Delay optimization method and system for cooperative video analysis of unmanned aerial vehicle group

Abstract

The invention discloses a delay optimization method and system for cooperative video analysis of an unmanned aerial vehicle group, and the method comprises the following steps: constructing an unmanned aerial vehicle cooperative video analysis system comprising a plurality of unmanned aerial vehicles, collecting real-time videos from the ground by each unmanned aerial vehicle, and generating a reasoning task; generating a directed acyclic graph (DAG) for each unmanned aerial vehicle according to the reasoning task, initializing a segmentation point of the DAG of each unmanned aerial vehicle, and performing task unloading at the segmentation point by each unmanned aerial vehicle; and establishing an optimization objective function by taking minimization of task processing delay as an objective, solving the optimization objective function, and updating an unloading decision, communication resource allocation and a DAG segmentation point of each unmanned aerial vehicle according to a solving result. According to the method, the problem of delay minimization of video analysis in the centerless unmanned aerial vehicle group is solved, and task scheduling, communication resource allocation and DAG segmentation point selection can be optimized at the same time.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) collaborative video analysis, and in particular to a delay optimization method and system for unmanned aerial vehicle (UAV) swarm collaborative video analysis. Background Art

[0002] Compared with traditional monitoring systems, aerial monitoring systems can reduce manual involvement, effectively improve system efficiency and reduce operating costs. However, due to resource constraints, it is difficult for a single drone to provide long-term and reliable monitoring services. Therefore, the use of a collaborative network consisting of multiple drones is essential to achieve effective monitoring coverage. Compared with a single high-performance drone, the multi-drone network UAV-Net is gradually becoming a better choice in the field of aerial video surveillance because of its cost-effectiveness, high reliability, wide coverage, and enhanced robustness through information sharing and data fusion.

[0003] Although drone technology has made significant progress in the field of video surveillance, they are still limited by resource constraints when performing tasks, especially in high-performance collaborative computing and real-time network communication. To deal with these limitations, existing technologies mainly adopt two strategies: one is to offload the computing tasks of drones to ground base stations to reduce the computing burden of drones; the other is to deploy deep neural networks on drones to directly process the data captured by drones.

[0004] Edge computing offloading provides a solution to the resource limitations of drones, but the reliability of this solution may be affected due to the dynamic characteristics of wireless channels and the energy consumption limitations of drones when processing delay-sensitive images. At the same time, although the application of deep neural networks on drones to analyze captured data has brought significant improvements to system performance, in multi-drone systems, how to effectively partition deep neural network models and select key partition points are still challenges that need to be overcome in technology development. Summary of the invention

[0005] The present invention proposes a delay optimization method and system for collaborative video analysis of drone swarms, which respectively optimizes the task scheduling, communication resource allocation and directed acyclic graph (DAG) segmentation points of the drone network through a deep reinforcement learning algorithm and a heuristic algorithm, so as to improve the computing efficiency of drones and reduce energy consumption.

[0006] To achieve the above object, the present invention provides a delay optimization method for drone swarm collaborative video analysis, comprising the following steps:

[0007] Step S1: constructing a UAV cooperative video analysis system including multiple UAVs, each of which collects real-time video from the ground and generates an inference task;

[0008] Step S2: generating a DAG for each drone according to the inference task, initializing the split point of the DAG of each drone, and each drone performs task offloading at the split point;

[0009] Step S3: establishing an optimization objective function with the goal of minimizing task processing delay, solving the optimization objective function, and updating the unloading decision, communication resource allocation and DAG split point of each UAV according to the solution result;

[0010] Step S4: Repeat step S3 until the set maximum number of iterations is reached, and obtain the drone swarm task allocation plan with the minimum delay in the video analysis task.

[0011] Preferably, the task processing delay in step S2 includes communication delay and calculation delay, and the expression of the communication delay is:

[0012]

[0013] L v,v' =Θ v,v' +Γ LoS ;

[0014]

[0015]

[0016] In the above formula, is the communication delay between UAVs; S v,i (par v ) is the DAG partition point par of drone v v The output data size of R v→v′ B is the transmission rate from UAV v to UAV v'; v,v′ is the available bandwidth between UAV v and UAV v'; P v The power required to transmit the UAV v; G v,v′ is the channel gain between UAV v and UAV v'; σ 2 is the Gaussian white noise power; L v,v' is the path loss between UAV v and UAV v'; Θ v,v' = LoS is the additional attenuation factor of the line-of-sight link between UAV v and UAV v'; d v,v' is the distance between UAV v and UAV v'; f c is the carrier frequency; c is the speed of light; x v ,y v are the position coordinates of the drone v; x v' ,y v' is the position coordinate of the UAV v';

