Real-time UAV path optimization system based on edge computing

Abstract

The present invention is applicable to the technical field of drone path planning, and in particular to a real-time drone path optimization system based on edge computing, the system comprising: a data acquisition unit for collecting raw environmental data from a multimodal sensor and transmitting it to an edge processing unit; an edge processing unit for performing data fusion on the raw environmental data input from the data acquisition unit and outputting fusion features; a path optimization unit for performing path search and path optimization based on the fusion features output by the edge processing unit, generating a flight path, and controlling the drone through the flight path. The present invention deeply couples the deeply fused spatiotemporal characteristics of the environment with the path optimization algorithm, breaking the separation between perception and planning, realizing an end-to-end closed-loop optimization process, greatly improving the adaptability and overall robustness of the system, and being able to output a continuous and feasible drone flight path in real time through local smoothing and multi-objective evaluation mechanisms.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned aerial vehicle (UAV) path planning, and in particular relates to a real-time UAV path optimization system based on edge computing. Background Art

[0002] Path optimization technology aims to calculate the optimal flight path in real time based on environmental data and mission requirements to achieve efficient mission execution and safe flight. This method mainly explores how to dynamically generate and adjust the flight path under conditions of rapid and complex environmental changes to maximize the efficiency and safety of drone flights. In the present invention, for complex environments with dynamic, intensive and high-frequency changes, path optimization technology uses an edge computing module to achieve high-frequency, concurrent real-time processing of environmental data and multi-objective joint optimization. It can adaptively adjust and reconstruct the flight path based on the latest perception information in scenarios where multiple factors such as obstacles and mission requirements are constantly changing, and quickly respond to various emergencies, thereby ensuring that the drone can complete its mission safely and efficiently under dynamic and complex environmental conditions.

[0003] Existing drone path optimization technologies have significant shortcomings in terms of real-time performance and adaptability to dynamic environments. Most methods rely on planning strategies based on preset paths, which are only suitable for relatively static and predictable environments. These strategies struggle to respond promptly to unexpected obstacles or environmental changes that may arise during flight, leading to increased flight safety risks and reduced mission efficiency in practical applications.

[0004] In addition, although cloud computing-based optimization methods have certain flexibility and dynamic adjustment capabilities, they are limited by network communication delays and insufficient bandwidth. According to relevant industry standards, the typical response delay of existing mainstream cloud-based drone path planning solutions is generally greater than 300ms, and in some scenarios exceeds 800ms, which is far higher than the real-time requirements of rigid scenarios such as dynamic obstacle avoidance (usually <100ms). It is difficult to achieve rapid interaction between environmental perception data and path optimization instructions, and cannot meet the real-time response needs of drones in complex dynamic environments. Summary of the Invention

[0005] The purpose of the present invention is to provide a real-time path optimization system for UAVs based on edge computing, aiming to solve the problem that existing path planning methods are difficult to achieve rapid interaction between environmental perception data and path optimization instructions, and cannot meet the real-time response requirements of UAVs in complex dynamic environments.

[0006] The present invention is implemented as follows: a real-time path optimization system for unmanned aerial vehicles based on edge computing, the system comprising:

[0007] A data acquisition unit, configured to collect raw environmental data from a multimodal sensor and transmit the data to an edge processing unit;

[0008] An edge processing unit, configured to perform data fusion on the raw environmental data input from the data acquisition unit and output fusion features;

[0009] The path optimization unit is used to perform path search and path optimization based on the fusion features output by the edge processing unit, generate a flight path, and control the drone through the flight path.

[0010] Preferably, the data acquisition unit acquires observation data from various sensors and performs preliminary data processing, synchronously acquiring multimodal raw environmental data at each time t. The raw measurement data of the i-th sensor at time t is:

[0011] ;

[0012] in, represents the original observation of the first sensor, for dimensional real vector space;

[0013] After being transmitted via the internal bus, the raw measurement data undergoes unified time synchronization and light preprocessing, using the mapping function:

[0014] ;

[0015] in, Indicates the Mapping function for unified time synchronization, denoising, and lightweight preprocessing of the original observation data. for dimensional real vector space;

[0016] Generate a consistent and denoised observation vector:

[0017] ;

[0018] in, Indicates the Road sensor at all times The observation vector after preprocessing.

