stopped and the panoramic image around the point to be stopped are image data, and the convolutional neural network model has excellent performance in processing the image data, in the technical solution of the present application, the image of the point to be stopped and the panoramic image around the point to be stopped are encoded using a deep convolutional neural network model for feature extraction for classification.
Correspondingly, the image of the point to be stopped is coded by a first convolution neural network. When selecting a point to be stopped, not only the shallow characteristics of the point to be stopped, including shape characteristics (e.g., whether there is a protrusion, a depression, etc.) and texture characteristics (e.g., whether there is a line, a crack, etc.) but also the deep characteristics of the point to be stopped, including the type of road surface (asphalt road, highway, mud road, etc.), etc., need to be concerned. When the convolutional neural network is coded, as the coding depth is deepened, the shallow features of the convolutional neural network are submerged or lost. Therefore, in the technical solution of the present application, the encoding method of the first convolutional neural network is adjusted. Specifically, at least one shallow feature map is extracted from a shallow layer (generally 1-6) of the first convolutional neural network and at least one deep feature map is extracted from a deep layer (for example, the last layer of the first convolutional neural network) of the first convolutional neural network, and then the at least one shallow feature map and the at least one deep feature map are fused by using a deep-shallow feature fusion module of the first convolutional neural network to fuse the shallow feature and the deep feature of the point to be stopped so as to improve the accuracy of subsequent classification judgment.
And similarly, coding the panoramic image around the point to be stopped by using a second convolutional neural network. The panoramic image has a larger view field relative to the overhead image of the point to be stopped, and considering that when the classification judgment is performed, the influence of each spatial position (relative to the position of the point to be stopped) in the panoramic image on the classification judgment is different, for example, in the process of actually selecting the point to be stopped, the panoramic image focuses more on an area close to the stopping point, and if a traffic sign exists around the point to be stopped, the spatial positions are focused. This can be achieved by a spatial attention mechanism. That is, a second convolutional neural network with a spatial attention mechanism is used to encode the panoramic image around the point to be stopped to obtain a second feature map. And then, fusing the first feature map and the second feature map and carrying out classification judgment through a classifier.
However, although the local feature in the second feature map is focused by using the spatial attention mechanism in the feature extraction process of the second feature map, the entire feature expression is still focused more than the first feature map, and therefore, if the first feature map and the second feature map are fused directly using the point addition method, balance may not be achieved between the local feature and the entire feature. Based on this, global normalization of the first feature map focused on the local feature expression with respect to the second feature map is performed, namely:
Figure BDA0003652728560000091
wherein
Figure BDA0003652728560000092
Respectively, a first characteristic diagram F 1 The second characteristic diagram F 2 The (i, j, k) th position of (a).
Through the global normalization, robustness surrounding the global characteristic information minimum loss for the second characteristic diagram can be introduced into the characteristic value of the first characteristic diagram, clustering performance of the local characteristic expression of the first characteristic diagram towards the global characteristic expression of the second characteristic diagram is achieved, and therefore the dependency of the position-based characteristic expression of the first characteristic diagram on the global expected distribution is improved. Thus, the accuracy of classification judgment can be improved.
Based on this, this application provides a pull type unmanned aerial vehicle removes airport, and it includes: the system comprises a to-be-stopped point image acquisition unit, a to-be-stopped point image acquisition unit and a to-be-stopped point image acquisition unit, wherein the to-be-stopped point image acquisition unit is used for acquiring an image of a to-be-stopped point through an overlooking camera deployed in a mobile airport of the towed unmanned aerial vehicle; the panoramic image acquisition unit is used for acquiring panoramic images around the point to be stopped through a panoramic camera deployed in the towed unmanned aerial vehicle mobile airport; the first convolution coding unit is used for enabling the image of the point to be stopped to pass through a first convolution neural network to obtain a first feature map, wherein the first convolution neural network comprises a deep-shallow feature fusion module; a second convolution coding unit, configured to pass the panoramic image around the point to be stopped through a second convolution neural network using a spatial attention mechanism to obtain a second feature map; a global normalization modification unit configured to perform global normalization of the first feature map with respect to the second feature map to generate a corrected second feature map, wherein the global normalization is performed based on a ratio between a logarithmic value of a sum of feature values at respective positions in the first feature map and a logarithmic value of a sum of feature values at all positions in the second feature map; a feature map fusion unit, configured to fuse the first feature map and the corrected second feature map to generate a classification feature map; and the decision unit is used for enabling the classification characteristic graph to pass through a classifier to obtain a classification result, and the classification result is used for indicating whether the point to be stopped is suitable for parking.
Fig. 1 illustrates an application scenario diagram of a mobile airport for a towed drone according to an embodiment of the present application. As shown in fig. 1, in this application scenario, first, an image of a point to be stopped is acquired by a camera (e.g., a general RGB camera C as illustrated in fig. 1) disposed in an overhead manner deployed at a towed unmanned mobile airport (e.g., M as illustrated in fig. 1), and a panoramic image of the surroundings of the point to be stopped is acquired by a panoramic camera (e.g., F as illustrated in fig. 1) deployed at the towed unmanned mobile airport. Then, the obtained image of the point to be stopped and the panoramic image of the surrounding of the point to be stopped are input into a server (for example, a cloud server S as illustrated in fig. 1) deployed with a towed unmanned aerial vehicle mobile airport algorithm, wherein the server can process the image of the point to be stopped and the panoramic image of the surrounding of the point to be stopped with the towed unmanned aerial vehicle mobile airport algorithm to generate a classification result representing whether the point to be stopped is suitable for parking.