[0017] The expression of the task processing delay is:

[0018]

[0019] Where d is the task processing delay of the drone swarm performing the video analysis task; is a drone swarm; i∈M v , M v Video analysis tasks generated for drone v; is the delay of executing task i on UAV v; is the latency of task i generated on UAV v and transmitted to UAV v’ for execution; Indicates whether the task i generated by drone v is offloaded to drone v', Indicates uninstallation. Indicates that it has not been uninstalled; The latency of executing task i generated by drone v on drone v'.

[0020] Preferably, the energy consumption of the drone needs to be considered when solving the optimization objective function in step S3. Including the computing energy consumption of executing tasks Transmission energy consumption of offloading tasks to other drones And the hovering energy consumption E of the drone hovering at a fixed height v,hov ;

[0021] The calculation energy consumption E v,exe The expression is:

[0022]

[0023] In the formula, The computational energy consumption of UAV v performing task i; The computational energy consumption of executing task i generated by drone v’ on drone v;

[0024] The transmission energy consumption The expression is:

[0025]

[0026] The hovering energy consumption E v,hov The expression is:

[0027]

[0028] Where P v,hov is the hovering power consumption of UAV v; t v,hovis the maximum hovering time of the UAV v; η is a value proportional to the mass of the UAV; is the power efficiency of UAV v; q is the number of rotors of UAV v; r is the diameter of each rotor; κ is the air density;

[0029] The energy consumption of the drone v The expression is:

[0030]

[0031] Preferably, solving the optimization objective function in step S3 includes the following steps: using a distributed partially observable Markov decision process Dec-POMDP to define the state space, observation space, action space and reward function; based on the defined state space, observation space, action space and reward function, using a multi-agent proximal policy optimization algorithm MAPPO to solve the optimization objective function.

[0032] Preferably, the expression of the state space is:

[0033] S(t)={C(t),P(t)};

[0034] C(t)={(x 1,t ,y 1,t ),...(x v,t ,y v,t ),...,(x V,t ,y V,t )};

[0035] P(t)={p1(t),...p v (t),...,p V (t)};

[0036] Among them, S(t) is the state space; C(t) is the coordinate of each drone in the drone group at time slot t, (x v,t ,y v,t ) is the coordinate of UAV v at time slot t; P(t) is the DAG segmentation point of each UAV in the UAV swarm at time slot t; p v (t) is the DAG segmentation point of UAV v in time slot t;

[0037] The expression of the observation space is:

[0038] O(t)={C v (t),P v (t)};

[0039] C v (t) = {c v (t),c v' (t)};

[0040] P v (t) = {p v (t),p v' (t)};

[0041] Where O(t) is the observation space; c v (t) is the coordinate of UAV v at time slot t; c v' (t) is the coordinate of UAV v' observable by UAV v at time slot t, where And v≠v'; represents the drone v′ associated with drone v, where p v' (t) is the DAG partition point of UAV v' observable by UAV v at time slot t;

[0042] The expression of the action space is:

[0043]

[0044] Where A(t) is the action space; Indicates whether the task i generated by UAV v is offloaded to UAV v'; B v,v′ is the available bandwidth between UAV v and UAV v';

[0045] The expression of the reward function of the drone swarm is:

[0046]

[0047] Where R(t) is the reward function; β1 and β2 are weight coefficients respectively; d t The rewards obtained by all agents.

[0048] Preferably, the expression of the cumulative reward of the MAPPO algorithm is:

[0049]

[0050] Where V is the number of drones in the drone swarm; T is the total number of time slots for the video analysis task; r t (θ) is the cumulative reward of the drone group in time slot t; τ v is the experience sequence of the drone v during its interaction with the environment; R(τ) is the discounted reward, μ is the discount coefficient; Represents the sequence τ v The probability of occurrence.

