A method and system for video analysis oriented to unmanned aerial vehicle assisted edge computing network

Abstract

The application belongs to the technical field of video analysis, in order to solve the limitations of the existing unmanned aerial vehicle assisted edge computing network video analysis, and the problem of low efficiency, a kind of unmanned aerial vehicle assisted edge computing network video analysis method and system is provided, for each unmanned aerial vehicle independent decision-making distributed framework;Unmanned aerial vehicle computing in local computing scene task queue length, task delay and transmission reliability, and unloading to remote server in remote computing scene transmission time, task queue length, task delay and transmission reliability;Based on the calculation result, the optimization target function of inference accuracy rate is constructed with the optimization target of task successful execution and inference accuracy rate maximization, the optimization problem is converted into a partially observable Markov decision process, and the optimal decision strategy of learning adaptation dynamic environment is solved through reinforcement learning, which improves the video analysis efficiency in unmanned aerial vehicle assisted network.

Description

Technical Field

[0001] This invention belongs to the field of video analysis technology, and in particular relates to a video analysis method and system for UAV-assisted edge computing networks. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Video analytics is increasingly important in fields such as video surveillance, smart cities, and intelligent manufacturing. Given that efficient and accurate target detection requires substantial computing resources, performing all computations on a single device is impractical; therefore, video analytics supported by mobile edge computing (MEC) has been widely adopted. However, information gathering is often limited by the geographical location of cameras, and edge devices deployed in remote or isolated areas perform poorly, often struggling to maintain continuous video analytics operations in extreme environments lacking infrastructure (e.g., natural disasters). In contrast, unmanned aerial vehicles (UAVs) offer unique advantages in this regard: their mobility allows them to be deployed flexibly regardless of terrain, providing reliable real-time computing support for a variety of tasks. UAVs equipped with high-resolution cameras represent a promising solution for mobile edge video analytics in complex situations such as smart agriculture, natural disaster response, and emergency management.

[0004] Several factors need to be considered when conducting video analytics in UAV-assisted networks. First, UAVs are energy-constrained aircraft with limited battery capacity. A large portion of a UAV's energy is used for hovering and flight operations, leaving very little energy available for video analytics. Therefore, energy consumption constraints are crucial in UAV-assisted networks. Furthermore, UAVs have significantly limited computing resources compared to ground-based edge servers. This inherent limitation presents considerable challenges to effective decision-making regarding task offloading within UAV-related systems. Second, completing tasks in complex situations requires multiple UAVs, necessitating optimization of the overall system performance, not just the performance of individual UAVs.

[0005] Current research focuses on leveraging drones to support video analytics, but many challenges remain. The operating environment of drones is extremely complex; the computation and communication of their networked operating units are unstable, potentially leading to malfunctions during task processing. Furthermore, for efficient collaborative video analytics, drones need to make decisions in three key areas: selecting appropriate DNN models, configuring video frame preprocessing parameters, and making optimal offloading decisions. These decisions are inherently interconnected, requiring a system-level optimization strategy.

[0006] In summary, providing a video analysis method based on UAV edge computing to improve the efficiency and reliability of video analysis in UAV-assisted networks is a problem that needs to be solved. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, this invention provides a video analysis method and system for UAV-assisted edge computing networks, enabling each UAV to make independent offloading decisions, thereby improving the efficiency and reliability of video analysis in UAV-assisted networks.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] In a first aspect, the present invention provides a video analysis method for UAV-assisted edge computing networks, comprising:

[0010] After the drone acquires the video analysis task, the task queue length, task latency and transmission reliability are calculated in the local computing scenario, and the transmission time, task queue length, task latency and transmission reliability are calculated when the drone is offloaded to the remote server in the remote computing scenario.

[0011] Based on the task calculation results, an inference accuracy objective function is constructed with the goal of maximizing the inference accuracy while successfully executing the task. The delay threshold constraint is used as the constraint condition for the inference accuracy objective function, and the problem of optimizing the inference accuracy objective function is transformed into a partially observable Markov decision process.