Having described the general principles of the present application, various non-limiting embodiments of the present application will now be described with reference to the accompanying drawings.
Exemplary System
Fig. 2 illustrates a block diagram of a towed drone mobile airport according to an embodiment of the present application. As shown in fig. 2, a mobile airport 200 for a towed drone according to an embodiment of the present application includes: a to-be-stopped point image obtaining unit 210, configured to obtain an image of a to-be-stopped point through an overhead camera deployed in a mobile airport of the towed unmanned aerial vehicle; a panoramic image acquisition unit 220, configured to acquire a panoramic image of the periphery of the point to be stopped through a panoramic camera deployed in the mobile airport of the towed unmanned aerial vehicle; a first convolution coding unit 230, configured to pass the image of the point to be stopped through a first convolution neural network to obtain a first feature map, where the first convolution neural network includes a deep-shallow feature fusion module; a second convolution encoding unit 240, configured to pass the panoramic image around the point to be stopped through a second convolution neural network using a spatial attention mechanism to obtain a second feature map; a global normalization modification unit 250 configured to perform global normalization on the first feature map with respect to the second feature map to generate a corrected second feature map, wherein the global normalization is performed based on a ratio between a logarithmic value of a sum of feature values at respective positions in the first feature map and a logarithmic value of a sum of feature values at all positions in the second feature map; a feature map fusion unit 260, configured to fuse the first feature map and the corrected second feature map to generate a classification feature map; and a decision unit 270, configured to pass the classification feature map through a classifier to obtain a classification result, where the classification result is used to indicate whether the point to be stopped is suitable for parking.
Specifically, in the embodiment of the present application, the image obtaining unit 210 for a point to be stopped and the panoramic image obtaining unit 220 are configured to obtain an image of the point to be stopped through a camera deployed in an overhead view of a mobile airport of a trailer-type unmanned aerial vehicle, and obtain a panoramic image of the surroundings of the point to be stopped through a panoramic camera deployed in the mobile airport of the trailer-type unmanned aerial vehicle. As mentioned earlier, in current unmanned aerial vehicle removes the nest, need put the removal nest at the car hopper, like this, after vehicle manufacture factory repacking, can't pass through the requirement and the regulation that the car pipe was required, and then lead to the unable license plate of vehicle or the difficulty of license plate. And when the nest is installed, the nest which is not fixed on the vehicle is inconvenient to carry and install and is inconvenient to fix, and meanwhile, the nest has certain requirements on pickup trucks, so that the utilization rate of the vehicle is low, the size of a hopper behind the vehicle has specified requirements, and the safety of the unmanned aerial vehicle in the landing process is also reduced. Therefore, in the technical scheme of this application, choose to separate the vehicle with the quick-witted nest, present in the form of pulling unmanned aerial vehicle mobile machine nest. Can make car hopper itself need not reequip like this to avoid the difficulty of registering license, pull formula unmanned aerial vehicle removes the quick-witted nest and matches with the multi-vehicle type simultaneously, be fit for all kinds of vehicle and pull, its light in weight is not bulky, with the horizontal leveling of floorpan easily, does benefit to taking off and descending of unmanned aerial vehicle.
That is, in particular, in the technical solutions of the present application, it should be understood that, when selecting the stopping point, not only the case of the stopping point, for example, whether the road surface is smooth, whether the road surface is tight, etc., but also the case around the stopping point, needs to be considered. This is essentially a classification problem, i.e. a classification decision is made based on the features of the acquisition stopping point and the surrounding stopping point: whether the point to be stopped is suitable for parking.
Specifically, in the technical solution of the present application, the camera arranged in the overhead manner deployed at the towed unmanned aerial vehicle mobile airport may be used to capture an image of a point to be stopped, for example, a common RGB camera and a panoramic camera deployed at the towed unmanned aerial vehicle mobile airport may capture a panoramic image of the surroundings of the point to be stopped. In particular, the pull-type unmanned aerial vehicle mobile nest requirement coupler and the hook are connected with a towing vehicle such as an off-road vehicle, an SUV, a business vehicle and a small truck, and synchronous running and continuity are finished through synchronous braking and the towing vehicle. And, the pull type unmanned aerial vehicle removes the aircraft nest outward appearance side is just, regular, symmetry, has cuboid, quasi-cuboid. Meanwhile, the mobile nest of the trailer type unmanned aerial vehicle is provided with one-side expansion equipment, and the distance that the carriage body can slide outwards by up to one meter is widened by only lightly pressing a button so as to increase the taking-off and landing safety of the unmanned aerial vehicle.
Specifically, in this embodiment of the present application, the first convolution encoding unit 230 is configured to pass the image of the point to be stopped through a first convolution neural network to obtain a first feature map, where the first convolution neural network includes a deep-shallow feature fusion module. It should be understood that, in the technical solution of the present application, the image of the point to be stopped and the panoramic image around the point to be stopped are encoded using the deep convolutional neural network model for feature extraction for classification, considering that the image of the point to be stopped and the panoramic image around the point to be stopped are image data, and the convolutional neural network model has an excellent performance in processing the image data.