[0051] Preferably, the multi-agent proximal policy optimization algorithm MAPPO is used to solve the optimization objective function, comprising the following steps:

[0052] Step S31: define the objective function of the target network as:

[0053]

[0054] In the formula, In the strategy π θ' The expectation of all trajectories under θ (α t ∣s t ) means that in a given state s t When strategy π θ Next, take action α t The probability of θ' (α t ∣o t ) means that in a given state s t When strategy π θ' Next, take action α t The probability of A θ '(s t ,α t ) represents the state s in a given t and action α t , the action-value function under the policy with parameter θ';

[0055] Step S32: Define the objective function of the value network as:

[0056]

[0057] In the formula, is the objective function of the value network of the MDP; T is the same number of time slots; Buffer for global experience traces; Indicates that in state s t Next, by the value function V φ The calculated state value; represents the reward at time t;

[0058] Step S33: define the objective function of the policy network, and update the parameter θ of each drone according to the objective function of the policy network. The objective function of the policy network is:

[0059]

[0060] In the formula, J θ' (θ) is the objective function of the policy network; Represents strategy π θ' Lower track τ v Expected value; In the strategy π θ Given the observation When you select Action probability; In the strategy π θ' Given the observation When you select Action probability; Indicates that at a given observation and actions Next, the advantage function is calculated based on the policy parameters θ';

[0061] Step S34: Limit the update amplitude of parameter θ to:

[0062]

[0063] A θ =π θ (s t ,a t )-V Φ (s t );

[0064]

[0065] In the formula, is the update amplitude of parameter θ; E t is the expected function; A θ is the substitution advantage function; r t (θ) is the cumulative reward of the drone group in time slot t; clip(r t (θ), 1-∈, 1+∈) is a clipping function used to restrict the input to the interval [1-∈, 1+∈]; ∈ is a small positive number; π θ (s t ,a t ) is state s t Next select action a t The probability of V Φ (s t ) means in state s t The expected reward obtained by acting according to the strategy Φ.

[0066] The present invention also provides a delay minimization system for collaborative video analysis of drone swarms, which is implemented based on the above-mentioned delay optimization method for collaborative video analysis of drone swarms, and includes a system construction module, a task generation module, a task execution module, a resource management module and a monitoring module.

[0067] System building blocks are used to build a video analytics system that includes multiple drones.

[0068] The task generation module is used to treat the video frames generated by the drone as inference tasks.

[0069] The task execution module is used to perform video analysis tasks, process the captured data using a deep learning model, and output video analysis results. The task-specific module is represented as a DAG. Each vertex in the DAG represents a deep neural network DNN classifier, which is responsible for classifying specific attributes of objects in the video frame.

[0070] Preferably, the resource management module uses deep reinforcement learning and heuristic algorithms to implement the system's task offloading decisions and resource allocation.

[0071] The monitoring module is used to monitor system performance and drone status in real time, collect feedback data, provide information support for the resource management module, and ensure that video analysis tasks are completed on time.

[0072] The benefits of the present invention include at least:

[0073] 1. This invention solves the delay minimization problem of video analysis in a decentralized drone swarm based on the multi-agent proximal policy optimization algorithm MAPPO and the heuristic algorithm, and can simultaneously optimize task scheduling, communication resource allocation, and DAG segmentation point selection;

[0074] 2. By updating the DAG split points, dynamic adjustment of task allocation between drones is achieved, which not only optimizes the use of computing resources, but also significantly reduces processing delays due to the reasonable allocation of tasks, thereby accelerating the video analysis process; in addition, the update of DAG split points also provides drones with the ability to flexibly offload tasks according to their current resource status and network conditions, further improving the responsiveness and efficiency of the system, allowing drones to work together efficiently and complete complex video analysis tasks even when resources are limited or network conditions change. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 A schematic diagram of a method flow of an embodiment of the present invention;

[0076] Figure 2 A schematic diagram of a directed acyclic graph DAG constructed for a drone in an embodiment of the present invention;

[0077] Figure 3 It is the convergence curve of the multi-agent proximal strategy optimization algorithm adopted in the embodiment of the present invention. DETAILED DESCRIPTION

[0078] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of the present invention.

[0079] like Figure 1 As shown, an embodiment of the present invention provides a delay optimization method for drone group collaborative video analysis, comprising the following steps:

[0080] Step S1: Construct a UAV cooperative video analysis system including multiple UAVs, the system including V UAVs. The UAVs are connected to each other wirelessly. Each UAV Real-time video can be collected from the ground and equipped with an engine to perform video inference tasks. These drones are not only able to handle tasks independently, but can also offload tasks to neighboring drones for additional processing. represents the drone v′ associated with drone v, where The multi-UAV system is described using a Cartesian coordinate system, where the position of the UAV is represented by d v ={x v ,y v Throughout the flight, the drone maintains a constant flight altitude to prevent collisions and minimize interference with other drones.

