Summary of the invention
Inventor studies the method for avoidance realizing unmanned aerial vehicle group.The prior art controls between unmanned plane progress group
Avoidance process is all that the control progress avoidance of platform is passively controlled by unmanned plane, and there are errors with timely for such avoidance mode
The problems such as prolonging.
It has been recognised by the inventors that a kind of mode is, artificial designed image feature extraction algorithm detects obstacle and complete from image
It is positioned at obstacle, UAV Flight Control rule is then designed according to obstacle positioning result, guidance unmanned plane is completed independently to keep away between group
Barrier process.Another way is to detect deep learning network model using trained image object, barrier is identified from image
Hinder and complete obstacle in image and position, UAV Flight Control rule is then designed according to obstacle positioning result, guidance unmanned plane is complete
Automatic obstacle avoiding process between in groups.
However, either traditional method for designing special image detection algorithm for specific objective, or using deeply
The obstacle spent in learning network model prediction image and the method for attempting obstacle positioning, have all only completed and have detected to hinder from image
The part hindered, if to apply the scene for independently completing task in unmanned plane, unmanned vehicle, robot etc., it is still necessary to according to obstacle
Detection as a result, special design control law generates corresponding control amount, complete specific task.The design of suitable control law and
Optimizing and revising for parameter is all the work taken time and effort, and effect is not necessarily highly desirable.Therefore, it needs to find a kind of nothing at present
The man-machine method that can be realized automatic obstacle avoiding between group.
The technical problem that the present invention solves is how to realize automatic obstacle avoiding of the unmanned plane between group.
According to an aspect of an embodiment of the present invention, a kind of unmanned plane barrier-avoiding method is provided, comprising: take the photograph using unmanned plane
As head acquires obstacle unmanned plane image;By obstacle unmanned plane image input deep learning network model trained in advance, kept away
Hinder flight control amount, deep learning network model is practical by input obstacle unmanned plane training image and corresponding unmanned plane
Flight control amount is trained;Flight according to avoidance flight control amount control unmanned plane.
In some embodiments, avoidance flight control amount is four dimensional vectors, including three-dimensional line rate controlling amount and
Rotate horizontally control amount.
In some embodiments, unmanned plane barrier-avoiding method further includes trained deep learning network model, training deep learning
Network model includes: that control unmanned plane hides obstacle unmanned plane in flight course;By unmanned plane in flight course unmanned plane
The obstacle unmanned plane training image of camera acquisition and corresponding unmanned plane practical flight control amount input deep learning network
Model, to be trained to deep learning network model, enable deep learning network model according to the obstacle of input nobody
Machine image prediction avoidance flight control amount.
In some embodiments, training deep learning network model further include: by last of deep learning network model
The output fruiting quantities of a full articulamentum are set as 4, to respectively indicate three-dimensional line rate controlling amount and rotate horizontally control amount.
In some embodiments, training deep learning network model further include: calculate the avoidance flight control that prediction obtains
Error between amount and unmanned plane practical flight control amount;Error is fed back to the loss layer of deep learning network model, so as to
Loss layer calculates the undated parameter of deep learning network model using Euclidean distance loss function, and using undated parameter to depth
Learning network model is updated.
In some embodiments, training deep learning network model further include: in last of deep learning network model
Layer addition normalization layer, normalization layer are normalized using Sigmoid function, to export normalized avoidance flight
Control amount.
Other side according to an embodiment of the present invention provides a kind of avoidance unmanned plane, comprising: camera, for adopting
Collect obstacle unmanned plane image;Flight control amount prediction meanss, for the depth that the input of obstacle unmanned plane image is trained in advance
Practise network model, obtain avoidance flight control amount, deep learning network model be by input obstacle unmanned plane training image with
And corresponding unmanned plane practical flight control amount is trained;Flight controller, for according to avoidance flight control amount
Control the flight of unmanned plane.
In some embodiments, avoidance flight control amount is four dimensional vectors, including three-dimensional line rate controlling amount and
Rotate horizontally control amount.
In some embodiments, avoidance unmanned plane further includes deep learning network model training device, deep learning network
Model training apparatus is used for: control unmanned plane hides obstacle unmanned plane in flight course;By unmanned plane in flight course nothing
The obstacle unmanned plane training image of man-machine camera acquisition and corresponding unmanned plane practical flight control amount input deep learning
Network model enables deep learning network model according to the obstacle of input to be trained to deep learning network model
Unmanned plane image prediction avoidance flight control amount.
In some embodiments, deep learning network model training device is also used to: most by deep learning network model
The output fruiting quantities of the full articulamentum of the latter are set as 4, to respectively indicate three-dimensional line rate controlling amount and rotate horizontally control
Amount processed.
