Disclosure of Invention
(one) solving the technical problems
Aiming at the defects of the prior art, the invention provides the unmanned aerial vehicle inspection tracking distribution network line flight method based on edge calculation, which has the advantages of reducing the unmanned aerial vehicle calculation pressure and power consumption and prolonging the unmanned aerial vehicle operation time, thereby solving the problem of shorter operation time of the existing unmanned aerial vehicle.
(II) technical scheme
In order to achieve the advantages of reducing the calculated pressure and power consumption of the unmanned aerial vehicle and prolonging the operation time of the unmanned aerial vehicle, the invention adopts the following specific technical scheme:
an unmanned aerial vehicle inspection tracking distribution network line flight method based on edge calculation comprises the following steps:
s1, acquiring power distribution network image data by using a terminal node and analyzing and calculating;
s2, adaptively distributing the computing task of the terminal node to the edge cloud by adopting a preset method;
s3, judging whether the terminal node loses a power distribution network target according to a preset scheme;
and S4, fusing calculation results of the terminal nodes and the edge cloud through a prefabrication method and analyzing network distribution conditions.
Further, the edge cloud is an edge computing platform and is provided with a target tracking module, an information fusion module, an information management module and a communication module.
Furthermore, the terminal nodes are unmanned aerial vehicles connected with the edge computing platform, and each terminal node is provided with a target tracking module, a computing task segmentation module, an information fusion module and a motion detection module.
Further, in the step S2, adaptively distributing the computing task of the terminal node to the edge cloud by adopting a preset method includes judging the equipment temperature, the CPU utilization rate, the memory occupancy rate and the energy status of the unmanned aerial vehicle, and adaptively distributing the computing task to the edge cloud according to the real-time network status and the resource occupancy condition of the edge server.
Further, the computing any includes processing the power distribution network image data, planning the flight path, and analyzing the power distribution network equipment faults.
Further, the edge cloud further comprises an unmanned aerial vehicle with high computing capacity as an edge cloud node, and when the unmanned aerial vehicle cannot be normally connected with the edge cloud, the unmanned aerial vehicle moves the power distribution network image data to the edge cloud node.
Further, in the step S3, determining whether the terminal node loses the power distribution network target according to a preset scheme further includes the following steps:
s31, modeling the motion condition of a power distribution network target by a terminal node through a Kalman filter;
s32, setting a displacement threshold of a power distribution network target by adopting a motion model;
s33, judging whether the target displacement of the power distribution network exceeds a threshold range, if so, executing S34, otherwise, executing S35;
s34, judging whether a moving object exists by utilizing a motion detection module of the terminal node, if so, executing S35, otherwise, stopping the calculation task;
s35, continuing to calculate the task.
Furthermore, the terminal node adopts a correlation filter algorithm, and the edge computing platform adopts a tracking algorithm of a full convolution twin neural network deep learning model.
Further, the calculation formula of the full convolution twin neural network is as follows:
wherein x represents a current picture data sample, z represents next picture data, and the functionFor feature extractor embedding, the function g is used for similarity measurement, b1 represents offset, and the algorithm is a cross-correlation operation.
Further, in the step S4, the steps of fusing the calculation results of the terminal node and the edge cloud by a prefabricating method and analyzing the network distribution situation further include the following steps:
s41, obtaining a response diagram according to a full convolution twin neural network algorithm and determining an algorithm fusion coefficient;
s42, integrating a correlation filter algorithm of the terminal node and a deep learning algorithm of the edge cloud to update an algorithm model;
s43, fusing the calculation results returned by the terminal nodes and the edge cloud by using an information fusion module, and finally predicting the specific position of the power distribution network target.
(III) beneficial effects
Compared with the prior art, the unmanned aerial vehicle inspection tracking distribution network line flight method based on edge calculation has the following beneficial effects: the computing tasks are reasonably unloaded to the edge cloud by utilizing the task segmentation module, and the computing results are analyzed and fused by utilizing the information fusion module, so that the computing pressure and the power consumption of the end node are reduced; in addition, the calculation pressure and the power consumption of the terminal node can be further reduced by adding the motion detection module, and the operation time of the unmanned aerial vehicle is prolonged, so that the efficiency of the unmanned aerial vehicle in inspecting the distribution network cable is improved, and the safety of the distribution network cable in conveying electric energy is ensured.