[0081] Step S2: Each drone generates a reasoning task. The reasoning task generated by drone v is recorded as and That is, drone v generates M v The tasks generated by the drone v are considered as delay-sensitive computing tasks M v,i ={S v,i ,T v,i},S v,i represents the size of the mission i of drone v, T v,i represents the maximum tolerable delay of task i for UAV v, emphasizing the importance of timely processing.

[0082] Step S3: Represent the task-specific module as a directed acyclic graph DAG. Figure 2As shown, each vertex in the DAG represents a deep neural network DNN classifier, which is responsible for classifying objects in video frames by specific attributes. Each directed edge in the DAG represents the flow of tasks between different DNN classifiers, that is, the output of one classifier is used as the input of the next classifier. In order to handle different types of detection objects, the embodiment of the present invention designs two branches for the DAG, one for vehicle recognition and the other for personnel recognition. According to the type of the detection object, the system first selects the corresponding branch. If the detection object is a vehicle, the DAG will further analyze its type, brand and color attributes; if the detection object is a person, the DAG will further analyze its gender and behavior attributes.

[0083] The embodiment of the present invention organizes the video frames into a series of inference tasks arranged in time order. Each drone processes these generated tasks in order according to the flow direction indicated by its DAG until the DAG split point is reached. At the same time, the tasks will be offloaded to the adjacent drones for further classification after the DAG split point.

[0084] Step S4: Establish an optimization objective function with the goal of minimizing task processing delay, use the distributed partially observable Markov decision process Dec-POMDP and the multi-agent proximal policy optimization algorithm MAPPO to solve the optimization objective function, and update the unloading decision and communication resource allocation of each drone according to the solution results. And solve the DAG split point of each drone according to the minimized delay.

[0085] Specifically, the task processing delay of the multi-UAV system is the time from receiving the workload to generating the inference result. Taking this into account, the task processing delay can be divided into two parts: communication delay and calculation delay. Since the UAVs are flying in the air, the embodiment of the present invention assumes that the communication between UAVs is a direct link. Then the communication delay of the multi-UAV system can be expressed as:

[0086]

[0087] L v,v' =Θ v,v' +Γ LoS ;

[0088]

[0089] In the above formula, is the communication delay between UAV v and UAV v'; S v,i (par v ) is based on the DAG partition point par v , the output data size of UAV v; R v→v' B is the transmission rate from UAV v to UAV v';v,v' is the available bandwidth between UAV v and UAV v'; P v The power required to transmit the UAV v; G v,v' is the channel gain between UAV v and UAV v'; σ 2 is the Gaussian white noise power; L v,v' is the path loss between UAV v and UAV v'; Γ LoS is the additional attenuation factor of the line-of-sight link between UAV v and UAV v'; d v,v' is the distance between UAV v and UAV v'; x v ,y v are the position coordinates of the drone v; x v' ,y v' is the position coordinate of the drone v'; f c is the carrier frequency, which is 2 GHz in the embodiment of the present invention; c is the speed of light.

[0090] Then the expression of task reasoning delay is:

[0091]

[0092] Where d is the inference delay of the drone swarm performing the video analysis task; is the delay for UAV v to perform its mission i; is the delay of transmitting task i generated on UAV v to UAV v’ for execution; Indicates whether the task i generated by drone v is offloaded to drone v', Indicates uninstallation. Indicates that it is not uninstalled.

[0093] Because the energy of the drone is limited and is provided by the built-in battery, the energy consumption of the drone needs to be considered when solving the optimization objective function. Including the computing energy consumption E of executing tasks v,exe , the transmission energy consumption of offloading tasks to other drones And the hovering energy consumption E of the drone hovering at a fixed height v,hov .

[0094] Calculate energy consumption E v,exe The expression is:

[0095]

[0096] In the formula, The computational energy consumption of UAV v performing task i; The computational energy consumption of executing task i generated by drone v’ on drone v;

[0097] Transmission energy consumption The expression is:

[0098]

[0099] Hover energy consumption E v,hov The expression is:

[0100]

[0101] Where P v,hov is the hovering power consumption of UAV v; t v,hov is the maximum hovering time of the UAV v; η is a value proportional to the mass of the UAV; is the power efficiency of UAV v; q is the number of rotors of UAV v; r is the diameter of each rotor; κ is the air density.