[0019] Preferably, the edge processing unit performs data fusion on the original environmental data input from the data acquisition unit and outputs the fusion feature process, which specifically includes:

[0020] The sensor data is preliminarily encoded through a lightweight neural network mapping function to construct a unified output d-dimensional feature vector;

[0021] Calculate the cosine similarity between any two features and construct a dynamic adjacency matrix;

[0022] Perform local fusion of graph attention network on the dynamic adjacency matrix and output the local fusion result;

[0023] Multiple sets of local fusion results are used as sequences, mapped into query matrix, key matrix and value matrix respectively, and multi-head self-attention calculation is performed to generate a global fusion feature sequence;

[0024] Collect the nearest preset number of global features, construct the temporal query matrix, key matrix and value matrix, perform a self-attention calculation, and output the global spatiotemporal fusion features.

[0025] Preferably, in the step of preliminarily encoding the sensor data through a lightweight neural network mapping function and constructing a unified output d-dimensional feature vector, the lightweight neural network mapping function is expressed as:

[0026] ;

[0027] in, Indicates the Feature encoding mapping function of road sensor data, For the Road sensor at all times The denoised observation vector, express dimensional real number space;

[0028] For the first Road sensor data is preliminarily encoded and output uniformly -dimensional feature vector:

[0029] ;

[0030] in, For the corresponding -dimensional feature vector.

[0031] Preferably, the process of calculating the cosine similarity between any two features and constructing a dynamic adjacency matrix includes:

[0032] Calculate the cosine similarity of any two features:

[0033] ;

[0034] in, Indicates at time No. Road and The cosine similarity between the path feature vectors reflects the correlation between the two path features. represents the eigenvector of the i-th path;

[0035] Introducing temperature parameters Normalized by softmax:

[0036] ;

[0037] in, Indicates the first Rank The adjacency weight of the column reflects the Road characteristics Contribution of road features, is a positive temperature parameter that controls the smoothness of the softmax distribution. is the total number of features;

[0038] Get the dynamic adjacency matrix

[0039] ;

[0040] The sum of its i-th row is 1. When the sensor signal quality fluctuates or fails, the weight of its corresponding row and column is automatically lowered.

[0041] Preferably, the process of performing local fusion of the graph attention network on the dynamic adjacency matrix and outputting the local fusion result includes:

[0042] For each node i, select the adjacent weighted The index set of the first k , and adopts a learnable attention vector and transformation matrix Calculate the attention coefficient:

[0043] ;

[0044] in, is the learnable attention vector, is the feature transformation matrix, For the Nodes at time The neighbor index set of For the The feature vector of each node, For the The feature vector of each node, is the set of neighbor nodes The traversal subscript in is used to normalize the denominator accumulation of softmax, Represents vector concatenation operation;

[0045] Then the neighbor features are weighted and aggregated to obtain the local fusion result:

[0046] ;

[0047] in, For the The local fusion feature vector of the nodes after weighted aggregation.

[0048] Preferably, multiple sets of local fusion results are used as sequences, mapped into query matrices, key matrices, and value matrices respectively, and multi-head self-attention calculations are performed to generate a global fusion feature sequence. The process includes:

[0049] Will Treated as a sequence of length n, it is mapped into query matrix, key matrix and value matrix respectively:

[0050] ;

[0051] in, are the linear transformation matrices for query, key, and value, respectively, is the number of nodes, is the feature dimension, parallel The head performs multi-head self-attention calculation:

[0052] ;

[0053] in, are the query matrix, key matrix and value matrix of the first head respectively, It is a row-normalized softmax operation; the outputs of each head are concatenated and linearly mapped , generate a global fusion feature sequence .