[0012] A reinforcement learning model is used to solve the partially observable Markov decision process to obtain the optimal task processing decision strategy adapted to the dynamic environment. The UAV performs optimal processing on the video analysis task according to the optimal task processing decision strategy.

[0013] Secondly, the present invention provides a video analytics system for unmanned aerial vehicle (UAV) assisted edge computing networks, comprising:

[0014] The computing module is configured to: after the UAV acquires the video analysis task, calculate the task queue length, task latency and transmission reliability in the local computing scenario, and the transmission time, task queue length, task latency and transmission reliability when offloaded to the remote server in the remote computing scenario.

[0015] The transformation module is configured to: construct an inference accuracy objective function based on the task calculation results, with the goal of maximizing the inference accuracy while successfully executing the task, and use the delay threshold constraint as the constraint condition of the inference accuracy objective function, and transform the problem of optimizing the inference accuracy objective function into a partially observable Markov decision process;

[0016] The solution module is configured to: use a reinforcement learning model to solve the partially observable Markov decision process to obtain the optimal task processing decision strategy adapted to the dynamic environment, and the UAV performs optimal processing on the video analysis task according to the optimal task processing decision strategy.

[0017] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0018] Fourthly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in the first aspect.

[0019] The above one or more technical solutions have the following beneficial effects:

[0020] In this invention, a distributed framework for independent decision-making by each UAV is adopted, which does not rely on complete information from the global system. This adapts to the scenario requirements of UAVs autonomously executing tasks and avoids the drawbacks of centralized control in multi-UAV collaboration, such as slow response and reliance on communication links. The UAV calculates the task queue length, task latency, and transmission reliability in the local computing scenario, and the transmission time, task queue length, task latency, and transmission reliability when offloaded to the remote server in the remote computing scenario. These quantitative indicators provide objective basis for decision-making. Based on the calculation results, an inference accuracy objective function is constructed with the optimization goal of maximizing task success and inference accuracy. The optimization problem is transformed into a partially observable Markov decision process, which fits the actual observation scenario. The optimal decision strategy adapted to the dynamic environment is learned through reinforcement learning, thereby maximizing long-term system performance.

[0021] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0023] Figure 1 This is a flowchart of a video analysis method for UAV-assisted edge computing networks in Embodiment 1 of the present invention. Detailed Implementation

[0024] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0026] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0027] Example 1

[0028] This embodiment discloses a video analysis method for UAV-assisted edge computing networks, including:

[0029] After the drone acquires the video analysis task, the task queue length, task latency and transmission reliability are calculated in the local computing scenario, and the transmission time, task queue length, task latency and transmission reliability are calculated when the drone is offloaded to the remote server in the remote computing scenario.

[0030] Based on the task calculation results, an inference accuracy objective function is constructed with the goal of maximizing the inference accuracy while successfully executing the task. The delay threshold constraint is used as the constraint condition for the inference accuracy objective function, and the problem of optimizing the inference accuracy objective function is transformed into a partially observable Markov decision process.

[0031] A reinforcement learning model is used to solve the partially observable Markov decision process, and the optimal task processing decision strategy adapted to the dynamic environment is learned. The UAV performs optimal processing on the video analysis task according to the optimal task processing decision strategy.

[0032] This embodiment presents a distributed framework for independent decision-making by each UAV, eliminating the need to rely on complete information from the global system. It adapts to scenarios requiring autonomous UAV task execution, avoiding the drawbacks of centralized control in multi-UAV collaboration, such as slow response and reliance on communication links. The UAV calculates task queue length, task latency, and transmission reliability in local computing scenarios, and offloads these metrics to a remote server in remote computing scenarios, providing objective data for decision-making. Based on the calculation results, an inference accuracy objective function is constructed with the goal of maximizing task success and inference accuracy. This transforms the optimization problem into a partially observable Markov decision process, closely aligning with actual observation scenarios. Reinforcement learning is used to learn the optimal decision-making strategy adapted to the dynamic environment, maximizing long-term system performance.