In some embodiments, deep learning network model training device is also used to: calculating the avoidance flight that prediction obtains
Error between control amount and unmanned plane practical flight control amount;Error is fed back to the loss layer of deep learning network model,
So that loss layer is using the undated parameter of Euclidean distance loss function calculating deep learning network model, and utilize undated parameter pair
Deep learning network model is updated.
In some embodiments, deep learning network model training device is also used to: deep learning network model most
Later layer addition normalization layer, normalization layer is normalized using Sigmoid function, to export normalized avoidance
Flight control amount.
Another aspect according to an embodiment of the present invention provides a kind of unmanned plane obstacle avoidance apparatus, wherein includes: storage
Device;And it is coupled to the processor of memory, processor is configured as executing above-mentioned based on instruction stored in memory
Unmanned plane barrier-avoiding method.
Another aspect according to an embodiment of the present invention provides a kind of computer readable storage medium, wherein computer
Readable storage medium storing program for executing is stored with computer instruction, and instruction realizes unmanned plane barrier-avoiding method above-mentioned when being executed by processor.
Unmanned plane barrier-avoiding method provided by the invention can be realized automatic obstacle avoiding of the unmanned plane between group.
By referring to the drawings to the detailed description of exemplary embodiment of the present invention, other feature of the invention and its
Advantage will become apparent.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.Below
Description only actually at least one exemplary embodiment be it is illustrative, never as to the present invention and its application or make
Any restrictions.Based on the embodiments of the present invention, those of ordinary skill in the art are not making creative work premise
Under all other embodiment obtained, shall fall within the protection scope of the present invention.
Fig. 1 is combined to introduce the unmanned plane barrier-avoiding method of one embodiment of the invention first.
Fig. 1 shows the flow diagram of the unmanned plane barrier-avoiding method of one embodiment of the invention.As shown in Figure 1, this reality
The unmanned plane barrier-avoiding method applied in example includes:
Step S102 acquires obstacle unmanned plane image using unmanned plane camera.
For example, fixing forward sight camera in front of unmanned plane body, unmanned plane is obtained by forward sight camera between group
The image of automatic obstacle avoiding.
Obstacle unmanned plane image input deep learning network model trained in advance is obtained avoidance flight by step S104
Control amount, deep learning network model are by input obstacle unmanned plane training image and corresponding unmanned plane practical flight control
What amount processed was trained.
Wherein, avoidance flight control amount is four dimensional vectors, including X, Y, the three-dimensional wire velocity control of Z-space reference axis
Amount and horizontal rotation control amount about the z axis.Rotate horizontally the yaw angle that control amount can be used to control unmanned plane, i.e. control nothing
The angle of man-machine heading and X-axis.
Step S106, the flight according to avoidance flight control amount control unmanned plane.
For example, flight control amount can be sent to the flight controller of unmanned plane, by flight controller according to the flight
Control amount controls the practical flight of unmanned plane.
Above example implements a kind of unmanned planes using deep learning end to end control from barrier-avoiding method between main group.?
In whole process, the Flight Control Law of artificial designed image feature extraction, disorder detection method and unmanned plane is not needed, end is utilized
The process that the output of UAV Flight Control amount is directly arrived in camera image input is completed to end deep learning network model, realizes nobody
Automatic obstacle avoiding under machine cluster state of flight, between each unmanned plane.
It will be understood by those skilled in the art that above-described embodiment also has during automatic obstacle avoiding between realizing unmanned aerial vehicle group
The characteristics of being simple and efficient.Because on every frame unmanned plane, it is only necessary to the camera that shooting direction is heading is equipped with, in flight side
Avoidance between realization group upwards.Between every unmanned plane can be realized the group on heading in the case where avoidance, unmanned plane cluster
Also it is achieved that avoidance effect.Meanwhile inventor devises feathering angle in flight control amount, in unmanned plane according to level
In the case that rotation angle realizes rotation, the heading and camera of unmanned plane change with the horizontal rotation of unmanned plane,
Therefore the avoidance that still can be realized heading, avoids the collision of unmanned plane cluster internal.
The method that deep learning network model is trained below with reference to Fig. 2 introduction.
Fig. 2 shows the flow diagrams of one embodiment of the invention being trained to deep learning network model.Such as
Shown in Fig. 2, the method being trained in the present embodiment to deep learning network model includes:
Step S202, control unmanned plane hide obstacle unmanned plane in flight course.