Detailed Description
For the purpose of further illustrating the various embodiments, the present invention provides the accompanying drawings, which are a part of the disclosure of the present invention, and which are mainly used to illustrate the embodiments and, together with the description, serve to explain the principles of the embodiments, and with reference to these descriptions, one skilled in the art will recognize other possible implementations and advantages of the present invention, wherein elements are not drawn to scale, and like reference numerals are generally used to designate like elements.
According to the embodiment of the invention, an unmanned aerial vehicle inspection tracking distribution network line flight method based on edge calculation is provided.
The invention is further described with reference to the accompanying drawings and the specific embodiments, as shown in fig. 1-16, according to an embodiment of the invention, an unmanned aerial vehicle inspection tracking distribution network line flight method based on edge calculation comprises the following steps:
s1, acquiring power distribution network image data by using a terminal node and analyzing and calculating;
s2, adaptively distributing the computing task of the terminal node to the edge cloud by adopting a preset method;
s3, judging whether the terminal node loses a power distribution network target according to a preset scheme;
and S4, fusing calculation results of the terminal nodes and the edge cloud through a prefabrication method and analyzing network distribution conditions.
By means of the scheme, the computing tasks are reasonably unloaded to the edge cloud by utilizing the task segmentation module, and the computing results are analyzed and fused by utilizing the information fusion module, so that the computing pressure and the power consumption of the end nodes are reduced; in addition, by adding the motion detection module, the calculation pressure and the power consumption of the terminal node can be further reduced, and the operation time of the unmanned aerial vehicle can be prolonged.
In one embodiment, the edge cloud is an edge computing platform, and is provided with a target tracking module, an information fusion module, an information management module and a communication module.
In one embodiment, the terminal nodes are unmanned aerial vehicles accessing to an edge computing platform, and each terminal node is provided with a target tracking module, a computing task segmentation module, an information fusion module and a motion detection module.
Specifically, the overall design scheme framework of the unmanned aerial vehicle inspection tracking distribution network algorithm is divided into three layers, namely a terminal node, an edge cloud and a cloud data center. The terminal node layer is mainly an unmanned aerial vehicle accessed to the edge computing platform, and each node deployment application comprises a target tracking module, a computing task segmentation module, an information fusion module, a motion detection module and the like. The edge cloud layer is mainly an edge computing platform resource deployment layer and comprises hardware resource devices such as an edge computing server, a file server and the like, necessary programs required for completing computing tasks of terminal nodes are deployed, and the edge cloud layer mainly comprises a target tracking module, an information fusion module, an information management module, a communication module and the like. The cloud data center is a data center far away from the mobile terminal node, and the layer only communicates with the edge cloud.
In one embodiment, the step of adaptively distributing the computing task of the terminal node to the edge cloud in S2 by adopting a preset method includes judging the equipment temperature, the CPU utilization, the memory occupancy rate and the energy status of the unmanned aerial vehicle, and adaptively distributing the computing task to the edge cloud according to the real-time network status and the resource occupancy condition of the edge server.
Specifically, the unmanned aerial vehicle tracking distribution network algorithm framework design main body based on calculation unloading is a target tracking algorithm based on computer vision, and input data is video stream data, specifically video frame pictures. The task segmentation module judges whether to perform local computation according to a local computation load state, such as equipment temperature, CPU utilization rate, memory occupancy rate, energy state and the like, and meanwhile judges whether to transmit edge cloud to perform computation according to a network state and an edge server resource occupancy condition, and meanwhile, determines computation task segmentation points to finally determine task computation amounts of terminal nodes and the edge cloud. Because the network condition is difficult to ensure for a long time in a field environment and the specific application environment of the equipment is complex, when the mobile terminal node and the edge cloud are difficult to keep smooth connection, the target tracking module deployed by the terminal node has independent operation capability, otherwise, the tracking task is directly caused to fail. The terminal node adopts an algorithm based on a correlation filter to deploy so as to meet the characteristics of weaker computing capacity and higher energy consumption requirement of the mobile node.
In addition, in the task segmentation strategy, two segmentation strategies are adopted in the subsequent experiments, namely the complete unloading and the partial unloading of the neural network model.
(1) Model complete offload strategy: the feature extraction part of the algorithm is mainly unloaded to the edge cloud for calculation, and the local terminal node is responsible for preprocessing and post-processing of data. The calculation task of the data preprocessing part mainly comprises the calculation of the super parameters related to the input image data; and cutting out images with different sizes by taking the position of a tracking target of a previous frame as a center for subsequent scale estimation calculation and the like. The post-processing part mainly comprises the steps of calculating the scale and specific estimated position of the tracking target according to the response diagram returned by the edge cloud.