[0054] Preferably, the process of collecting the latest preset number of global features, constructing a temporal query matrix, a key matrix, and a value matrix, performing a self-attention calculation, and outputting a global spatiotemporal fusion feature includes:

[0055] Collect the global fusion features of the latest T frames , construct time series query matrix, key matrix and value matrix , perform a self-attention calculation:

[0056] ;

[0057] in, Represents the history The contribution weight of the frame to the current fusion result.

[0058] Preferably, the path optimization unit performs path search and path optimization based on the fusion features output by the edge processing unit, generates a flight path, and controls the UAV through the flight path, including:

[0059] The path optimization unit obtains the global spatiotemporal fusion features and constructs a comprehensive state vector based on the UAV’s state information;

[0060] A policy sampling network trained by deep reinforcement learning is used, which takes a comprehensive state vector as input and outputs candidate actions or target nodes for path extension.

[0061] Based on the candidate nodes output by the strategy sampling network, combined with the current path tree structure, according to the RRT expansion principle, the tree node closest to the sampling point is selected, and new nodes are generated along this direction to generate multiple candidate paths;

[0062] A multi-objective evaluation function is constructed to screen candidate paths, output the optimal path, and control the UAV through the optimal path.

[0063] Preferably, during the flight of the UAV, it continuously receives new perception states and judges in real time whether there are sudden changes in the environment. When a new obstacle is detected, a new state perception, sampling, tree expansion and path optimization process is triggered to dynamically correct the flight path.

[0064] The edge computing-based real-time path optimization system for drones provided by the present invention has the following beneficial effects:

[0065] Deeply coupling the deeply integrated spatiotemporal characteristics of the environment with the path optimization algorithm breaks the separation between perception and planning, realizes an end-to-end closed-loop optimization process, and greatly improves the system's adaptability and overall robustness. Through local smoothing and multi-objective evaluation mechanisms, a continuous and highly feasible drone flight path can be output in real time. This invention has significant technical advantages and practical value in handling dynamic environments, improving multi-objective optimization capabilities, and achieving the integration of perception and planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is an architecture diagram of a real-time UAV path optimization system based on edge computing provided by an embodiment of the present invention;

[0067] Figure 2 Schematic diagram of the DGTF algorithm flow for the edge processing unit;

[0068] Figure 3 Schematic diagram of the RL-RRT algorithm flow of the path optimization unit. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0070] Edge computing technology refers to the process of processing data close to the data source or data collection end, aiming to achieve efficient real-time data analysis and decision-making. Research in this field focuses on how to deploy effective computing resources on the device side to reduce data transmission latency and improve response speed. In the application scenario of this invention, edge computing uses localized data processing and path optimization algorithms to quickly convert the environmental data collected by drones in real time into executable path adjustment strategies, thereby improving the mission efficiency and safety of drones.

[0071] Environmental perception technology integrates sensor data to perceive changes in the drone's surroundings in real time, including obstacle locations and dynamic weather conditions. This information is then generated accurately and comprehensively to enhance the timeliness and precision of the drone's flight path adjustments. Research in this field typically focuses on the design of sensor fusion and data analysis algorithms to maximize the accuracy and real-time nature of environmental perception. In this invention, environmental perception technology, through the close collaboration of real-time data collection and edge computing, provides rich environmental data support, ensuring the accuracy and reliability of drone path optimization.

[0072] Through the synergy of edge computing, environmental perception and path optimization technologies, the present invention constructs a localized real-time path optimization system for drones. The edge computing module in the system realizes the real-time synchronous acquisition and parallel preprocessing of multi-source sensor data, providing efficient data support for environmental perception and path optimization. The environmental perception unit dynamically extracts high-dimensional features that reflect changes in the flight environment by fusing multi-type sensor data, and can characterize obstacles and risk information in complex environments. The path optimization unit uses the fused features as input, combines historical path information, and realizes real-time dynamic adjustment of the drone's flight path based on a multi-objective optimization algorithm. Through data flow and feedback, the three realize the timely transmission of information and high-frequency update of the path, effectively breaking through the previous problem of delayed response of a single optimization link and difficulty in adapting to a dynamic environment. The present invention can continuously ensure the efficient and safe flight of drones in complex dynamic environments, and is suitable for scenarios such as logistics distribution and agricultural monitoring.