The step is mainly used for avoidance database between acquisition unmanned aerial vehicle group.The control process of unmanned plane can have manually and from
Dynamic two schemes, manual approach are that avoidance process between group is completed using remote controler remote manual control unmanned plane, and automatic scheme is will to hinder
Hinder the hovering of object unmanned plane aerial, the GPS of the unmanned plane based on acquisition data and both unmanned planes for serving as barrier is positioned in real time
Information controls unmanned plane automatic obstacle avoiding.Avoidance database needs while acquiring forward sight camera image and flight speed between unmanned aerial vehicle group
Control amount is spent, has to serve as barrier comprising another frame unmanned plane in image, also may include other background informations.Database
After the completion of acquisition, data can be split as to training set, verifying collection and test set three parts.
Step S204, by unmanned plane in flight course unmanned plane camera acquire obstacle unmanned plane training image and
Corresponding unmanned plane practical flight control amount inputs deep learning network model, to be trained to deep learning network model,
Enable deep learning network model according to the obstacle unmanned plane image prediction avoidance flight control amount of input.
The step is mainly used for the training of deep learning network model.Wherein, deep learning network model can specifically be adopted
With such as caffenet, vgg16, resnet etc..Wherein it is possible to by the last one full articulamentum of deep learning network model
Output fruiting quantities be set as 4, with respectively indicate three-dimensional line rate controlling amount and rotate horizontally control amount.Matching flight control
The output of amount processed.Optionally, the input layer of deep learning network model can be modified, original RGB triple channel input is changed to ash
Spend single channel input.Single channel inputs the operand that can reduce network, the operation of acceleration model network.It can also be in depth
The last layer addition normalization layer of network model is practised, normalization layer is normalized using Sigmoid function, so as to defeated
Normalized avoidance flight control amount out, to accelerate the convergence of deep learning neural network model.
It should be strongly noted that in the obstacle unmanned plane training image and unmanned plane of input unmanned plane camera acquisition
When practical flight control amount, it is ensured that the corresponding relationship of image and rate controlling amount.That is, each frame image should be with this
It is corresponding that the rate controlling amount at moment carries out stringent synchronization in time.Meanwhile in the training process, due to every obstacle unmanned plane
It all include obstacle unmanned plane in training image, there is no need to be labeled to target in image, deep learning neural network model
It can be realized automatically extracting for obstacle unmanned plane.
Step S206 calculates the mistake between the obtained avoidance flight control amount of prediction and unmanned plane practical flight control amount
Error is fed back to the loss layer of deep learning network model by difference, so that loss layer is calculated deeply using Euclidean distance loss function
The undated parameter of learning network model is spent, and deep learning network model is updated using undated parameter.Using it is European away from
From the convergence that loss function can accelerate deep learning network model.
In the training process, the effect of upper testing model training can be collected in verifying simultaneously, completes model parameter and finely tuned
Journey.Image and corresponding rate controlling amount are input to deep learning network model simultaneously when training, and by speed control
Amount is used as supervisory signals, and image passes through the reasoning operation of entire neural network model, provides predetermined speed control output of model
Amount carries out error calculation with the rate controlling amount supervisory signals inputted simultaneously with image.Then, model is carried out according to control information
The update of parameter carries out the input of next group of data, iterates after the completion of updating.Verifying collection is in the training process, examining
The effect of model training is tested, verifying collection data input network model can be carried out output prediction, checked by every certain number of iteration
The control information of collection is verified, the above training process is autonomous by deep learning frame after setting network model and relevant parameter
It carries out, while can be according to the effect of the error size judgment models training of the error convergence situation and verifying collection of training set.
Finally, the test process of trained model can be completed on database test set.It equally will test when test
It concentrates image and corresponding rate controlling amount to input network model, but no longer carries out model parameter renewal process at this time.According to mould
The error of corresponding speed control amount, carrys out the mould that final training of judgement is completed when the rate controlling amount output and data acquisition of type prediction
The indexs such as the error rate of type can be applied to practical flight after reaching certain threshold value.
The avoidance unmanned plane of one embodiment of the invention is introduced below with reference to Fig. 3.
Fig. 3 shows the structural schematic diagram of the avoidance unmanned plane of one embodiment of the invention.As shown in figure 3, the present embodiment
Avoidance unmanned plane 30 include:
Camera 302, for acquiring obstacle unmanned plane image.
Flight control amount prediction meanss 304, for the deep learning network that the input of obstacle unmanned plane image is trained in advance
Model, obtains avoidance flight control amount, and deep learning network model is by input obstacle unmanned plane training image and correspondence
Unmanned plane practical flight control amount be trained.
Flight controller 306, for the flight according to avoidance flight control amount control unmanned plane.
In some embodiments, avoidance flight control amount is four dimensional vectors, including three-dimensional line rate controlling amount and
Rotate horizontally control amount.