Correspondingly, in the technical solution of the present application, the first convolutional neural network is used to encode the image of the point to be stopped. When selecting a point to be stopped, not only the shallow characteristics of the point to be stopped, including shape characteristics (e.g., whether a protrusion, a depression, etc. exist) and texture characteristics (e.g., whether a line of the point to be stopped, whether a crack exists, etc.), but also the deep characteristics of the point to be stopped, including the type of road surface (asphalt road, highway, mud road, etc.), etc., need to be concerned. However, when the convolutional neural network is coded, as the coding depth is increased, the shallow features of the convolutional neural network are submerged or lost. Therefore, in the technical solution of the present application, the encoding method of the first convolutional neural network is adjusted. Specifically, at least one shallow feature map is extracted from a shallow layer (generally 1 to 6) of the first convolutional neural network and at least one deep feature map is extracted from a deep layer (for example, the last layer of the first convolutional neural network) of the first convolutional neural network, and then the at least one shallow feature map and the at least one deep feature map are fused by using a deep-shallow feature fusion module of the first convolutional neural network to fuse the shallow feature and the deep feature of the point to be stopped so as to improve the accuracy of subsequent classification judgment, thereby obtaining the first feature map.
More specifically, in an embodiment of the present application, the first convolution encoding unit includes: and the shallow feature extraction subunit is used for obtaining a shallow feature map from the Mth layer of the first convolutional neural network, wherein M is greater than or equal to 1 and less than or equal to 6. A deep feature map extraction subunit, configured to obtain a deep feature map from a last layer of the first convolutional neural network. A fusion subunit, configured to fuse the shallow feature map and the deep feature map through a deep-shallow feature fusion module of the first convolutional neural network to obtain the first feature map. Accordingly, in one particular example, the shallow feature map and the deep feature map are fused using the deep-shallow feature fusion module to obtain the first feature map with the formula:
F 1 ＝αF s +βF d
wherein, F 1 Is the first characteristic diagram, F s As the shallow feature map, F d For the deep feature map, "+" indicates the addition of elements at corresponding positions of the shallow feature map and the deep feature map, and α and β are weighting parameters for controlling the balance between the shallow feature map and the deep feature map in the first feature map.
Fig. 3 illustrates a block diagram of a first volume-coded unit in a towed drone mobile airport according to an embodiment of the present application. As shown in fig. 3, the first convolution encoding unit 230 includes: a shallow feature extraction subunit 231, configured to obtain a shallow feature map from an mth layer of the first convolutional neural network, where M is greater than or equal to 1 and less than or equal to 6; a deep feature map extraction subunit 232, configured to obtain a deep feature map from the last layer of the first convolutional neural network; a fusion subunit 233, configured to fuse the shallow feature map and the deep feature map through a deep-shallow feature fusion module of the first convolutional neural network to obtain the first feature map.
Specifically, in this embodiment of the application, the second convolution encoding unit 240 is configured to pass the panoramic image around the point to be stopped through a second convolution neural network using a spatial attention mechanism to obtain a second feature map. That is, in the technical solution of the present application, similarly, the panoramic image around the point to be stopped is encoded by the second convolutional neural network. It should be understood that the panoramic image has a larger field of view relative to the top view image of the point to be stopped, and considering that there is a difference in the influence of each spatial position (relative to the position of the point to be stopped) in the panoramic image on the classification judgment when performing the classification judgment, for example, in the process of actually selecting the point to be stopped, a region closer to the stopping point is more focused, and if there is a traffic sign around the point to be stopped, these spatial positions should be focused. This can be achieved by a spatial attention mechanism. That is, a second convolutional neural network with a spatial attention mechanism is used to encode the panoramic image around the point to be stopped to obtain a second feature map.
More specifically, in this embodiment of the present application, the second convolutional encoding unit is further configured to: each layer of the second convolutional neural network model performs, in forward pass of the layer, on input data: performing convolution processing based on a two-dimensional convolution kernel on the input data to generate a convolution characteristic diagram; pooling the convolved feature map to generate a pooled feature map; performing activation processing on the pooled feature map to generate an activated feature map; performing global average pooling along a channel dimension on the activation feature map to obtain a spatial feature matrix; performing convolution processing and activation processing on the spatial feature matrix to generate a weight vector; weighting each feature matrix of the activation feature map by using the weight value of each position in the weight vector to obtain a generated feature map; wherein the generated feature map output by the last layer of the second convolutional neural network model is the second feature map.
Specifically, in this embodiment of the present application, the global normalization modification unit 250 is configured to perform global normalization on the first feature map with respect to the second feature map to generate a corrected second feature map, where the global normalization is performed based on a ratio between a logarithmic value of a sum of feature values of respective positions in the first feature map and a logarithmic value of a sum of feature values of all positions in the second feature map. It should be understood that, in the technical solution of the present application, originally, the first feature map and the second feature map are fused and classified and determined by a classifier. However, considering that although the local features in the second feature map are focused by using the spatial attention mechanism in the feature extraction process of the second feature map, the entire feature expression is still focused more than the first feature map, and therefore if the first feature map and the second feature map are fused directly using the point addition method, balance may not be achieved between the local features and the overall features. Therefore, in the technical solution of the present application, global normalization is further performed on the first feature map focused on the local feature expression with respect to the second feature map to generate the corrected second feature map.