(2) Model part offload strategy: compared with completely offloading the feature extraction task to edge cloud computing, the first part of the convolutional neural network model, namely the first convolutional layer, the batch normalization layer and the pooling layer, is distributed to terminal node computing.
In addition, the task segmentation module divides the computing tasks mainly according to the states of local computing node load conditions, hardware equipment utilization rate, network delay and the like, and the computing tasks are partially or completely unloaded to the edge cloud, so that the combined optimization purpose of energy consumption, computing speed and accuracy is achieved. When the task is calculated and split, on one hand, the data volume is smaller after local processing is considered, so that the network transmission delay is reduced; on the other hand, when the network is in a high-delay state, the subsequent calculation tasks are continuously processed locally, so that the current high-delay period is avoided, and unnecessary waiting delay is reduced.
After the module is probed by the environment information, the obtained information is fused, the calculation task dividing points are judged, and the local terminal nodes and the edge cloud calculation task quantity are decided.
The computational task partitioning is mainly between the first few layers of data preprocessing or neural network models. And the task segmentation module judges that the local terminal node performs preprocessing and clipping operation, or continuously performs first-layer or several-layer feature extraction, and the processed data is uploaded to the edge cloud for further calculation. And after the response graph is obtained by the edge cloud computing, returning an extremum sequence in the response graph, and transmitting the data to a local information fusion module for reconstructing the response graph. And finally judging the target position after fusing the processing result of the terminal node target tracking algorithm. The edge cloud is only responsible for feature extraction and response diagram related calculation tasks, and the local terminal node is responsible for data preprocessing, shallow feature extraction, scale judgment after obtaining the response diagram and other calculation tasks.
In one embodiment, the computing any includes processing power distribution network image data, planning a flight path, and analyzing power distribution network equipment faults.
In one embodiment, the edge cloud further comprises an unmanned aerial vehicle with high computing power as an edge cloud node, and when the unmanned aerial vehicle cannot be normally connected with the edge cloud, the unmanned aerial vehicle moves the power distribution network image data to the edge cloud node.
Specifically, the cloud resource is moved down to the edge end or the edge cloud layer is continuously split, and in the platform design, the mobile terminal node with stronger computing capability is moved up to the edge cloud layer. In a field unmanned aerial vehicle inspection distribution network environment, a working environment may be poor, network delay may cause that terminal node equipment and edge cloud cannot be smoothly connected, and therefore, an acquired distribution network image is moved to node equipment with stronger computing capability to serve as an edge cloud node of the equipment. The unmanned aerial vehicle mobile terminal node can freely select a task unloading position, so that the influence of a network environment on a computing task is reduced.
In one embodiment, the step of determining whether the terminal node loses the power distribution network target according to the preset scheme in S3 further includes the following steps:
s31, modeling the motion condition of a power distribution network target by a terminal node through a Kalman filter;
s32, setting a displacement threshold of a power distribution network target by adopting a motion model;
s33, judging whether the target displacement of the power distribution network exceeds a threshold range, if so, executing S34, otherwise, executing S35;
s34, judging whether a moving object exists by utilizing a motion detection module of the terminal node, if so, executing S35, otherwise, stopping the calculation task;
s35, continuing to calculate the task.
Specifically, an algorithm fusion coefficient is determined according to a response diagram calculated by an algorithm, and whether a terminal algorithm model is updated at the moment is determined. And the information fusion module carries out final prediction on the specific position of the tracking target through fusing the calculation results returned by the terminal nodes and the edge cloud. In addition, the terminal node adopts a Kalman filter to model the motion condition of the tracking target, on the other hand, adopts a motion model to dynamically set the displacement threshold of the tracking target, and when the displacement of the continuous multi-frame target exceeds the threshold range, the motion detection module is started to determine whether a moving object exists or not, so that the calculation pressure of the terminal is further reduced, and unnecessary calculation is avoided.