[0033] This embodiment relates to a drone-assisted MEC system for video frame processing and analysis, comprising N drones (UAVs) and an edge cloud server (EC). Each UAV includes a controller, a frame preprocessing unit, a local inference queue, and several deep neural network models for video analysis. The edge cloud server (EC) includes a remote inference queue and deep neural network models. Considering that the UAVs perform target reconnaissance and generate analysis requests, these requests can be processed locally or offloaded to the edge cloud server (EC). The discrete-time system is modeled as t = {0, 1, 2, ...}, where each time slot has a very short duration. Typically, it only accommodates one or zero requests. Each drone The queue responsible for managing task execution follows a first-in, first-out (FIFO) principle. If a task in the queue has its computing resources occupied, it must wait for its turn to execute. The computing power of drone n is represented by F. n The computing power of the edge cloud server (EC) is represented as F0. When a video analytics task arrives, the controller manages the processing, including setting video frame preprocessing parameters, determining the offloading operation, and selecting the DNN model.

[0034] The decision-making process involves the following stages:

[0035] Video preprocessing: In each task, video frames are first preprocessed, including operations such as resizing, cropping, and quality reduction. For inference requests, the video resolution can be reduced to a resolution v in a set V, including options such as 360P, 480P, and 720P. The average latency of reducing a video frame to resolution v is D. v The corresponding data size is represented as B. v .

[0036] Offloading Decision: This is a crucial decision, determining whether inference should be performed on the drone or remotely at an edge computing center. If the task can be processed locally, it will be added to the local inference queue, and the reliability of the computation will be analyzed on the drone. Otherwise, the task will be added to the remote inference queue and sent to the edge computing center for processing by the video analytics model.

[0037] Model Selection: This stage involves selecting the most suitable deep neural network model. The set of deployed deep neural network models is denoted as M. The computational cost of inference is affected by the complexity of the selected deep neural network model and the video frame resolution, where the complexity of the deep neural network model is represented by m, and the video frame resolution is represented by v. The computational cost of one video frame is defined as C. m,v The corresponding raw accuracy is defined as A. m,v .

[0038] The optimization objectives mainly include three aspects:

[0039] Delay: If the task cannot be completed within a certain delay threshold, it means that the video cannot be analyzed in a timely manner, which may result in invalid data.

[0040] Accuracy: Accuracy is a measure of whether an object is correctly classified and positioned within a frame.

[0041] Reliability: The environment in which drones operate is extremely complex. Therefore, the transmission reliability of drones is also an important metric.

[0042] First, this embodiment uses a Poisson probability model to construct a reliability model for the UAV communication link. Assume the mission transmission time of UAV n is... During task transmission (0, Within this framework, system faults are modeled as having strength parameters. The Poisson event, the cumulative number of failures of drone n is represented as Therefore, in (0, Exactly zero faults were observed within the timeframe. The probability is denoted as :

[0043]

[0044] Next, we will discuss the local computation model. For locally processed requests, the length of the inference task queue for drone n in time slot t is denoted as... :

[0045]

[0046] in, Let be the complexity of the deep neural network model for UAV n in time slot t. Let n be the resolution of the drone video in time slot t. The model complexity is The task resolution is The computational cost of the next video frame, Let t represent the computing power of drone n, and t represent the time slot. This means that if the value inside the symbol is greater than or equal to 0, the original value will be output directly; if the value inside the symbol is less than 0, 0 will be output.

[0047] Task latency includes task queue inference time and video preprocessing time, i.e.:

[0048]

[0049] in, This indicates a task delay; To reduce video frames to the resolution The average delay; Let n be the computing power of the drone.

[0050] The following is the remote computing model. If the task is unloaded to a remote cloud server, the required transmission time is denoted as... :

[0051]

[0052] in, It refers to the amount of data corresponding to the task. This refers to the transmission rate.

[0053] Due to the complex environment, the transmission process of drones is unreliable; the reliability of remote computing transmission is denoted as... :

[0054]

[0055] in, Let n be the intensity parameters of the n Poisson event of the UAV.

[0056] Since the EC receives all the unloading tasks from the drones, the length of the EC's inference task queue at time slot t is denoted as: :

[0057]

[0058] in, Indicates local computation. This indicates remote computing, where N is the total number of drones, and the value of n ranges from [1, N]. It is EC's computing power. Indicates the duration of each time slot; The model complexity is The task resolution is The computational cost of the next video frame.