More specifically, in this embodiment of the application, the global normalization correction unit is further configured to: globally normalizing the first feature map relative to the second feature map to generate the corrected second feature map in the following formula;
wherein the formula is:
Figure BDA0003652728560000141
wherein
Figure BDA0003652728560000142
Respectively, the first characteristic diagram F 1 The second characteristic diagram F 2 The (i, j, k) th position of (a). It will be appreciated that by the global normalization, the feature values for the first feature map may be introduced around the feature values for the second feature mapAnd global characterization information minimizes the robustness of loss to realize the clustering performance of the local feature expression of the first feature map to the global feature expression of the second feature map, thereby improving the dependency of the local feature expression of the first feature map on global expected distribution. Thus, the accuracy of classification judgment can be improved.
Specifically, in this embodiment of the application, the feature map fusing unit 260 and the decision unit 270 are configured to fuse the first feature map and the corrected second feature map to generate a classification feature map, and pass the classification feature map through a classifier to obtain a classification result, where the classification result is used to indicate whether the point to be stopped is suitable for parking. That is, in the technical solution of the present application, after the corrected second feature map is obtained, the first feature map and the corrected second feature map are further fused to generate a classification feature map, and the classification feature map is passed through a classifier to obtain a classification result indicating whether the point to be stopped is suitable for parking. Accordingly, in one specific example, the classifier processes the classification feature map to generate a classification result according to the following formula: softmax { (W) n ,B n ):…:(W 1 ,B 1 ) L project (F), where project (F) represents the projection of the classification feature map as a vector, W 1 To W n As a weight matrix for each fully connected layer, B 1 To B n A bias matrix representing the layers of the fully connected layer.
In conclusion, based on this application embodiment the mobile airport 200 of pull-type unmanned aerial vehicle is clarified, it is right through using the depth convolution neural network model the image of waiting to stop point and the panoramic image around waiting to stop point encode in order to be used for categorised feature extraction, just so can be based on wait to stop point and wait to stop point around the feature information come synthesize right wait to stop point whether suitable parking carries out accurate categorised judgement, and then make the mobile airport of pull-type unmanned aerial vehicle can intelligently select suitable stop point when patrolling and examining automatically to guarantee unmanned aerial vehicle's normal work and security.
As described above, the towed drone mobile airport 200 according to the embodiment of the present application may be implemented in various terminal devices, such as a server of a towed drone mobile airport algorithm, and the like. In one example, the towed drone mobile airport 200 according to embodiments of the present application may be integrated into the terminal device as one software module and/or hardware module. For example, the towed drone mobile airport 200 may be a software module in the operating system of the terminal device, or may be an application developed for the terminal device; of course, the towed drone mobile airport 200 may also be one of many hardware modules of the terminal device.
Alternatively, in another example, the towed drone mobile airport 200 and the terminal device may also be separate devices, and the towed drone mobile airport 200 may be connected to the terminal device through a wired and/or wireless network and transmit the interaction information in an agreed data format.
Exemplary method
Fig. 4 illustrates a flow chart of a method of operating a towed drone mobile airport. As shown in fig. 4, the method for operating a mobile airport with towed drones according to the embodiment of the present application includes the following steps: s110, acquiring an image of a point to be stopped through a downward camera deployed in a mobile airport of the towed unmanned aerial vehicle; s120, acquiring panoramic images around the point to be stopped through a panoramic camera deployed at the mobile airport of the trailer type unmanned aerial vehicle; s130, enabling the image of the point to be stopped to pass through a first convolutional neural network to obtain a first feature map, wherein the first convolutional neural network comprises a deep-shallow feature fusion module; s140, enabling the panoramic image around the point to be stopped to pass through a second convolutional neural network using a spatial attention mechanism to obtain a second feature map; s150, performing global normalization on the first feature map relative to the second feature map to generate a corrected second feature map, wherein the global normalization is performed based on a ratio of a logarithmic value of a summation value of feature values of all positions in the first feature map to a logarithmic value of a summation value of global summation values of feature values of all positions in the second feature map; s160, fusing the first feature map and the corrected second feature map to generate a classification feature map; and S170, passing the classification characteristic map through a classifier to obtain a classification result, wherein the classification result is used for indicating whether the point to be stopped is suitable for parking.
Fig. 5 illustrates an architecture diagram of an operation method of a mobile airport for a towed drone according to an embodiment of the application. As shown in fig. 5, in the network architecture of the operation method of the towed unmanned mobile airport, first, the acquired image of the point to be stopped (e.g., P1 as illustrated in fig. 5) is passed through a first convolutional neural network (e.g., CNN1 as illustrated in fig. 5) to obtain a first feature map (e.g., F1 as illustrated in fig. 5); then, passing the obtained panoramic image (e.g., P2 as illustrated in fig. 5) around the point to be stopped through a second convolutional neural network (e.g., CNN2 as illustrated in fig. 5) using a spatial attention mechanism to obtain a second feature map (e.g., F2 as illustrated in fig. 5); then, global normalization of the first feature map with respect to the second feature map is performed to generate a corrected second feature map (e.g., F as illustrated in fig. 5); then, fusing the first feature map and the corrected second feature map to generate a classification feature map (e.g., FC as illustrated in fig. 5); and finally, passing the classification feature map through a classifier (e.g., a classifier as illustrated in fig. 5) to obtain a classification result, wherein the classification result is used for indicating whether the point to be stopped is suitable for stopping.