In addition, as the unmanned aerial vehicle patrol tracking distribution network has higher requirements on calculation pressure and power consumption, and when a tracking algorithm loses a target, continuous calculation causes serious resource waste, however, the tracking algorithm is usually difficult to judge when the tracking algorithm correctly tracks the target or the target which is lost, so that a motion model is adopted to model the motion state of the tracking target, a displacement threshold range is dynamically set through the motion model, when the tracking algorithm predicts that a continuous multi-frame of positions exceeds the threshold range, the lost target is preliminarily judged, and a motion detection module is started to detect whether a moving object exists or not so as to determine whether the target is lost or not. The module may assist the algorithm in re-detecting when the target is lost.
Because the motion trail of the tracked target has strong positive significance on target tracking, the invention adopts a Kalman filter to model the motion state of the tracked target, and the method is summarized as follows.
Prediction stage:
(1) state prediction equation:
wherein,the symbol "-" indicates an estimated value, a is a state transition matrix, B is a control matrix, and u is a control amount, which is obtained by estimating the value from the previous state.
(2) Error prediction equation:
p is an error correlation matrix, and a covariance matrix Q represents inherent noise of the model.
Updating:
(3) kalman filter gain:
k is called Kalman coefficient, H is observation matrix, and R is measurement noise covariance matrix.
(4) State correction equation:
z is the true observed value of the value,representing realityResidual between the observed value and the model predicted value.
(5) Correcting an error correlation matrix:
during target tracking, the algorithm can acquire real information of a tracked object from the first frame data, and in the follow-up picture data, the position can be predicted only through the algorithm to acquire the information, so that in the tracking process, the early tracking effect is good, and the later stage gradually generates a drift problem until the tracking fails. Therefore, we use the first K frames to detect position and perform filter model training. In the later tracking process, a motion model prediction result is added into a filter response diagram so as to improve the algorithm tracking effect.
When information is fused, a two-dimensional Gaussian spectrum is generated by taking the predicted position of the motion model as the center, and the peak height is s times of the peak value of the response spectrum of the correlation filter.
In the related paper, when multiple peaks appear in the filter response chart, a common strategy is to select the coordinate value closest to the target position of the previous frame as the tracking result, such as an algorithm LDES, and the problem of the method is that when the target moves rapidly, a discrimination error occurs, namely, a far peak is a real target, and a near peak is a false measurement, so that the target tracking fails. In contrast, an independent motion model is adopted, and a filter result and a motion model result are combined at the same time, so that a correct result can be judged when both have higher response values.
The motion detection algorithm is deployed by adopting a frame difference method so as to meet the requirements of small calculation pressure, high calculation speed and low energy consumption. In addition, a motion model is built to adaptively set displacement thresholds for different tracked targets and for the targets at different stages of the tracking process.
(1) Motion model:
for the target tracking algorithm, the motion state of the object has positive influence on model tracking, especially when conditions such as shielding and motion blurring exist in the tracking process. Thus, tracking target motion states are modeled to further improve algorithm performance. In addition, due to dynamic changes of the target during tracking, it is not appropriate to set a fixed threshold for the target displacement to determine whether the filter prediction result is within a normal range. Based on the reasons, a motion model is adopted to dynamically set the target displacement threshold value so as to adapt to the dynamic change of the target in the tracking process.
The motion model setup is modeled using a kalman filter.
The threshold value is set as shown below,
(2) Motion detection algorithm:
when the displacement of the continuous multi-frame predicted target of the tracking algorithm exceeds the threshold range, the motion detection algorithm is adopted to perform motion detection on the input data so as to further reduce the energy consumption of the model, improve the processing speed of the algorithm and reduce unnecessary calculation. Common motion detection algorithms include a frame difference method, an optical flow method, a background subtraction method, a ViBe algorithm and the like. The frame difference method has the characteristics of simplicity, rapidness, insensitivity to illumination environment, strong adaptability to dynamic environment and the like, and is suitable for a hardware environment of a terminal platform and an application program running environment.
The main theoretical basis of the frame difference method is that after two or three adjacent frames are subjected to gray value difference operation, gray residues are generated by the moving object due to gray value change, and the static object is removed by the difference operation due to gray invariance. The three-frame difference method is improved based on the adjacent frame difference method, so that the problem that the boundary contour of an object in two-frame difference is thicker is solved, and the detection performance is better. The algorithm mainly comprises the following three steps: three frames of image data are continuously input, and after gray level difference values are calculated by the first two frames and the second two frames respectively, the result is subjected to bitwise AND operation to obtain the result.
