Unmanned aerial vehicle area boundary automatic identification and obstacle avoidance method and device

Abstract

The invention relates to an unmanned aerial vehicle area boundary automatic identification and obstacle avoidance method and device, and the method comprises the steps: obtaining environment perception data of an unmanned aerial vehicle sensor, carrying out the type division of the environment perception data, and obtaining corresponding image data, point cloud data and continuous image frames; performing boundary recognition processing on the image data to obtain a corresponding boundary recognition result, and performing obstacle feature extraction on the point cloud data to obtain corresponding obstacle features; performing long-range dependency feature prediction on the continuous image frames through a preset Transform model to obtain corresponding motion prediction information; performing obstacle avoidance path planning on the three-dimensional environment map according to the dynamic boundary information to obtain a corresponding obstacle avoidance path; and performing decision analysis on the obstacle avoidance path, the dynamic boundary information and a preset task target to generate an automatic obstacle avoidance decision. According to the invention, environment information can be updated in real time and a corresponding obstacle avoidance decision can be made.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicles, and in particular to a method and a device for automatically identifying and avoiding obstacles of an unmanned aerial vehicle region boundary. Background Art

[0002] With the rapid development of drone technology, its application in multiple fields is becoming increasingly widespread. However, how to ensure that drones can perform tasks safely and efficiently in complex environments has become one of the current research focuses. In particular, in terms of regional boundary recognition and obstacle avoidance, existing technologies still have some limitations. Traditional drone navigation and obstacle avoidance systems usually rely on single sensor data or simplified environmental models, which may lead to inaccurate recognition or untimely response in complex and changing actual environments. In addition, existing obstacle avoidance methods often lack the ability to fully perceive the environment and track dynamic changes in real time, which may cause drones to face potential safety risks when performing tasks. Especially in rapidly changing environments, how to update environmental information in a timely manner and make corresponding obstacle avoidance decisions has become an urgent problem to be solved. Summary of the invention

[0003] The main purpose of the present invention is to provide a method and device for automatic identification and obstacle avoidance of UAV area boundaries, which can update environmental information in real time and make corresponding obstacle avoidance decisions.

[0004] To achieve the above object, the present invention provides a method for automatically identifying and avoiding obstacles in a drone area, comprising: Acquire environmental perception data from the drone sensor, classify the environmental perception data into types, and obtain corresponding image data, point cloud data, and continuous image frames; Performing boundary recognition processing on the image data to obtain corresponding boundary recognition results, and performing obstacle feature extraction on the point cloud data to obtain corresponding obstacle features; Predicting the long-range dependent features of the continuous image frames through a preset Transformer model to obtain corresponding motion prediction information; Performing three-dimensional construction based on the boundary recognition result, the obstacle characteristics and the motion prediction information to obtain a three-dimensional environment map; Perform boundary tracking based on the three-dimensional environment map to obtain corresponding dynamic boundary information, and perform obstacle avoidance path planning on the three-dimensional environment map according to the dynamic boundary information to obtain a corresponding obstacle avoidance path; A decision analysis is performed on the obstacle avoidance path, the dynamic boundary information and the preset task goal to generate an automatic obstacle avoidance decision.

[0005] Furthermore, the environmental perception data of the drone sensor is obtained, and the environmental perception data is classified into types to obtain corresponding image data, point cloud data and continuous image frames, including: Acquire the original environmental data collected by the drone sensor, the original environmental data including single-frame image data, multiple-frame continuous image data, original infrared image data, original laser radar data and original ultrasonic data, Performing time stamp alignment and image fusion processing on the single-frame image data and the original infrared image data to obtain the fused image data; Performing Euclidean clustering fusion on the original laser radar data and the original ultrasonic data to obtain the structured point cloud data; The multiple frames of continuous image data are arranged in order of timestamps, and an image sequence is constructed to obtain the corresponding continuous image frames.