[0073] The proposed edge computing-based real-time drone path optimization system integrates environmental perception, intelligent data fusion, and dynamic path optimization locally on the drone, creating an efficient closed-loop structure with collaborative data acquisition, edge processing, and path optimization. The system's core innovation lies in the fact that all data perception, analysis, and decision-making processes are performed locally on the drone, significantly improving the real-time performance and environmental adaptability of path optimization.

[0074] like Figure 1 FIG. 1 is an architecture diagram of a real-time UAV path optimization system based on edge computing provided by an embodiment of the present invention. The system includes:

[0075] The data acquisition unit is used to collect raw environmental data from the multimodal sensor and transmit it to the edge processing unit.

[0076] In this system, the data acquisition unit is responsible for synchronously collecting raw environmental data from multimodal sensors (including cameras, lidar, and IMUs). It ensures the consistency of multi-source data through a unified timestamp and hardware synchronization mechanism. The collected multimodal raw data is input to the edge processing unit via an internal bus.

[0077] Specifically, the data acquisition unit is responsible for acquiring observation data from various sensors and performing preliminary data processing. At each time t, multimodal raw data is synchronously collected. The raw measurement of the i-th sensor at time t is:

[0078] ;

[0079] After transmission via the internal bus, the system performs unified time synchronization and light preprocessing, which is done by the mapping function:

[0080] ;

[0081] Generate consistent and denoised observation vectors:

[0082] ;

[0083] The images, point clouds and inertial measurements contained in the raw data are all time-scaled and filtered at this stage and output as .

[0084] The edge processing unit is used to perform data fusion on the original environmental data input from the data acquisition unit and output fusion features.

[0085] In this system, if Figure 2 As shown, in the edge processing unit, for the pre-processed data , proposed the DGTF fusion algorithm, which maps these multimodal data end-to-end into high-dimensional features containing spatiotemporal information , its process mainly uses lightweight neural network mapping function:

[0086] ;

[0087] Perform preliminary encoding on the i-th sensor data and uniformly output the d-dimensional feature vector:

[0088] ;

[0089] This step ensures that different modal data have consistent metrics in the same feature space.

[0090] Next, in order to reflect the correlation between sensors in real time, the system calculates the cosine similarity of any two features:

[0091] ;

[0092] And introduce the temperature parameter Normalized by softmax:

[0093] ;

[0094] This gives the dynamic adjacency matrix:

[0095] ;

[0096] The sum of its i-th row is 1, which can be updated online. If the signal quality of a certain sensor fluctuates or fails, the weight of the corresponding row and column will be automatically lowered.

[0097] In the construction Afterwards, the graph attention network (GAT) local fusion is further performed on the graph structure.

[0098] Specifically, for each node i, select the adjacent weighted The index set of the first k , and utilizes learnable attention vectors and transformation matrix Calculate the attention coefficient:

[0099] ;

[0100] Then the neighbor features are weighted and aggregated to obtain the local fusion result:

[0101] ;

[0102] This step not only highlights the channel information most closely related to node i, but also dynamically suppresses noise and redundancy during parameter training.

[0103] In order to further integrate global context information, Treated as a sequence of length n, it is mapped into query, key, and value matrices respectively:

[0104] ;

[0105] in, , and perform multi-head self-attention calculations in parallel:

[0106] ;

[0107] Then the outputs of each head are concatenated and linearly mapped , and finally generate a global fusion feature sequence This layer captures the long-range complementary information between non-neighboring nodes and makes up for the limitation of local aggregation.

[0108] After spatial fusion is completed, in order to smooth the jitter between frames and utilize short-term dynamic information, the algorithm collects the global features of the most recent T frames. , construct time series query, key matrix , perform a self-attention:

[0109] ;

[0110] Among them, the coefficient Represents the history The contribution weight of the frame to the current fusion result. It has been fully integrated in both spatial and temporal dimensions, reflecting environmental characteristics such as obstacle distribution while also taking into account the reliability of each sensor. , flight target position Together, they form the state input of the reinforcement learning-rapid random tree (RL-RRT) path optimization unit, achieving seamless connection from perception to planning, such as Figure 3 As shown, the DGTF algorithm of the present invention can be used on edge devices by adjusting parameters As well as model quantization / pruning and other means, it meets the real-time update requirements of 20-30 Hz and demonstrates good robustness and accuracy in complex dynamic environments.