[0059] Therefore, the total latency of the remote computing task is denoted as... :

[0060]

[0061] in, It is to reduce the video frames to the resolution. average latency, It is the computing power of EC.

[0062] Combining local and remote computing models, if a task executes successfully, the task latency is recorded as... :

[0063]

[0064] One of the optimization variables is where the task is executed. If =1, the task is executed remotely, and the task latency is reduced. = ;like If =0, the task is executed locally, and the task latency is reduced. = .

[0065] To avoid invalid video analytics data, task latency should not exceed the allowable latency threshold, which can be expressed as:

[0066]

[0067] in The allowed delay threshold.

[0068] The optimization objective is to maximize inference accuracy while ensuring successful task execution, and is defined as follows:

[0069]

[0070] Among them, if the condition If true, then , if the condition If it is false, then , The original accuracy rate is used as the target value for easier differentiation. Known as inference accuracy, For the reliability of task execution; if =1 means the task is executed remotely. If the value is 0, the task is executed locally.

[0071] To address the aforementioned optimization problem, this embodiment proposes to use deep reinforcement learning to solve it. Therefore, it is first transformed into a partially observable Markov decision process, which is mainly described from four aspects: system state, action space, reward function, and observation space.

[0072] System status: The current workload of drones and edge cloud servers is a key factor influencing decision-making. The current workload is described by the task queue length, defined as follows:

[0073]

[0074] in, This means that if the value inside the symbol is greater than or equal to 0, the original value is output directly; if the value inside the symbol is less than 0, 0 is output. This indicates the current workload of the edge cloud server. This represents the current workload of drone n. It is the queue length before the task arrives at time t. It is the queue length after the task arrives at time t.

[0075] Use a vector To represent all drone mission requests:

[0076]

[0077] in, This represents the number of task requests made by drone n.

[0078] Therefore, the system state The definition is as follows:

[0079]

[0080] in, Represents the current load vector. Represents the computational capability vector. Represents the failure rate parameter vector. Represents the transmission rate vector, { } is the parameter vector corresponding to video preprocessing and model selection.

[0081] Action space: For each drone, the unloading decision, frame resolution, and model selection for time slot t need to be determined.

[0082] Action decision-making for a single drone It can be represented as:

[0083]

[0084] System Actions It is the collection of all drone actions, defined as:

[0085]

[0086] in, For the actions of drone 1, For the actions of drone N.

[0087] Reward function: The goal is to maximize inference accuracy. Therefore, if an action leads to higher inference accuracy and satisfies all constraints, it will receive a higher reward. Conversely, if constraints are violated, the reward function will include a penalty.

[0088] Reward function The definition is as follows:

[0089]

[0090] in, -Pu The penalty for not satisfying all constraints.

[0091] Observation space: Each drone can only acquire partial information, not all data from the entire system. Observation information This includes local information about the drone and status information from the edge cloud server, defined as follows:

[0092]

[0093] in, This indicates the current workload of the edge cloud server (EC). Used to describe the current workload of drone n; The number of task requests for drone n; This indicates the computing power of the EC; The computing power of drone n; The intensity parameter represents the n Poisson event of a drone. This represents the transmission rate of drone n; It is the parameter vector corresponding to video preprocessing and model selection; t represents the time slot.

[0094] Based on the partially observable Markov decision process constructed above, this embodiment proposes a video analysis algorithm based on reinforcement learning. As shown in Algorithm 1, specifically, firstly, all evaluation networks are initialized. All execution networks and hybrid networks ,in To evaluate the network parameters, To execute the network parameters, These are the parameters for the hybrid network. In each iteration, each drone is first... n It will transmit its own observation information The input values ​​are fed into the execution network, which then executes the network and outputs the corresponding actions. Based on the current system status and the actions of all drones You can receive the system reward at this time. After the action is performed, the system will enter the next state. , the quintuple ( , , , , Stored in the relay buffer.