When in motion detection, after a three-frame difference method result is obtained, the target tracking algorithm predicts the position, the target size region is taken, and whether the object motion exists in the region is determined by summing the values in the region and comparing the values with a threshold value.
In one embodiment, the terminal node adopts a correlation filter algorithm, and the edge computing platform adopts a tracking algorithm of a full convolution twin neural network deep learning model.
In one embodiment, the calculation formula of the full convolution twin neural network is:
wherein x represents a current picture data sample, z represents next picture data, and the functionFor feature extractor embedding, the function g is used for similarity measurement, b1 represents offset, and the algorithm is a cross-correlation operation.
In one embodiment, in the step S4, the fusing the calculation results of the terminal node and the edge cloud through a prefabrication method and analyzing the network distribution situation further includes the following steps:
s41, obtaining a response diagram according to a full convolution twin neural network algorithm and determining an algorithm fusion coefficient;
s42, integrating a correlation filter algorithm of the terminal node and a deep learning algorithm of the edge cloud to update an algorithm model;
s43, fusing the calculation results returned by the terminal nodes and the edge cloud by using an information fusion module, and finally predicting the specific position of the power distribution network target.
In particular, ensemble learning has become a strategy that is often employed in engineering applications to aggregate the advantages of different algorithms. For different algorithms deployed by the terminal nodes and the edge cloud in the tracking platform, the integration strategy or the information fusion strategy of the different algorithms still needs to be further researched. In actual deployment, the response patterns calculated by different algorithms are measured based on peak confidence indexes, so that specific fusion coefficients are determined. At the same time, it is determined whether the local correlation filter based tracking algorithm performs model updating.
The peak confidence definition is as follows.
Where Var () represents the variance, D θ D for extreme value data point set in response diagram θ Max Value represents the extreme point set after removing the maximum Value in the response diagram, and set D θ The medium filtering is smaller than the threshold value theta p To enhance robustness, super parameter θ p And presetting. Delta is a super parameter for avoiding numerical overflow in the operation to define a minimum value in advance. The filter is updated only when the peak confidence is above a threshold.
In addition, as the transmission time of the response graph has a great influence on the overall processing time of the algorithm, the algorithm only transmits an extremum set higher than a certain threshold in the response graph to reduce the information transmission quantity, and the terminal node rebuilds the response graph by returning extremum information.
Wherein G (x, y) is a reconstructed surface, G i (x, y) is a two-dimensional Gaussian surface centered on the ith extreme point coordinate, V i Is the corresponding extremum.
Because the exponential operation of the Gaussian function makes the calculation pressure larger, the strategy cannot meet the real-time requirement, the method is further optimized, and the translation strategy is adopted to replace the related operation for creating the two-dimensional Gaussian function.
Wherein G (x, y) is a reconstructed surface, G (x, y) is a two-dimensional Gaussian surface which is pre-created and takes (0, 0) coordinates as the center, V i To correspond to extreme value, P g (x i ,y i ) To translate the curved surface g (x, y) (x i ,y i ) And (5) a subsequent Gaussian curved surface.
In addition, the information fusion module is mainly used for information storage. The edge cloud stores the state of the target in the tracking process, records the motion path of the object and the activity of the target, and the historical record information is mainly used for decision judgment or subsequent further information extraction in a specific scene. For example, for the unmanned aerial vehicle inspection distribution network, the inspection path can be determined by analyzing the motion record, so that the next inspection operation is assisted to be judged. In addition, the record distribution network equipment and surrounding obstacle information can be combined with other application information to perform higher-layer information fusion or send other terminal nodes for other subsequent operations.
In addition, as the unmanned aerial vehicle inspection distribution network working environment is diversified, the network state is difficult to ensure, when equipment cannot be smoothly connected with the edge cloud, the terminal node still has independent operation capability, and serious consequences caused by task failure are avoided. For the above reasons, a target tracking module with smaller calculation pressure is deployed at the terminal node. On the other hand, since the algorithm is usually focused on solving the problem of one aspect of a specific task, it is difficult to optimize all the problems of the task, and therefore, a mode of integrating a plurality of algorithms is usually adopted to combine the advantages of different algorithms, so as to construct an algorithm model with better performance to meet the actual requirements. For example, the main difficulties of the target tracking algorithm based on computer vision include illumination, deformation, scale change, background clutter and the like, and different algorithms are usually designed and optimized for a certain problem, so that in practical application, an integrated mode is usually adopted to obtain better tracking performance. The ensemble learning is used for obtaining an algorithm model with better performance by combining a series of algorithm models, and the strategy is widely applied in engineering deployment. In the unmanned aerial vehicle inspection distribution network module, a relevant filter algorithm of the mobile terminal node and a deep learning algorithm of the edge cloud are integrated to obtain a more stable algorithm model with better performance.