More specifically, in steps S110 and S120, an image of a point to be stopped is captured by an overhead camera disposed at a towed drone mobile airport, and a panoramic image of the surroundings of the point to be stopped is captured by a panoramic camera disposed at the towed drone mobile airport. It can be understood that in current unmanned aerial vehicle removes the nest, need put the removal nest at the car hopper, like this, through vehicle manufacture factory repacking back, can't pass through the requirement and the regulation that the car pipe was required, and then lead to the unable tablet of registering or the tablet difficulty of registering of vehicle. And when the nest is installed, the nest which is not fixed on the vehicle is inconvenient to carry and install and is inconvenient to fix, and meanwhile, the nest has certain requirements on pickup trucks, so that the utilization rate of the vehicle is low, the size of a hopper behind the vehicle has specified requirements, and the safety of the unmanned aerial vehicle in the landing process is also reduced. Therefore, in the technical scheme of the application, the vehicle and the nest are separated and presented in the form of a mobile nest of a trailer type unmanned aerial vehicle. Can make car hopper itself need not reequip like this to avoid the difficulty of registering license, pull formula unmanned aerial vehicle removes the quick-witted nest and matches with the multi-vehicle type simultaneously, be fit for all kinds of vehicle and pull, its light in weight is not bulky, with the horizontal leveling of floorpan easily, does benefit to taking off and descending of unmanned aerial vehicle.
That is, in particular, in the technical solutions of the present application, it should be understood that, when selecting the stopping point, not only the case of the stopping point, for example, whether the road surface is smooth, whether the road surface is tight, etc., but also the case around the stopping point, needs to be considered. This is essentially a classification problem, i.e. a classification decision is made based on the features of the acquisition stopping point and the surrounding stopping point: whether the point to be stopped is suitable for parking.
Specifically, in the technical solution of the present application, the camera arranged in the overhead manner deployed at the towed unmanned aerial vehicle mobile airport may be used to capture an image of a point to be stopped, for example, a common RGB camera and a panoramic camera deployed at the towed unmanned aerial vehicle mobile airport may capture a panoramic image of the surroundings of the point to be stopped. In particular, the pull-type unmanned aerial vehicle mobile nest requirement coupler and the hook are connected with a towing vehicle such as an off-road vehicle, an SUV, a business vehicle and a small truck, and synchronous running and continuity are finished through synchronous braking and the towing vehicle. And, the pull type unmanned aerial vehicle removes the aircraft nest outward appearance side is just, regular, symmetry, has cuboid, quasi-cuboid. Meanwhile, the mobile nest of the trailer type unmanned aerial vehicle is provided with one-side expansion equipment, and the distance that the carriage body can slide outwards by up to one meter is widened by only lightly pressing a button so as to increase the taking-off and landing safety of the unmanned aerial vehicle.
More specifically, in step S130, the image of the point to be stopped is passed through a first convolutional neural network to obtain a first feature map, where the first convolutional neural network includes a deep-shallow feature fusion module. It should be understood that, in the technical solution of the present application, the image of the point to be stopped and the panoramic image around the point to be stopped are encoded using the deep convolutional neural network model for feature extraction for classification, considering that the image of the point to be stopped and the panoramic image around the point to be stopped are image data, and the convolutional neural network model has an excellent performance in processing the image data.
Correspondingly, in the technical solution of the present application, the first convolutional neural network is used to encode the image of the point to be stopped. When selecting a point to be stopped, not only the shallow characteristics of the point to be stopped, including shape characteristics (e.g., whether a protrusion, a depression, etc. exist) and texture characteristics (e.g., whether a line of the point to be stopped, whether a crack exists, etc.), but also the deep characteristics of the point to be stopped, including the type of road surface (asphalt road, highway, mud road, etc.), etc., need to be concerned. However, when the convolutional neural network is coded, as the coding depth is increased, the shallow features of the convolutional neural network are submerged or lost. Therefore, in the technical solution of the present application, the encoding method of the first convolutional neural network is adjusted. Specifically, at least one shallow feature map is extracted from a shallow layer (generally 1 to 6) of the first convolutional neural network and at least one deep feature map is extracted from a deep layer (for example, the last layer of the first convolutional neural network) of the first convolutional neural network, and then the at least one shallow feature map and the at least one deep feature map are fused by using a deep-shallow feature fusion module of the first convolutional neural network to fuse the shallow feature and the deep feature of the point to be stopped so as to improve the accuracy of subsequent classification judgment, thereby obtaining the first feature map.
More specifically, in S140, the panoramic image around the point to be stopped is passed through a second convolutional neural network using a spatial attention mechanism to obtain a second feature map. That is, in the technical solution of the present application, similarly, the panoramic image around the point to be stopped is encoded by the second convolutional neural network. It should be understood that the panoramic image has a larger field of view relative to the top view image of the point to be stopped, and considering that there is a difference in the influence of each spatial position (relative to the position of the point to be stopped) in the panoramic image on the classification judgment when performing the classification judgment, for example, in the process of actually selecting the point to be stopped, a region closer to the stopping point is more focused, and if there is a traffic sign around the point to be stopped, these spatial positions should be focused. This can be achieved by a spatial attention mechanism. That is, a second convolutional neural network with a spatial attention mechanism is used to encode the panoramic image around the point to be stopped to obtain a second feature map.