[0095] During the training of a reinforcement learning model, samples are randomly drawn from the replay buffer to execute the network's actions. and system status The input values ​​are fed into the evaluation network, which outputs a score for the current action. The outputs of all evaluation networks are then fed into the hybrid network to output an overall evaluation of all actions. Gradient descent is then performed using the difference between the hybrid network's output at the next time step and its expected output value. The expected output value... for:

[0096] , (t), );

[0097] in, , (t), )= ( , ) represents the current output value of the hybrid network. To update the step size, the loss function is defined as follows:

[0098] Loss=

[0099] in, This represents the output value of the hybrid network at time t+1.

[0100] Gradient descent is performed based on this loss function to update the parameters of the hybrid network. And evaluate the parameters of the network Based on the output value of the evaluation network Perform gradient descent to update network parameters By repeating the above steps multiple times, a trained reinforcement learning model can be obtained.

[0101] Algorithm 1:

[0102] 1. Initialize all evaluation networks All execution networks and hybrid networks ;

[0103] 2. Initialize the experience pool ;

[0104] 3. Forepisode=1 to Kdo

[0105] 4. Obtain the initial global state s0 and all local information. ;

[0106] 5. Fort=1 to max-lengthdo

[0107] 6. Parameterized Actions in the Decision-Making Department ;

[0108] 7. Perform all actions Receive global rewards Update global state and local observation status ;

[0109] 8. The experience of the sample ( , , , , Stored in experience pool D;

[0110] 9. IF Training then

[0111] 10. Randomly sample from experience pool D;

[0112] 11. , (t), );

[0113] 12. Calculate the loss function Loss = Based on this loss function, gradient descent is performed to update the parameters of the hybrid network. And evaluate the parameters of the network ;

[0114] 13. Based on the evaluation network output results Perform gradient descent to update network parameters ;

[0115] 14. End

[0116] 15. End

[0117] 16. End

[0118] In Algorithm 1, episode is used as the loop iteration variable, max-length represents the upper limit of the loop, and Training represents training.

[0119] Example 2

[0120] The purpose of this embodiment is to provide a video analytics system for UAV-assisted edge computing networks, including:

[0121] The computing module is configured to: after the UAV acquires the video analysis task, calculate the task queue length, task latency and transmission reliability in the local computing scenario, and the transmission time, task queue length, task latency and transmission reliability when offloaded to the remote server in the remote computing scenario.

[0122] The transformation module is configured to: construct an inference accuracy objective function based on the task calculation results, with the goal of maximizing the inference accuracy while successfully executing the task, and use the delay threshold constraint as the constraint condition of the inference accuracy objective function, and transform the problem of optimizing the inference accuracy objective function into a partially observable Markov decision process;

[0123] The solution module is configured to: use reinforcement learning to solve the partially observable Markov decision process, learn the optimal task processing decision strategy adapted to the dynamic environment, and the UAV performs optimal processing on the video analysis task according to the optimal task processing decision strategy.

[0124] In further embodiments, the following is also provided:

[0125] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When executed by the processor, the computer instructions perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0126] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0127] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0128] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0129] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0130] A computer program product includes a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0131] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0132] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0133] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0134] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0135] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A video analytics method for UAV-assisted edge computing networks, characterized in that, include: After the drone acquires the video analysis task, the task queue length, task latency and transmission reliability are calculated in the local computing scenario, and the transmission time, task queue length, task latency and transmission reliability are calculated when the drone is offloaded to the remote server in the remote computing scenario. Based on the task calculation results, an inference accuracy objective function is constructed with the goal of maximizing task success and inference accuracy. A delay threshold constraint is used as the constraint condition for this objective function. The optimization problem of the inference accuracy objective function is then transformed into a partially observable Markov decision process, specifically: The system state is constructed based on the current task queue length vector of the UAV, the task request vector of all UAVs, the UAV computing power vector, the UAV failure rate parameter vector, the UAV transmission rate vector, and the parameter vectors corresponding to UAV video preprocessing and model selection. The drone unloading decision, frame resolution, and model selection are defined as the action space; The reward function is set based on whether the action performed maximizes the objective function of inference accuracy. The local information of the drone and the state information of the edge cloud server are defined as the observation space; A reinforcement learning model is used to solve the partially observable Markov decision process to obtain the optimal task processing decision strategy adapted to the dynamic environment. The UAV performs optimal processing on the video analysis task according to the optimal task processing decision strategy.