[0102] Energy consumption of drone v The expression is:

[0103]

[0104] In the embodiment of the present invention, the task offloading problem of the UAV is formalized as a distributed partially observable Markov decision process, Decentralized Partial Observable Markov Decision Process, or Dec-POMDP for short. In this case, the UAV is regarded as a distributed intelligent agent that implements the monitoring strategy by utilizing its local observations. Specifically, information exchange occurs between the UAV and its neighboring nodes within the communication range. The observation space, state space, action space, and reward function of this Dec-POMDP at time slot t are represented as O, S, A, and R.

[0105] Specifically, the system state includes the state information of all drones in the environment, whose positions change over time, so the state space S(t) at time slot t can be defined as:

[0106] S(t)={C(t),P(t)};

[0107] C(t)={(x 1,t ,y 1,t ),...(x v,t ,y v,t ),...,(x V,t ,y V,t )};

[0108] P(t)={p1(t),...p v (t),...,p V (t)};

[0109] Among them, C(t) is the coordinate of each drone in the drone group at time slot t, (x v,t ,y v,t ) is the coordinate of UAV v at time slot t; P(t) is the DAG segmentation point of each UAV in the UAV swarm at time slot t; p v (t) is the DAG split point of UAV v in time slot t.

[0110] In time slot t, the agent collects environmental information within the observation range. This information includes the location of the drone and the DAG partition point. Therefore, the observation space O(t) at time slot t can be expressed as:

[0111] O(t)={C v (t),P v (t)};

[0112] C v (t) = {c v (t),c v' (t)};

[0113] P v (t) = {p v (t),p v' (t)};

[0114] In the formula, c v (t) is the coordinate of UAV v at time slot t; c v' (t) is the coordinate of UAV v' observable by UAV v at time slot t, where And v≠v';p v' (t) is the DAG partition point of UAV v’ observable by UAV v at time slot t.

[0115] The action space A(t) is defined as:

[0116]

[0117] In the formula, Indicates whether the task i generated by UAV v is offloaded to UAV v'; B v,v' is the available bandwidth between UAV v and UAV v'.

[0118] Because the imbalance of energy consumption distribution between drones may lead to the neglect of certain tasks, resulting in performance degradation. Therefore, energy consumption balance is also crucial for managing power consumption within the drone network. In the embodiment of the present invention, the Jain fairness equation is used to evaluate the fairness of drone action selection. The rewards of all drones are defined as:

[0119]

[0120] Among them, β1 and β2 are weight coefficients, which are key factors to ensure the balanced distribution of energy consumption of UAV during the mission, and to achieve mission success; d t represents the rewards obtained by all agents;

[0121] Then the global cumulative reward of all drones is:

[0122]

[0123] Where V is the number of drones in the drone swarm, T is the total number of time slots for the video analysis task; r t (θ) is the global cumulative reward of the drone swarm in time slot t; τ v is the experience sequence of the drone v during its interaction with the environment; R(τ) is the discounted reward, μ is the discount coefficient; Represents the sequence τ v The probability of occurrence.

[0124] The objective function of the target network in the MAPPO algorithm is:

[0125]

[0126] In the formula, In the strategy π θ' The expectation of all trajectories under θ (α t ∣s t ) means that in a given state s t When strategy π θ Next, take action α t The probability of θ' (α t ∣o t ) means that in a given state s t When strategy π θ' Next, take action α t The probability of A θ '(s t ,α t ) represents the state s in a given t and action α t , using the action-value function under the policy with parameter θ'.

[0127] In the centralized training phase of Dec-POMDP, the embodiment of the present invention converts a part of the value network of the global state into MDP. In this phase, the objective function of the value network of MDP is first defined as:

[0128]

[0129] In the formula, is the objective function of the value network of MDP; is the global experience pool trajectory buffer, used to store the trajectory τ s ,in Indicates that in state s t Next, by the value function V φ The calculated state value; represents the reward at time t.

[0130] Furthermore, the objective function of the MDP policy network is defined to update the parameters θ of each drone. The main optimization goal of the MDP policy network is:

[0131]

[0132] In the formula Represents strategy π θ' Lower track τ i Expected value; In the strategy π θ Given the observation When you select Action probability; In the strategy π θ' Given the observation When you select Action probability; Indicates that at a given observation and actions Below, the advantage function is calculated based on the policy parameters θ'.