More specifically, in step S150, the first feature map is subjected to global normalization with respect to the second feature map to generate a corrected second feature map, wherein the global normalization is performed based on a ratio between a logarithmic value of a sum of feature values of respective positions in the first feature map and a logarithmic value of a sum of feature values of all positions in the second feature map. It should be understood that, in the technical solution of the present application, originally, the first feature map and the second feature map are fused and classified and determined by a classifier. However, considering that although the local features in the second feature map are focused by using the spatial attention mechanism in the feature extraction process of the second feature map, the entire feature expression is still focused more than the first feature map, and therefore if the first feature map and the second feature map are fused directly using the point addition method, balance may not be achieved between the local features and the overall features. Therefore, in the technical solution of the present application, global normalization is further performed on the first feature map focused on the local feature expression with respect to the second feature map, that is, the global normalization is performed by:
Figure BDA0003652728560000181
wherein
Figure BDA0003652728560000182
Respectively, the first characteristic diagram F 1 The second characteristic diagram F 2 The (i, j, k) th position of (a).
It should be understood that, through the global normalization, robustness surrounding a minimum loss of global characterizing information for the second feature map can be introduced into the feature values of the first feature map to achieve clustering performance of the local feature expression of the first feature map towards the global feature expression of the second feature map, thereby improving the dependency of the per-position feature expression of the first feature map on the global expected distribution. Thus, the accuracy of classification judgment can be improved.
More specifically, in step S160 and step S170, the first feature map and the corrected second feature map are fused to generate a classification feature map, and the classification feature map is passed through a classifier to obtain a classification result, where the classification result is used to indicate whether the point to be stopped is suitable for stopping. That is, in the technical solution of the present application, after the corrected second feature map is obtained, the first feature map and the corrected second feature map are further fused to generate a classification feature map, and the classification feature map is passed through a classifier to obtain a classification result indicating whether the point to be stopped is suitable for stopping. Accordingly, in one specific example, the classifier processes the classification feature map to generate a classification result according to the following formula: softmax { (W) n ,B n ):…:(W 1 ,B 1 ) L project (F), where project (F) represents the projection of the classification feature map as a vector, W 1 To W n As a weight matrix for each fully connected layer, B 1 To B n A bias matrix representing the layers of the fully connected layer.
In summary, the operation method of the mobile airport for the trailer unmanned aerial vehicle based on the embodiment of the application is clarified, and the image of the point to be stopped and the panoramic image around the point to be stopped are coded by using the deep convolutional neural network model to perform the feature extraction for classification, so that whether the point to be stopped is suitable for parking can be comprehensively classified and judged accurately based on the point to be stopped and the feature information around the point to be stopped, and a suitable stopping point can be intelligently selected when the mobile airport for the trailer unmanned aerial vehicle is automatically patrolled, so that the normal work and the safety of the unmanned aerial vehicle can be ensured.
The foregoing describes the general principles of the present application in conjunction with specific embodiments, however, it is noted that the advantages, effects, etc. mentioned in the present application are merely examples and are not limiting, and they should not be considered essential to the various embodiments of the present application. Furthermore, the foregoing disclosure of specific details is for the purpose of illustration and description and is not intended to be limiting, since the foregoing disclosure is not intended to be exhaustive or to limit the disclosure to the precise details disclosed.
The block diagrams of devices, apparatuses, systems referred to in this application are only given as illustrative examples and are not intended to require or imply that the connections, arrangements, configurations, etc. must be made in the manner shown in the block diagrams. These devices, apparatuses, devices, systems may be connected, arranged, configured in any manner, as will be appreciated by those skilled in the art. Words such as "including," "comprising," "having," and the like are open-ended words that mean "including, but not limited to," and are used interchangeably therewith. The words "or" and "as used herein mean, and are used interchangeably with, the word" and/or, "unless the context clearly dictates otherwise. The word "such as" is used herein to mean, and is used interchangeably with, the phrase "such as but not limited to".
It should also be noted that in the devices, apparatuses, and methods of the present application, the components or steps may be decomposed and/or recombined. These decompositions and/or recombinations are to be considered as equivalents of the present application.
The previous description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other aspects without departing from the scope of the application. Thus, the present application is not intended to be limited to the aspects shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
The foregoing description has been presented for purposes of illustration and description. Furthermore, the description is not intended to limit embodiments of the application to the form disclosed herein. While a number of example aspects and embodiments have been discussed above, those of skill in the art will recognize certain variations, modifications, alterations, additions and sub-combinations thereof.