[0006] Further, the performing boundary recognition processing on the image data to obtain a corresponding boundary recognition result, and performing obstacle feature extraction on the point cloud data to obtain a corresponding obstacle feature, includes: Performing edge feature detection on the image data to obtain corresponding edge feature information; Performing image segmentation on the image data according to the edge feature information to obtain a regional image; Performing contour analysis on the regional image according to the edge feature information to obtain boundary contour information; Performing boundary fitting and recognition on the image data according to the boundary contour information and the edge feature information to obtain the boundary recognition result; Performing vector extraction on the point cloud data to obtain an obstacle vector; Feature classification is performed on the obstacle vector to obtain the obstacle feature.

[0007] Furthermore, the performing long-range dependency feature prediction on the continuous image frames through a preset Transformer model to obtain corresponding motion prediction information includes: Extracting features from the continuous image frames through the input layer of the Transformer model to obtain an image feature sequence; Performing sinusoidal position coding on the image feature sequence to obtain feature sequence coding; Input the feature sequence encoding into a Transformer encoder, perform self-attention calculation through a multi-head self-attention mechanism layer of the Transformer encoder, and obtain a self-attention output; The self-attention output is encoded and transformed by an encoding feedforward neural network layer, and input into an encoding residual connection layer for stacking processing to obtain an encoding feature representation; Inputting the encoded feature representation into the global self-attention mechanism layer to calculate the correlation between different time steps to obtain long-range dependency features; Inputting the long-range dependency feature into a Transformer decoder, performing cross-attention calculation on the long-range dependency feature and the encoding feature representation through a multi-head cross-attention mechanism layer of the Transformer decoder, and obtaining a cross-attention output; The cross attention output is encoded and transformed by a decoding feedforward neural network layer, and input into an encoding residual connection layer for decoding processing to obtain a decoding feature sequence; Performing time series prediction on the decoded feature sequence to obtain a predicted feature sequence; Inputting the predicted feature sequence into the fully connected layer to perform motion state prediction to obtain a preliminary motion prediction result; Performing uncertainty analysis on the preliminary motion prediction result to obtain prediction confidence; The prediction confidence and the preliminary motion prediction result are input into the output layer of the Transformer model for prediction correction, and the motion prediction information is output.

[0008] Furthermore, the three-dimensional construction is performed based on the boundary recognition result, the obstacle characteristics and the motion prediction information to obtain a three-dimensional environment map, including: Extracting boundary points from the boundary recognition result to obtain a boundary point set; Delineating the boundary according to the boundary point set to obtain a corresponding boundary curve; Performing three-dimensional analysis on the obstacle features to obtain three-dimensional obstacle data; Performing three-dimensional modeling based on the obstacle characteristics and the three-dimensional obstacle data to obtain a three-dimensional model of the obstacle; Performing coordinate transformation on the motion prediction information to obtain a motion trajectory in a global coordinate system; An initial three-dimensional environment map is constructed according to the boundary curve, the obstacle three-dimensional model and the motion trajectory, wherein the process of constructing the initial three-dimensional environment map includes: Generate a boundary plane mesh according to the boundary curve to obtain a boundary mesh model; Calculating the position and posture information of the UAV at each time point according to the motion trajectory; Projecting the boundary grid model and the obstacle three-dimensional model into the same coordinate system according to the posture information, and performing spatial registration to obtain an aligned environment model; Triangulating the environment model to obtain a triangular mesh model; Reconstructing the surface of the triangular mesh model to obtain an initial environment map; The initial environment map is subjected to three-dimensional gridding processing to obtain the three-dimensional environment map.

[0009] Further, the performing boundary tracking based on the three-dimensional environment map to obtain corresponding dynamic boundary information, and performing obstacle avoidance path planning on the three-dimensional environment map according to the dynamic boundary information to obtain a corresponding obstacle avoidance path, includes: Performing dynamic boundary calculation on the three-dimensional environment map to obtain a corresponding dynamic area edge contour; Calculating the curvature of contour points on the edge contour of the dynamic area to obtain corresponding dynamic contour points; Performing boundary analysis based on the dynamic contour points combined with the dynamic area edge contour to obtain the dynamic boundary information; Performing obstacle analysis on the three-dimensional environment map based on the dynamic boundary information to obtain corresponding dynamic obstacle information, and constructing a time-varying constraint field to obtain corresponding dynamic obstacle avoidance constraint conditions; Performing path planning on the three-dimensional environment map according to the dynamic obstacle avoidance constraint condition to obtain an initial path; The initial path is optimized and verified according to the dynamic obstacle information to obtain the obstacle avoidance path.