(1) Terminal node target tracking module
The terminal node adopts a target tracking algorithm based on a correlation filter CF to deploy so as to meet the characteristics of weaker computing capacity and higher requirements on cruising ability of the mobile terminal node, and the algorithm model adopts a correlation filter DCF.
DCF is proposed on the basis of CSK algorithm. The DCF algorithm solves the sparse sampling problem of the prior related filter by using the cyclic matrix characteristic through cyclic sampling, and introduces the multi-channel characteristic, so that the performance is greatly improved compared with the prior target tracking algorithm.
DCF uses ridge regression modeling to solve a function by minimizing the distance between the computation tag of the sampled data and the true location of the target.
min w ∑ i (f(x i )-y i ) 2 +λ||w|| 2 →f(z)=w T z;
w=(X T X+λI) -1 X T y→w=(X H X+λI) -1 X H y;
The W expression can be obtained after solving, the right expression on the upper side is in a complex domain form, and X H Representing the complex conjugate matrix of X. After introducing a cyclic matrix to carry out cyclic sampling, a sampleThe method is carried into the method for solving the problems,
a kernel function is introduced and the solution is available,
α=min α ||φ(X)φ(X) T α-y|| 2 +λ∥φ(X) T α∥ 2 =(φ(X)φ(X) T +λI) -1 y=(K+λI)-1y;
form in the frequency domain
DCF employs a linear kernel
Detection stage
The DCF tracking algorithm mainly comprises the following steps: and calculating a kernel function k (x, x) and a parameter alpha value according to the current picture data sample x, calculating k (z, x) and f (z) when the next picture data z is input, and taking the real part of f (z) as a response graph, wherein the maximum value position is the algorithm prediction target position.
The DCF algorithm has the main advantage of processing speed, and is suitable for mobile equipment with weaker computing power. The algorithm is deployed at the terminal node, so that the unmanned aerial vehicle still has independent operation capability when the unmanned aerial vehicle cannot be connected with the edge cloud, and task failure is avoided.
(2) Edge cloud target tracking module
Edge computation refers to performing deep learning computation at the edge of the drone using Jetson Nano, rasberry Pi devices with embedded GPUs. The basic idea of edge computing is to migrate a cloud computing platform from inside a mobile core network to the edge of a mobile access network, thereby enabling flexible utilization of computing and storage resources. Edge computing pushes mobile computing, network control, and data storage towards the network edge, thereby implementing computationally intensive and latency critical applications on resource-constrained edge devices.
Edge computing is actually an extension of cloud computing compared to cloud computing modes, as edge computing can extend cloud computing paradigms to network edges to compensate for lack of security in data storage and high latency in service delivery in cloud computing. Edge computing has many advantages and features, including the important role of wireless access, good mobility and scalability, lower latency and location awareness, wider geographic distribution and real-time applications, higher security, etc. Edge computing may provide advantages for industrial, entertainment, personal computing, and other applications with computing and storage capabilities. Through the integrated data collection, calculation and storage services, the unmanned aerial vehicle-cloud system based on the mobile edge brings various convenience to individual clients and enterprises.
The network environment of edge calculation is utilized, so that information transmission and call among nodes can be realized, and the effect of parallel tracking of the nodes is realized. In the parallel tracking process, the target position is determined according to the response values of a plurality of position nodes, so that the selection of the target tracking position is more credible. The algorithm is embedded into the tracking process, and in terms of time complexity, a small amount of time is consumed for generating and releasing the nodes, and the main node and the generating node are processed in parallel, so that the real-time performance of tracking is not affected by the designed algorithm 1. In terms of space complexity, the space complexity increment of the algorithm is not large due to the reasonable node generation and release principle.
Based on the characteristics of abundant edge computing resources, high processing speed and the like, the algorithm deployment does not need to consider excessive hardware resource overhead. Deployment is carried out by adopting an unmanned aerial vehicle routing inspection distribution network algorithm based on a deep learning model so as to meet the performance requirement of a tracking task, and the algorithm model adopts a full convolution twin neural network Siam FC.