In some embodiments, avoidance unmanned plane 30 further includes deep learning network model training device 301, deep learning
Network model training device 301 is used for: control unmanned plane hides obstacle unmanned plane in flight course;Unmanned plane was being flown
The obstacle unmanned plane training image and the input of corresponding unmanned plane practical flight control amount that unmanned plane camera acquires in journey are deep
It spends learning network model and enables deep learning network model according to input to be trained to deep learning network model
Obstacle unmanned plane image prediction avoidance flight control amount.
In some embodiments, deep learning network model training device 301 is also used to: by deep learning network model
The output fruiting quantities of the last one full articulamentum are set as 4, to respectively indicate three-dimensional line rate controlling amount and horizontal rotation
Control amount.
In some embodiments, deep learning network model training device 301 is also used to: being calculated the avoidance that prediction obtains and is flown
Error between row control amount and unmanned plane practical flight control amount;Error is fed back to the loss of deep learning network model
Layer so that loss layer is using the undated parameter of Euclidean distance loss function calculating deep learning network model, and utilizes update ginseng
Several pairs of deep learning network models are updated.
In some embodiments, deep learning network model training device 301 is also used to: in deep learning network model
The last layer addition normalization layer, normalization layer is normalized using Sigmoid function, to export normalized keep away
Hinder flight control amount.
Fig. 4 shows the structural schematic diagram of one embodiment of unmanned plane obstacle avoidance apparatus of the present invention.As shown in figure 4, the reality
The unmanned plane obstacle avoidance apparatus 40 for applying example includes: memory 410 and the processor 420 for being coupled to the memory 410, processor
420 are configured as the instruction based on storage in store 410, execute the unmanned plane avoidance side in any one aforementioned embodiment
Method.
Wherein, memory 410 is such as may include system storage, fixed non-volatile memory medium.System storage
Device is for example stored with operating system, application program, Boot loader (Boot Loader) and other programs etc..
Unmanned plane obstacle avoidance apparatus 40 can also include input/output interface 430, network interface 440, memory interface 450 etc..
It can for example be connected by bus 460 between these interfaces 430,440,450 and memory 410 and processor 420.Wherein,
The input-output equipment such as input/output interface 430 is display, mouse, keyboard, touch screen provide connecting interface.Network interface
440 provide connecting interface for various networked devices.The external storages such as memory interface 440 is SD card, USB flash disk provide connection and connect
Mouthful.
The invention also includes a kind of computer readable storage mediums, are stored thereon with computer instruction, and the instruction is processed
Device realizes the unmanned plane barrier-avoiding method in any one aforementioned embodiment when executing.
It should be understood by those skilled in the art that, the embodiment of the present invention can provide as method, system or computer program
Product.Therefore, complete hardware embodiment, complete software embodiment or reality combining software and hardware aspects can be used in the present invention
Apply the form of example.Moreover, it wherein includes the computer of computer usable program code that the present invention, which can be used in one or more,
The calculating implemented in non-transient storage medium (including but not limited to magnetic disk storage, CD-ROM, optical memory etc.) can be used
The form of machine program product.
The present invention be referring to according to the method for the embodiment of the present invention, the process of equipment (system) and computer program product
Figure and/or block diagram describe.It should be understood that every one stream in flowchart and/or the block diagram can be realized by computer program instructions
The combination of process and/or box in journey and/or box and flowchart and/or the block diagram.It can provide these computer programs
Instruct the processor of general purpose computer, special purpose computer, Embedded Processor or other programmable data processing devices to produce
A raw machine, so that being generated by the instruction that computer or the processor of other programmable data processing devices execute for real
The device for the function of being specified in present one or more flows of the flowchart and/or one or more blocks of the block diagram.
These computer program instructions, which may also be stored in, is able to guide computer or other programmable data processing devices with spy
Determine in the computer-readable memory that mode works, so that it includes referring to that instruction stored in the computer readable memory, which generates,
Enable the manufacture of device, the command device realize in one box of one or more flows of the flowchart and/or block diagram or
The function of being specified in multiple boxes.
These computer program instructions also can be loaded onto a computer or other programmable data processing device, so that counting
Series of operation steps are executed on calculation machine or other programmable devices to generate computer implemented processing, thus in computer or
The instruction executed on other programmable devices is provided for realizing in one or more flows of the flowchart and/or block diagram one
The step of function of being specified in a box or multiple boxes.
The foregoing is merely presently preferred embodiments of the present invention, is not intended to limit the invention, it is all in spirit of the invention and
Within principle, any modification, equivalent replacement, improvement and so on be should all be included in the protection scope of the present invention.