[0133] The goal here is to limit the magnitude of each θ update, which can be achieved by applying parameter ∈ clipping to the objective function, expressed as:

[0134]

[0135] A θ =π θ (s t ,a t )-V Φ (s t );

[0136]

[0137] In the formula, is the update amplitude of parameter θ; E t is the expected function; A θ is the substitution advantage function; r t (θ) is the cumulative reward of the drone group in time slot t; clip(r t(θ), 1-∈, 1+∈) is a clipping function used to restrict the input to the interval [1-∈, 1+∈]; ∈ is a small positive number; π θ (s t ,a t ) is state s t Next select action a t The probability of V Φ (s t ) means in state s t The expected reward obtained by acting according to the strategy Φ.

[0138] Step S5: Repeat step S4 until the maximum number of iterations is reached or the reward curve of the MAPPO algorithm converges, and obtain the drone swarm task allocation plan with the minimum delay for the video analysis task. And solve the DAG segmentation point of each drone based on the minimized delay.

[0139] The delay optimization method for cooperative video analysis of drone swarms provided by the present invention solves the delay minimization problem of video analysis in a decentralized drone swarm. The adopted MAPPO algorithm and heuristic algorithm have good convergence performance and optimization performance, and can simultaneously optimize task scheduling, communication resource allocation, and DAG segmentation point selection.

[0140] An embodiment of the present invention also provides a delay minimization system for collaborative video analysis of a drone swarm, which is implemented based on the above-mentioned delay optimization method for collaborative video analysis of a drone swarm, and includes a system construction module, a task generation module, a task execution module, a resource management module and a monitoring module.

[0141] System building blocks are used to build a video analytics system that includes multiple drones.

[0142] The task generation module is used to treat the video frames generated by the drone as inference tasks.

[0143] The task execution module is used to perform video analysis tasks, process the captured data using a deep learning model, and output video analysis results. The task-specific module is represented as a DAG. Each vertex in the DAG represents a deep neural network DNN classifier, which is responsible for classifying specific attributes of objects in the video frame.

[0144] The resource management module uses deep reinforcement learning and heuristic algorithms to make task offloading decisions and resource allocation based on the output results of the task execution module and the current state of the drone to optimize the inference latency and drone energy consumption.

[0145] The monitoring module is used to monitor system performance and drone status in real time, collect feedback data, provide information support for the resource management module, and ensure that video analysis tasks are completed on time.

[0146] Extensive simulations are performed to evaluate the performance of our algorithm under various conditions and scenarios and compare it with other baseline solutions. The simulations are implemented using PyTorch 2.3.0 and Python 3.8.19 on an Intel Core i7-13700KF CPU server.

[0147] The number of captured video frames was randomly selected in the range of 5 to 30. OpenCV 3.2.0 was used for real-time video processing. In terms of the inference engine, the efficient yolov3-tiny model was used for each classifier.

[0148] The target area is set as an area of ​​500m × 500m. There are 11 drones in total, each equipped with a camera to collect videos of the observed area and start the inference process. In addition, the monitoring radius covered by each drone is 100m. We used two types of drones, each equipped with a Raspberry Pi 3B+ and an Intel Neural Compute Stick NCS. Under the same working conditions, the advanced NCS2 shows superior performance than its predecessor in terms of frame rate and energy efficiency. The processing latency and energy consumption based on the YOLOv3-tiny model can be measured in advance. The latency of the Raspberry Pi 3B+ combined with the NCS1 execution model is 0.25s and the power consumption is 6w; the latency of the RaspberryPi 3B+ combined with the NCS2 execution model is 0.2s and the power consumption is 5.6w. P v 30dBm, P v,hov is 60w, η is 30N, is 70%, q is 4, r is 0.254m, and κ is 1.225kg / m 3 .

[0149] like Figure 3 Shown is the convergence curve of the MAPPO algorithm used in the embodiment of the present invention. Figure 3 The graph shows the changing trend of the average reward of the multi-UAV system, that is, the accumulation of the average reward obtained by each UAV in one step. The embodiment of the present invention uses 11 UAVs as experimental objects and evaluates the convergence curve of the algorithm when the learning rate is 0.005. The graph shows the good convergence performance of the algorithm.