2. The video analysis method for UAV-assisted edge computing networks as described in claim 1, characterized in that, The objective function for inference accuracy is: in, For the accuracy of reasoning, if the condition If true, then , if the condition If it is false, then , For the reliability of task execution, The allowable delay threshold, For task delay, This represents the original accuracy rate.

3. The video analysis method for UAV-assisted edge computing networks as described in claim 1, characterized in that, A reliability model for UAV communication links is constructed using the Poisson probability model, and the transmission reliability is calculated in both local and remote computing scenarios based on the Poisson probability model.

4. The video analysis method for UAV-assisted edge computing networks as described in claim 1, characterized in that, The training process for a reinforcement learning model is as follows: Each drone's current observation information is input into the execution network to obtain corresponding actions. The system reward is obtained by combining the current system status and the actions of all drones. After the UAV performs the corresponding action, the next system state is obtained, and a five-tuple consisting of the next system state, system reward, current system state, the UAV's current observation information and corresponding action is stored in the buffer. Samples are randomly drawn from the buffer to score the current action. The outputs of all evaluation networks are used as inputs to the hybrid network to obtain an overall evaluation of all actions. The parameters of the hybrid network and the evaluation network are updated using the gradient descent algorithm based on the difference between the output value of the hybrid network at the next time step and the expected output value. The parameters of the execution network are updated based on the output value of the evaluation network.

5. The video analysis method for UAV-assisted edge computing networks as described in claim 1, characterized in that, The drone performs optimal processing on the video analysis task according to the aforementioned optimal task processing decision strategy, specifically as follows: The drone adds the task to its local inference queue according to the optimal task processing decision strategy. Alternatively, the drone may add the task to the remote inference queue of the remote server according to the optimal task processing decision strategy.

6. A video analytics system for UAV-assisted edge computing networks, characterized in that, include: The computing module is configured to: after the UAV acquires the video analysis task, calculate the task queue length, task latency and transmission reliability in the local computing scenario, and the transmission time, task queue length, task latency and transmission reliability when offloaded to the remote server in the remote computing scenario. The transformation module is configured to: construct an inference accuracy objective function based on the task calculation results, with the optimization objective being the successful execution of the task and maximizing the inference accuracy; use a delay threshold constraint as the constraint condition for the inference accuracy objective function; and transform the optimization problem of the inference accuracy objective function into a partially observable Markov decision process, specifically: The system state is constructed based on the current task queue length vector of the UAV, the task request vector of all UAVs, the UAV computing power vector, the UAV failure rate parameter vector, the UAV transmission rate vector, and the parameter vectors corresponding to UAV video preprocessing and model selection. The drone unloading decision, frame resolution, and model selection are defined as the action space; The reward function is set based on whether the action performed maximizes the objective function of inference accuracy. The local information of the drone and the state information of the edge cloud server are defined as the observation space; The solution module is configured to: use a reinforcement learning model to solve the partially observable Markov decision process to obtain the optimal task processing decision strategy adapted to the dynamic environment, and the UAV performs optimal processing on the video analysis task according to the optimal task processing decision strategy.

7. A video analytics system for UAV-assisted edge computing networks as described in claim 6, characterized in that, In the transformation module, the problem of optimizing the inference accuracy objective function is transformed into a partially observable Markov decision process, specifically: The system state is constructed based on the current task queue length vector of the UAV, the task request vector of all UAVs, the UAV computing power vector, the UAV failure rate parameter vector, the UAV transmission rate vector, and the parameter vectors corresponding to UAV video preprocessing and model selection. The drone unloading decision, frame resolution, and model selection are defined as the action space; The reward function is set based on whether the action performed maximizes the objective function of inference accuracy. The local information of the drone and the state information of the edge cloud server are defined as the observation space.

8. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-5.

9. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-5.

Images