[0111] The path optimization unit is used to perform path search and path optimization based on the fusion features output by the edge processing unit, generate a flight path, and control the drone through the flight path.

[0112] In this system, the path optimization unit combines the high-dimensional global spatiotemporal fusion features output by the edge processing unit As a global perception expression of the current environment. At the same time, the current position of the drone , row target position , historical path The information together constitutes the comprehensive state vector:

[0113] ;

[0114] This state vector describes multiple factors such as the drone's perception environment, spatial position, target task, and path prior at the current moment, and is the basic input for path decision-making.

[0115] During the path optimization process, the system uses a policy sampling network trained by deep reinforcement learning ,by As input, output path expansion candidate action or target node:

[0116] ;

[0117] in, is the set of policy network parameters. This policy network can fully exploit historical experience and environmental characteristics to achieve adaptive and efficient optimization of path sampling in dynamic scenarios. Compared with the uniform random sampling in traditional RRT, it can more intelligently guide the path to expand to the best area.

[0118] Based on the candidate nodes output by the policy network, combined with the current path tree structure , the path optimization unit selects the tree node closest to the sampling point according to the RRT expansion principle , and generate new nodes along this direction :

[0119] ;

[0120] in, represents the expansion step. The feasibility of the new node is determined by the environmental model to ensure that constraints such as accessibility and compliance are met.

[0121] As the tree continues to expand, the system generates multiple path candidates in real time. , the present invention designs a multi-objective evaluation function , which is used to comprehensively measure multiple indicators such as the safety, distance, and energy consumption of the drone path, so as to achieve multi-objective optimization in path planning:

[0122] ;

[0123] in, represents the kth evaluation index, is a weight factor, which facilitates flexible adjustment of the focus of optimization objectives in different application scenarios. Finally, the system selects the optimal path from all feasible path candidates based on the comprehensive evaluation score:

[0124] ;

[0125] This optimal path not only meets the constraints and safety requirements, but also can adapt to the current environment and mission requirements to achieve global optimization of flight performance and mission efficiency.

[0126] During the actual flight execution, the path optimization unit continuously receives new perception states. , it can judge in real time whether there are sudden changes in the environment. For example, if a new obstacle is detected, it will immediately trigger a new state perception, sampling, tree expansion and path optimization process, and dynamically correct the flight path to ensure that the UAV is always under the guidance of the optimal path and completes the task autonomously, safely and efficiently.

[0127] The comprehensive expression is:

[0128] ;

[0129] in, Represents the full-process mapping relationship of path optimization based on the fusion of reinforcement learning and random tree search. is the fusion perception feature state at the current moment, Indicates the flight path at the current moment, Indicates the target point path, Represents the path information at a historical moment, is the set of parameters used in the reinforcement learning optimization process.

[0130] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A real-time path optimization system for drones based on edge computing, characterized in that: The system comprises: A data acquisition unit, configured to collect raw environmental data from a multimodal sensor and transmit the data to an edge processing unit; An edge processing unit, configured to perform data fusion on the raw environmental data input from the data acquisition unit and output fusion features; A path optimization unit, configured to perform path search and path optimization based on the fusion features output by the edge processing unit, generate a flight path, and control the UAV through the flight path; The edge processing unit performs data fusion on the raw environmental data input from the data acquisition unit and outputs the fusion feature process, which specifically includes: The sensor data is preliminarily encoded through a lightweight neural network mapping function to construct a unified output d-dimensional feature vector; Calculate the cosine similarity between any two features and construct a dynamic adjacency matrix; Perform local fusion of graph attention network on the dynamic adjacency matrix and output the local fusion result; Multiple sets of local fusion results are used as sequences, mapped into query matrix, key matrix and value matrix respectively, and multi-head self-attention calculation is performed to generate a global fusion feature sequence; Collect the latest preset number of global features, construct the time series query matrix, key matrix and value matrix, perform a self-attention calculation, and output the global spatiotemporal fusion features; The path optimization unit performs path search and path optimization based on the fusion features output by the edge processing unit, generates a flight path, and controls the drone through the flight path. The process includes: The path optimization unit obtains the global spatiotemporal fusion features and constructs a comprehensive state vector based on the UAV’s state information; A policy sampling network trained by deep reinforcement learning is used, which takes a comprehensive state vector as input and outputs candidate actions or target nodes for path extension. Based on the candidate nodes output by the strategic sampling network and the current path tree structure, the tree node closest to the sampling point is selected according to the RRT expansion principle, and new nodes are generated along the direction from the nearest tree node to the sampling point to generate multiple candidate paths; A multi-objective evaluation function is constructed to screen candidate paths, output the optimal path, and control the UAV through the optimal path.