Claims (10)

1. The utility model provides a pull type unmanned aerial vehicle removes airport which characterized in that includes:
the system comprises a to-be-stopped point image acquisition unit, a to-be-stopped point image acquisition unit and a to-be-stopped point image acquisition unit, wherein the to-be-stopped point image acquisition unit is used for acquiring an image of a to-be-stopped point through an overlooking camera deployed in a mobile airport of the towed unmanned aerial vehicle;
the panoramic image acquisition unit is used for acquiring panoramic images around the point to be stopped through a panoramic camera deployed in the towed unmanned aerial vehicle mobile airport;
the first convolution coding unit is used for enabling the image of the point to be stopped to pass through a first convolution neural network to obtain a first feature map, wherein the first convolution neural network comprises a deep-shallow feature fusion module;
a second convolution coding unit, configured to pass the panoramic image around the point to be stopped through a second convolution neural network using a spatial attention mechanism to obtain a second feature map;
a global normalization modification unit configured to perform global normalization of the first feature map with respect to the second feature map to generate a corrected second feature map, wherein the global normalization is performed based on a ratio between a logarithmic value of a sum of feature values at respective positions in the first feature map and a logarithmic value of a sum of feature values at all positions in the second feature map;
a feature map fusion unit, configured to fuse the first feature map and the corrected second feature map to generate a classification feature map; and
and the decision unit is used for enabling the classification characteristic graph to pass through a classifier to obtain a classification result, and the classification result is used for indicating whether the point to be stopped is suitable for parking.
2. The towed unmanned mobile airport of claim 1, wherein said first convolutional encoding unit is configured to pass said image of said point to be stopped through a first convolutional neural network to obtain a first feature map, wherein said first convolutional neural network comprises a deep-shallow feature fusion module, comprising:
a shallow feature extraction subunit, configured to obtain a shallow feature map from an mth layer of the first convolutional neural network, where M is greater than or equal to 1 and less than or equal to 6;
a deep feature map extraction subunit, configured to obtain a deep feature map from a last layer of the first convolutional neural network;
a fusion subunit, configured to fuse the shallow feature map and the deep feature map through a deep-shallow feature fusion module of the first convolutional neural network to obtain the first feature map.
3. The towed drone mobile airport of claim 2, wherein said blending subunit is further configured to: fusing the shallow feature map and the deep feature map using the deep-shallow feature fusion module to obtain the first feature map with the following formula:
F 1 ＝αF s +βF d
wherein, F 1 Is the first characteristic diagram, F s As the shallow feature map, F d For the deep feature map, "+" indicates the addition of elements at corresponding positions of the shallow feature map and the deep feature map, and α and β are weighting parameters for controlling the balance between the shallow feature map and the deep feature map in the first feature map.
4. The towed drone mobile airport of claim 3, wherein said second convolutional encoding unit is further to: each layer of the second convolutional neural network model performs, in forward pass of the layer, on input data:
performing convolution processing based on a two-dimensional convolution kernel on the input data to generate a convolution characteristic diagram;
pooling the convolved feature map to generate a pooled feature map;
performing activation processing on the pooled feature map to generate an activated feature map;
performing global average pooling along a channel dimension on the activation feature map to obtain a spatial feature matrix;
performing convolution processing and activation processing on the spatial feature matrix to generate a weight vector; and
weighting each feature matrix of the activation feature map by the weight value of each position in the weight vector to obtain a generated feature map;
wherein the generated feature map output by the last layer of the second convolutional neural network model is the second feature map.
5. The towed drone mobile airport of claim 4, wherein said global normalization correction unit is further to: globally normalizing the first feature map relative to the second feature map to generate the corrected second feature map in the following formula;
wherein the formula is:
Figure FDA0003652728550000021
wherein
Figure FDA0003652728550000022
Respectively, the first characteristic diagram F 1 The second characteristic diagram F 2 The (i, j, k) th position of (a).
6. The towed drone mobile airport of claim 5, wherein said decision unit is further configured to: the classifier processes the classification feature map to generate a classification result according to the following formula, wherein the formula is as follows: softmax { (W) n ,B n ):…:(W 1 ,B 1 ) L project (F), where project (F) represents the projection of the classification feature map as a vector, W 1 To W n As a weight matrix for each fully connected layer, B 1 To B n A bias matrix representing the layers of the fully connected layer.
7. An operation method of a mobile airport of a trailer type unmanned aerial vehicle is characterized by comprising the following steps:
the method comprises the steps of acquiring images of points to be stopped through an overlooking camera deployed at a towed unmanned aerial vehicle mobile airport;
acquiring panoramic images around the point to be stopped by a panoramic camera deployed at the mobile airport of the towed unmanned aerial vehicle;
the image of the point to be stopped is processed through a first convolution neural network to obtain a first feature map, wherein the first convolution neural network comprises a deep-shallow feature fusion module;
passing the panoramic image around the point to be stopped through a second convolutional neural network using a spatial attention mechanism to obtain a second feature map;
global normalization of the first feature map with respect to the second feature map to generate a corrected second feature map, wherein the global normalization is performed based on a ratio between a logarithmic value of a sum of feature values of respective positions in the first feature map and a logarithmic value of a sum of feature values of all positions in the second feature map;
fusing the first feature map and the corrected second feature map to generate a classification feature map; and
and passing the classification characteristic diagram through a classifier to obtain a classification result, wherein the classification result is used for indicating whether the point to be stopped is suitable for stopping.
8. The method of claim 7, wherein the step of passing the image of the point to be stopped through a first convolutional neural network to obtain a first feature map comprises:
obtaining a shallow feature map from an Mth layer of the first convolutional neural network, wherein M is greater than or equal to 1 and less than or equal to 6;
obtaining a deep profile from the last layer of the first convolutional neural network;
fusing, by a deep-shallow feature fusion module of the first convolutional neural network, the shallow feature map and the deep feature map to obtain the first feature map.
9. The method of operating a towed drone mobile airport of claim 8, wherein fusing the shallow feature map and the deep feature map by a deep-shallow feature fusion module of the first convolutional neural network to obtain the first feature map comprises: fusing the shallow feature map and the deep feature map using the deep-shallow feature fusion module to obtain the first feature map with the following formula:
F 1 ＝αF s +βF d
wherein, F 1 Is the first characteristic diagram, F s As the shallow feature map, F d For the deep feature map, "+" indicates the addition of elements at corresponding positions of the shallow feature map and the deep feature map, and α and β are weighting parameters for controlling the balance between the shallow feature map and the deep feature map in the first feature map.
10. The method of claim 9, wherein the step of passing the panoramic image of the surroundings of the point to be stopped through a second convolutional neural network using a spatial attention mechanism to obtain a second feature map comprises:
each layer of the second convolutional neural network model performs, in forward pass of the layer, on input data:
performing convolution processing based on a two-dimensional convolution kernel on the input data to generate a convolution characteristic diagram;
pooling the convolved feature map to generate a pooled feature map;
performing activation processing on the pooled feature map to generate an activated feature map;
performing global average pooling along a channel dimension on the activation feature map to obtain a spatial feature matrix;
performing convolution processing and activation processing on the spatial feature matrix to generate a weight vector; and
weighting each feature matrix of the activation feature map by the weight value of each position in the weight vector to obtain a generated feature map;
wherein the generated feature map output by the last layer of the second convolutional neural network model is the second feature map.
CN202210557467.0A 2022-05-19 2022-05-19 Pull type unmanned aerial vehicle mobile airport and operation method thereof Withdrawn CN114842362A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210557467.0A CN114842362A (en) 2022-05-19 2022-05-19 Pull type unmanned aerial vehicle mobile airport and operation method thereof