[0010] Furthermore, the performing decision analysis on the obstacle avoidance path, the dynamic boundary information and the preset task goal to generate an automatic obstacle avoidance decision includes: Splitting the obstacle avoidance path into sections to obtain a plurality of path sub-segments; Performing risk assessment on each of the path sub-segments to obtain a corresponding risk coefficient; Prioritizing the path sub-segments according to the risk coefficients to obtain an initial ranking result; Adjusting the initial sorting result according to the dynamic boundary information to obtain an optimized sorting result; Decomposing the preset task goal to obtain multiple sub-task goals; Associating and matching the optimization sorting results with the subtask objectives to obtain multiple task path association combinations; Performing feasibility analysis on each of the task path association combinations to obtain a corresponding feasibility score; The feasibility scores are screened according to a preset feasibility screening rule to obtain a feasibility score set; Selecting the corresponding task path association combination according to the feasibility score set to perform decision analysis to obtain an initial obstacle avoidance decision; Performing simulation on the initial obstacle avoidance decision to obtain a simulation result; The initial obstacle avoidance decision is optimized according to the simulation result to obtain the automatic obstacle avoidance decision.

[0011] The present invention also provides a device for automatically identifying and avoiding obstacles in a drone area boundary, which is applied to any of the above-mentioned methods for automatically identifying and avoiding obstacles in a drone area boundary, comprising: An acquisition module, which is used to obtain environmental perception data from drone sensors, classify the environmental perception data into types, and obtain corresponding image data, point cloud data, and continuous image frames; An analysis module, the analysis module is used to perform boundary recognition processing on the image data to obtain corresponding boundary recognition results, and to extract obstacle features from the point cloud data to obtain corresponding obstacle features; A prediction module, the prediction module is used to perform long-range dependency feature prediction on the continuous image frames through a preset Transformer model to obtain corresponding motion prediction information; An association module, the association module is used to perform three-dimensional construction based on the boundary recognition result, the obstacle characteristics and the motion prediction information to obtain a three-dimensional environment map; A processing module, the processing module is used to perform boundary tracking based on the three-dimensional environment map to obtain corresponding dynamic boundary information, and perform obstacle avoidance path planning on the three-dimensional environment map according to the dynamic boundary information to obtain a corresponding obstacle avoidance path; A control module is used to perform decision analysis on the obstacle avoidance path, the dynamic boundary information and the preset task objectives to generate an automatic obstacle avoidance decision.

[0012] The present invention provides a method and device for automatically identifying the boundary of a drone area and avoiding obstacles, which has the following beneficial effects: By comprehensively utilizing image data, point cloud data, and continuous image frames, multi-dimensional perception of the environment is achieved, the accuracy of boundary recognition and obstacle detection is improved, and the limitations that may be caused by a single data source are overcome. By fusing and processing multi-source data and constructing three-dimensional objects, a more accurate and comprehensive three-dimensional environmental map is generated, providing reliable basic information for the navigation and obstacle avoidance of the UAV. The real-time tracking and updating mechanism based on dynamic boundary information enables the system to respond to environmental changes in a timely manner, improving the adaptability of the UAV in complex and rapidly changing environments. By combining dynamic boundary information and three-dimensional environmental maps for obstacle avoidance path planning, a more intelligent and flexible obstacle avoidance strategy is achieved, improving the safety of the UAV and the efficiency of mission execution. By combining obstacle avoidance paths, dynamic boundary information, and preset mission goals, automatic obstacle avoidance decisions are generated, realizing the autonomous decision-making ability of the UAV during the mission execution process, and improving the overall intelligence level of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a flow chart of a method for automatic identification and obstacle avoidance of a drone area boundary provided by the present invention; Figure 2 This is a structural diagram of an automatic identification and obstacle avoidance system for a drone area boundary provided by the present invention; Figure 3 This is a structural diagram of an automatic identification and obstacle avoidance device for a drone area boundary provided by the present invention.