2. The UAV real-time path optimization system based on edge computing according to claim 1 is characterized in that: The data acquisition unit obtains observation data from various sensors and performs preliminary data processing. Synchronously collect multi-modal raw environmental data, The original measurement data of the road sensor at time is: ； in, represents the original observation of the first sensor, for dimensional real vector space; After being transmitted via the internal bus, the raw measurement data undergoes unified time synchronization and light preprocessing, using the mapping function: ； in, Indicates the Mapping functions for unified time synchronization, denoising and lightweight preprocessing of raw observation data. for dimensional real vector space; Generate a consistent and denoised observation vector: ； in, Indicates the Road sensor at all times The observation vector after preprocessing is -dimensional real vector.

3. The real-time path optimization system for UAVs based on edge computing according to claim 2 is characterized in that: In the step of preliminarily encoding the sensor data through a lightweight neural network mapping function and constructing a unified output d-dimensional feature vector, the lightweight neural network mapping function is expressed as: ； in, Indicates the Feature encoding mapping function of road sensor data, For the Road sensor at all times The denoised observation vector, express dimensional real number space; For the first Road sensor data is preliminarily encoded and output uniformly -dimensional feature vector: ； in, For the corresponding -dimensional feature vector.

4. The real-time path optimization system for UAVs based on edge computing according to claim 3 is characterized in that: The process of calculating the cosine similarity between any two features and constructing a dynamic adjacency matrix includes: Calculate the cosine similarity of any two features: ； in, Indicates at time No. Road and The cosine similarity between the path feature vectors reflects the correlation between the two path features. represents the eigenvector of the i-th path; Introducing temperature parameters Normalized by softmax: ； in, Indicates the first Rank The adjacency weight of the column reflects the Road characteristics Contribution of road features, is a positive temperature parameter that controls the smoothness of the softmax distribution. is the total number of features; Get the dynamic adjacency matrix; ； in, is a dynamic adjacency matrix, whose The sum of the rows is 1. When the sensor signal quality fluctuates or fails, the weight of the corresponding row and column is automatically lowered.

5. The real-time path optimization system for UAVs based on edge computing according to claim 4 is characterized in that: The process of performing local fusion of the graph attention network on the dynamic adjacency matrix and outputting the local fusion result includes: For each node , select the adjacency weight sort The index set of the first k , and adopts a learnable attention vector and transformation matrix Calculate the attention coefficient: ； in, is the learnable attention vector, is the feature transformation matrix, For the Nodes at time The neighbor index set of For the The feature vector of each node, For the The feature vector of each node, is the set of neighbor nodes The traversal subscript in is used to normalize the denominator accumulation of softmax, Represents vector concatenation operation; Then the neighbor features are weighted and aggregated to obtain the local fusion result: in, For the The local fusion feature vector of the nodes after weighted aggregation.

6. The real-time path optimization system for UAVs based on edge computing according to claim 1, characterized in that: During the flight of the drone, it continuously receives new perception states and judges in real time whether there are sudden changes in the environment. When new obstacles are detected, new state perception, sampling, tree expansion and path optimization processes are triggered to dynamically correct the flight path.

Images