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210557467.0A CN114842362A (en) 2022-05-19 2022-05-19 Pull type unmanned aerial vehicle mobile airport and operation method thereof

Publications (1)

Publication Number Publication Date
CN114842362A true CN114842362A (en) 2022-08-02

Family

ID=82571570

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210557467.0A Withdrawn CN114842362A (en) 2022-05-19 2022-05-19 Pull type unmanned aerial vehicle mobile airport and operation method thereof

Country Status (1)

Country Link
CN (1) CN114842362A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116630828A (en) * 2023-05-30 2023-08-22 中国公路工程咨询集团有限公司 Unmanned aerial vehicle remote sensing information acquisition system and method based on terrain environment adaptation

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116630828A (en) * 2023-05-30 2023-08-22 中国公路工程咨询集团有限公司 Unmanned aerial vehicle remote sensing information acquisition system and method based on terrain environment adaptation
CN116630828B (en) * 2023-05-30 2023-11-24 中国公路工程咨询集团有限公司 Unmanned aerial vehicle remote sensing information acquisition system and method based on terrain environment adaptation

Similar Documents

Publication Publication Date Title
US10678255B2 (en) Systems, methods and apparatuses are provided for enhanced surface condition detection based on image scene and ambient light analysis
US10223910B2 (en) Method and apparatus for collecting traffic information from big data of outside image of vehicle
WO2022206942A1 (en) Laser radar point cloud dynamic segmentation and fusion method based on driving safety risk field
KR101395094B1 (en) Method and system for detecting object in input image
CN112381784A (en) Equipment detecting system based on multispectral image
US11669947B2 (en) Machine learning assisted image analysis
CN114842362A (en) Pull type unmanned aerial vehicle mobile airport and operation method thereof
CN115909092A (en) Light-weight power transmission channel hidden danger distance measuring method and hidden danger early warning device
CN112489072A (en) Vehicle-mounted video perception information transmission load optimization method and device
CN115686063A (en) Mobile unmanned aerial vehicle and nest management and control system for distribution network circuit inspection
Yang et al. Recognition of taxi violations based on semantic segmentation of PSPNet and improved YOLOv3
CN112509321A (en) Unmanned aerial vehicle-based driving control method and system for urban complex traffic situation and readable storage medium
CN111898427A (en) Multispectral pedestrian detection method based on feature fusion deep neural network
CN114677848B (en) Perception early warning system, method, device and computer program product
CN116152750A (en) Vehicle feature recognition method based on monitoring image
CN114943903B (en) Self-adaptive clustering target detection method for aerial image of unmanned aerial vehicle
CN116229424A (en) Intelligent networking automobile pedestrian detection method based on improved YOLO v3
CN112115767B (en) Tunnel foreign matter detection method based on Retinex and YOLOv3 models
JP2007323177A (en) Image processor
CN116018814A (en) Information processing device, information processing method, and program
CN114255452A (en) Target ranging method and device
Hautière et al. Combination of roadside and in-vehicle sensors for extensive visibility range monitoring
Astudillo et al. Reducing the breach between simulated and real data for top view images
CN116978236B (en) Traffic accident early warning method, device and storage medium
Yi et al. Clothing color of surrogate bicyclist for pre-collision system evaluation

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
WW01 Invention patent application withdrawn after publication
WW01 Invention patent application withdrawn after publication

Application publication date: 20220802