[0014] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with 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.

[0016] The present invention is further described below in conjunction with the accompanying drawings and specific implementation methods.

[0017] Reference Figure 1 As shown, the present invention provides a method for automatically identifying and avoiding obstacles in a drone area boundary, comprising: Step S1: Acquire environmental perception data from the drone sensor, classify the environmental perception data into types, and obtain corresponding image data, point cloud data, and continuous image frames; Step S2: performing boundary recognition processing on the image data to obtain the corresponding boundary recognition result, and performing obstacle feature extraction on the point cloud data to obtain the corresponding obstacle feature; Step S3: Predicting the long-range dependent features of continuous image frames through a preset Transformer model to obtain corresponding motion prediction information; Step S4: Perform three-dimensional construction based on boundary recognition results, obstacle characteristics and motion prediction information to obtain a three-dimensional environment map; Step S5: performing boundary tracking based on the three-dimensional environment map to obtain corresponding dynamic boundary information, and performing obstacle avoidance path planning on the three-dimensional environment map according to the dynamic boundary information to obtain a corresponding obstacle avoidance path; Step S6: Perform decision analysis on the obstacle avoidance path, dynamic boundary information and preset mission objectives to generate an automatic obstacle avoidance decision.

[0018] Based on the above steps, the detailed process is as follows: Step S1: Start all sensors on the drone, make sure they are working properly and start collecting data. Collect environmental perception data from each sensor of the drone in real time. Separate the collected environmental perception data into image data, point cloud data and continuous image frames through data format or metadata.

[0019] Image data: Two-dimensional image data collected by a camera, usually RGB images or grayscale images. Point cloud data: Three-dimensional spatial data collected by a lidar or depth camera, representing objects in the environment and their distances. Continuous image frames: A series of image frames taken continuously by a camera in a short period of time to capture dynamically changing information.

[0020] Step S2: read the image data, convert it into grayscale, and apply Gaussian filtering to remove noise in the image.

[0021] Use edge detection algorithms (such as Canny) to detect edges in the image, and use morphological operations (such as dilation and erosion) to refine the edges to obtain clear boundary recognition results. Store the processed boundary recognition results as binary images.

[0022] Read point cloud data and perform voxel downsampling. Use statistical filtering and radius filtering to remove noise points in the point cloud.

[0023] The point cloud data is segmented by clustering algorithm to identify different obstacles. The RANSAC algorithm is used to fit the geometric shape of the obstacle and extract its location, size and other features.

[0024] The extracted obstacle features are labeled and stored as structured data.

[0025] Step S3: normalize the continuous image frames, extract motion information using inter-frame difference technology, and generate a difference image.

[0026] Perform feature extraction on each frame of image to obtain a high-level feature vector.

[0027] Initialize the parameters of the Transformer model, including the number of encoder and decoder layers, the number of attention heads, etc.

[0028] The extracted feature vector is used as a time series input and fed into the encoder of the Transformer model.

[0029] The self-attention mechanism of the Transformer model is used to capture long-range dependencies between image frames.

[0030] The decoder predicts future motion information based on the output of the encoder and obtains motion prediction information.

[0031] The predicted future position and speed of the UAV and other motion information are stored, the prediction results are analyzed for uncertainty, and the analysis results are saved for subsequent use.

[0032] Step S4: Convert the boundary recognition results, obstacle features and motion prediction information into a unified coordinate system.

[0033] Select a suitable 3D reconstruction algorithm, such as Marching Cubes or TSDF (Truncated Signed Distance Function). Initialize the parameters required by the 3D reconstruction algorithm.

[0034] Using a 3D reconstruction algorithm, the boundary recognition results and obstacle features are integrated into the 3D space, a preliminary 3D environment model is constructed, texture mapping is performed on the model, and motion prediction information is mapped onto the 3D model surface.

[0035] The initially constructed 3D model is subjected to noise filtering to remove redundant and erroneous data, and an optimization algorithm (such as BA or GTSAM) is applied to globally optimize the 3D model to ensure the accuracy and consistency of the model.