The siem FC compares an example image z with the candidate image x and returns a high score if both pictures describe the same object.Is about the template and candidate image full convolution layer for extracting the features of the template and search image x, due to the template and candidate image +>Is a shared parameter and is therefore also called a twin network. After the twin network extracts the features of the template and the candidate image respectively, the feature map of the template is finally generated to be 6 x 128, the feature map of the candidate image is 22 x 128, then the feature map of the template is used as a convolution kernel to carry out convolution operation with the feature map of the candidate image, a fraction mapping is output, the dimension of the fraction mapping is 17 x 1, and finally, the maximum value of the fraction mapping response is the position of the target. In the siem FC, the template is scaled to 127 x 3 after clipping, and only the image of the first frame is used in tracking. In the training stage, if only the target label is set to be 1 and the labels of the rest image parts are set to be 0, the sample imbalance problem can be caused, in order to solve the sample imbalance problem, the Siam FC marks the image part in the radius r as 1 by taking the target center point as the center of a circle when marking the sample, and the rest image part is set to be 0, so that the sample imbalance problem is solved to a certain extent. The convolution layer uses a five-layer AlexNet network, and no padding technique is used in the convolution in order to maintain the translational invariance.
Because the target tracking task can only acquire tracking target information from initial picture data, the tracking target can be any object, the form shown in the current data can be any posture of the target, even only a part of the tracking target, the object to be tracked has strong uncertainty, so that model training can not be carried out through pre-prepared target data, and the complete and comprehensive modeling of the target can not be carried out through an algorithm only by the information. For the above reasons, target tracking algorithms typically employ online update strategies to accommodate tracking.
Real-time changes in the target throughout the tracking process. The real-time requirement of the target tracking task puts a severe requirement on an online updating or online learning strategy of the model, and the updating strategy must be simple and effective to ensure the real-time performance of the algorithm, which is also a main reason that the target tracking algorithm based on the correlation filter is widely applied.
The full convolution twin neural network innovatively uses target tracking as a similarity learning problem, and after a model is trained offline by adopting a large amount of data, the model does not need to be updated when the model is actually deployed for tracking, so that complex calculation in the tracking process is avoided, and the algorithm performance meets the requirements of real-time tracking tasks.
The similarity between the sample picture and the candidate picture is calculated through the similarity measurement function, the similarity degree between the sample picture and the candidate picture can be measured, and the score is returned to be scored, wherein the score is higher as the similarity is higher. The algorithm adopts a full convolution twin neural network as a similarity measurement function, and the specific form is as follows.
Wherein x represents a current picture data sample, z represents next picture data, and the functionThe function g is used for similarity measurement, b1 represents the offset, and the algorithm is specifically a cross-correlation operation.
The model is trained using the following formula,
argmin θ ∑ (x,y,z) L(y,f(z,x;θ));
where l (y, v) represents the loss of each point, θ is the regularization parameter, y e (-1, +1) is the actual true value, v is the actual score, D is the final score map, and u is the position in the score map.
The tracking algorithm performance based on the full convolution twin neural network has good performance in terms of processing speed and tracking performance, the algorithm solves the tracking problem by adopting the angle of similarity learning, and the algorithm can adopt a large number of video sequences to train the model offline, so that the target tracking task under a specific scene can adopt the method of collecting similar scene data in advance to train the model better, and further, the better performance is achieved.
In the preprocessing process of the algorithm, regions with different dimensions are cut at the position of the tracking target to perform operation so as to estimate the dimensional change condition of the target in the tracking process. When the platform is actually deployed, the size and the number of the cutting areas are optimized, so that the data size is smaller than the original data size after the original picture is preprocessed.
In summary, by means of the above technical solution of the present invention, the task segmentation module is utilized to reasonably offload the computing task to the edge cloud, and the information fusion module is utilized to analyze and fuse the computing result, so as to reduce the computing pressure and power consumption of the end node; in addition, the calculation pressure and the power consumption of the terminal node can be further reduced by adding the motion detection module, and the operation time of the unmanned aerial vehicle is prolonged, so that the efficiency of the unmanned aerial vehicle in inspecting the distribution network cable is improved, and the safety of the distribution network cable in conveying electric energy is ensured.
The foregoing description of the preferred embodiments of the invention is not intended to be limiting, but rather is intended to cover all modifications, equivalents, alternatives, and improvements that fall within the spirit and scope of the invention.