[0150] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. Only the preferred embodiments of the present invention are expressed. The description is more specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. As long as there is no contradiction in the combination of these technical features, they should be considered as the scope recorded in this specification.

[0151] It should be noted that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these modifications and improvements all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A delay optimization method for drone swarm collaborative video analysis, characterized in that: The following steps are involved: Step S1: constructing a UAV cooperative video analysis system including multiple UAVs, each of which collects real-time video from the ground and generates an inference task; Step S2: generating a directed acyclic graph DAG for each drone according to the reasoning task, initializing the segmentation point of the DAG of each drone, and each drone performs task unloading at the segmentation point; Step S3: establishing an optimization objective function with the goal of minimizing task processing delay, solving the optimization objective function, and updating the unloading decision, communication resource allocation and DAG split point of each UAV according to the solution result; Step S4: Repeat step S3 until the set maximum number of iterations is reached, and obtain the drone swarm task allocation plan with the minimum delay in the video analysis task.

2. The delay optimization method for drone swarm collaborative video analysis according to claim 1 is characterized by: The task processing delay in step S2 includes communication delay and calculation delay. The expression of the communication delay is: L v,v' =Θ v,v' +C LoS ; In the above formula, is the communication delay between UAVs; S v,i (par v ) is the DAG partition point par of drone v v The output data size of R v→v′ B is the transmission rate from UAV v to UAV v'; v,v′ is the available bandwidth between UAV v and UAV v'; P v The power required to transmit the UAV v; G v,v′ is the channel gain between UAV v and UAV v'; σ 2 is the Gaussian white noise power; L v,v' is the path loss between UAV v and UAV v'; Θ v,v' = LoS is the additional attenuation factor of the line-of-sight link between UAV v and UAV v'; d v,v' is the distance between UAV v and UAV v'; f c is the carrier frequency; c is the speed of light; x v ,y v are the position coordinates of the drone v; x v' ,y v' is the position coordinate of the UAV v'; The expression of the task processing delay is: Where d is the task processing delay of the drone swarm performing the video analysis task; is a drone swarm; i∈M v , M v Video analysis tasks generated for drone v; is the delay of executing task i on UAV v; is the latency of task i generated on UAV v and transmitted to UAV v’ for execution; Indicates whether the task i generated by drone v is offloaded to drone v', Indicates uninstallation. Indicates that it has not been uninstalled; The latency for executing task i generated by drone v on drone v'.

3. The delay optimization method for drone swarm collaborative video analysis according to claim 2 is characterized by: When solving the optimization objective function in step S3, the energy consumption of the drone needs to be considered. Including the computing energy consumption of executing tasks Transmission energy consumption of offloading tasks to other drones Transmission energy consumption from UAV to ground base station E v→b And the hovering energy consumption E of the drone hovering at a fixed height v,hov ; The energy consumption of the calculation The expression is: In the formula, The computational energy consumption of UAV v performing task i; The computational energy consumption of executing task i generated by drone v’ on drone v; The transmission energy consumption The expression is: The hovering energy consumption E v,hov The expression is: Where P v,hov is the hovering power consumption of UAV v; t v,hov is the maximum hovering time of the UAV v; η is a value proportional to the mass of the UAV; is the power efficiency of UAV v; q is the number of rotors of UAV v; r is the diameter of each rotor; κ is the air density; The energy consumption of the drone v The expression is:

4. The delay optimization method for drone swarm collaborative video analysis according to claim 1 is characterized by: Solving the optimization objective function in step S3 includes the following steps: using a distributed partially observable Markov decision process Dec-POMDP to define the state space, observation space, action space and reward function; based on the defined state space, observation space, action space and reward function, using a multi-agent proximal policy optimization algorithm MAPPO to solve the optimization objective function.

5. The delay optimization method for drone swarm collaborative video analysis according to claim 4 is characterized by: The expression of the state space is: S(t)={C(t),P(t)}; C(t)={(x 1,t ,y 1,t ),...(x v,t ,y v,t ),...,(x V,t ,y V,t )}; P(t)={p1(t),...p v (t),...,p V (t)}; Among them, S(t) is the state space; C(t) is the coordinate of each drone in the drone group at time slot t, (x v,t ,y v,t ) is the coordinate of UAV v at time slot t; P(t) is the DAG segmentation point of each UAV in the UAV swarm at time slot t; p v (t) is the DAG segmentation point of UAV v in time slot t; The expression of the observation space is: O(t)={C v (t),P v (t)}; C v (t)={c v (t),c v' (t)}; P v (t)={p v (t),p v' (t)}; Where O(t) is the observation space; c v (t) is the coordinate of UAV v at time slot t; c v' (t) is the coordinate of UAV v' observable by UAV v at time slot t, where And v≠v'; represents the drone v′ associated with drone v, where v, v≠v′;p v' (t) is the DAG partition point of UAV v' observable by UAV v at time slot t; The expression of the action space is: Where A(t) is the action space; Indicates whether the task i generated by UAV v is offloaded to UAV v'; B v,v′ is the available bandwidth between UAV v and UAV v'; The expression of the reward function of the drone swarm is: Where R(t) is the reward function; β1 and β2 are weight coefficients respectively; d t The rewards obtained by all agents.

6. The delay optimization method for drone swarm collaborative video analysis according to claim 4 is characterized by: The expression of the cumulative reward of the MAPPO algorithm is: Where V is the number of drones in the drone swarm; T is the total number of time slots for the video analysis task; r t (θ) is the cumulative reward of the drone group in time slot t; τ v is the experience sequence of the drone v during its interaction with the environment; Represents the sequence τ v The probability of occurrence.

7. The delay optimization method for drone swarm collaborative video analysis according to claim 4 is characterized by: The method of using the multi-agent proximal strategy optimization algorithm MAPPO to solve the optimization objective function includes the following steps: Step S31: define the objective function of the target network as: In the formula, In the strategy π θ' The expectation of all trajectories under θ (α t ∣s t ) means that in a given state s t When strategy π θ Next, take action α t The probability of θ' (α t ∣o t ) means that in a given state s t When strategy π θ' Next, take action α t The probability of A θ '(s t ,α t ) represents the state s in a given t and action α t , the action-value function under the policy with parameter θ'; Step S32: Define the objective function of the value network as: In the formula, is the objective function of the value network of the MDP; T is the same number of time slots; Buffer for global experience traces; Indicates that in state s t Next, by the value function V φ The calculated state value; represents the reward at time t; Step S33: define the objective function of the policy network, and update the parameter θ of each drone according to the objective function of the policy network. The objective function of the policy network is: In the formula, J θ′ (θ) is the objective function of the policy network; Represents strategy π θ' Lower track τ v Expected value; In the strategy π θ Given the observation When you select Action probability; In the strategy π θ' Given the observation When you select Action probability; Indicates that at a given observation and actions Next, the advantage function calculated based on the policy parameter θ'; Step S34: Limit the update amplitude of parameter θ to: A θ =π θ (s t ,a t )-V Φ (s t ); In the formula, is the update amplitude of parameter θ; E t is the expected function; A θ is the substitution advantage function; r t (θ) is the cumulative reward of the drone group in time slot t; clip(r t (θ), 1-∈, 1+∈) is a clipping function used to restrict the input to the interval [1-∈, 1+∈]; ∈ is a small positive number; π θ (s t ,a t ) is state s t Next select action a t The probability of V Φ (s t ) means in state s t The expected reward obtained by acting according to the strategy Φ.

8. A delay optimization system for drone swarm collaborative video analysis, implemented based on a delay optimization method for drone swarm collaborative video analysis as claimed in any one of claims 1 to 7, characterized in that: It includes system construction module, task generation module, task execution module, resource management module and monitoring module; The system building module is used to build a video analysis system including multiple drones; The task generation module is used to decompose the tasks generated by the drone into tasks and construct corresponding DAGs to represent the execution order between tasks; The task execution module is used to execute the video analysis task, process the captured data using the deep learning model, and output the video analysis results; The resource management module makes task offloading decisions and resource allocation according to the output results of the task execution module and the current state of the UAV to optimize the overall energy consumption; The monitoring module is used to monitor system performance and drone status in real time, collect feedback data, provide information support for the resource management module, and ensure that the video analysis task is completed on time.

9. The delay optimization system for drone swarm collaborative video analysis according to claim 8, characterized in that: The task generation module decomposes the video analysis task generated by the drone into multiple tasks, takes the multiple tasks as vertices, takes the execution order between the multiple tasks as directed edges, and constructs a directed acyclic graph DAG.

10. The delay optimization system for drone swarm collaborative video analysis according to claim 8, characterized in that: The resource management module uses deep reinforcement learning and heuristic algorithms to implement the system's task offloading decisions and resource allocation.

Images