[0036] The optimized 3D environment map is stored in an appropriate file format (such as PLY, OBJ or PCD).

[0037] The 3D environment map is validated and if errors or inconsistencies are found, it is returned to the data preprocessing step for adjustment and reconstruction.

[0038] Step S5: In the three-dimensional environment map, a boundary detection algorithm (such as K-means or Mean Shift) is used to detect dynamic boundaries.

[0039] Update boundary information based on the real-time position and motion prediction information of the drone.

[0040] Select an appropriate path planning algorithm, such as A*, Dijkstra, or RRT, and initialize the parameters required by the path planning algorithm.

[0041] The cost function is designed by considering factors such as path length, obstacle distance and boundary distance, and the cost function is integrated into the path planning algorithm.

[0042] According to the dynamic boundary information and cost function, a path search is performed in the three-dimensional environment map to find the optimal obstacle avoidance path, and the initially planned path is verified to ensure that the path avoids all obstacles and boundaries.

[0043] The initially planned path is smoothed to eliminate sharp corners and discontinuities in the path, and the path is further optimized through optimization algorithms (such as B-spline or Bezier curve) to obtain the corresponding obstacle avoidance path.

[0044] Step S6: Read the preset mission objectives of the drone, including the target location, path, and mission priority, etc. Combine the obstacle avoidance path, dynamic boundary information, and mission objectives to perform multi-objective optimization analysis.

[0045] Design an optimization objective function, which involves factors such as path safety, task completion and time efficiency.

[0046] Assign appropriate weights to different decision factors and introduce weight parameters into the optimization objective function to guide the optimization process.

[0047] Select a suitable decision analysis method to perform decision analysis on the obstacle avoidance path, dynamic boundary information and preset mission objectives to generate automatic obstacle avoidance decisions.

[0048] The generated obstacle avoidance decision is verified, the final obstacle avoidance decision is output to the UAV control system, and the timestamp and related information of the decision generation are recorded.

[0049] The invention provides a method for automatic identification and obstacle avoidance of the regional boundary of a UAV, which realizes multi-dimensional perception of the environment by comprehensively utilizing image data, point cloud data and continuous image frames, improves the accuracy of boundary identification and obstacle detection, and overcomes the limitations that may be brought by a single data source. By fusing and three-dimensionally constructing multi-source data, a more accurate and comprehensive three-dimensional environmental map is generated, which provides reliable basic information for the navigation and obstacle avoidance of the UAV. Based on the real-time tracking and updating mechanism of dynamic boundary information, the system can respond to environmental changes in a timely manner, and improves the adaptability of the UAV in complex and rapidly changing environments. By combining dynamic boundary information and three-dimensional environmental maps for obstacle avoidance path planning, a more intelligent and flexible obstacle avoidance strategy is realized, which improves the safety and task execution efficiency of the UAV. By combining obstacle avoidance paths, dynamic boundary information and preset task targets, automatic obstacle avoidance decisions are generated, which realizes the autonomous decision-making ability of the UAV in the process of executing tasks and improves the overall intelligence level of the system.

[0050] In one embodiment, environmental perception data of a drone sensor is obtained, and the environmental perception data is classified into types to obtain corresponding image data, point cloud data, and continuous image frames, including: Drones are equipped with a variety of sensors, including visible light cameras, infrared cameras, lidar and ultrasonic sensors, which work simultaneously to collect raw data about the surrounding environment.

[0051] The original environmental data collected by each sensor is obtained. The original environmental data includes: single-frame image data and multiple-frame continuous image data collected by the visible light camera, original infrared image data collected by the infrared camera, original lidar data collected by the lidar, and original ultrasonic data collected by the ultrasonic sensor.

[0052] For single-frame image data and raw infrared image data, timestamp alignment and image fusion processing are performed. Timestamp alignment is to match the visible light image and infrared image collected at the same time. Image fusion processing uses a multi-scale fusion algorithm to combine the detail information of the visible light image with the thermal radiation information of the infrared image to generate fused image data containing more environmental information.

[0053] For the original lidar data and the original ultrasonic data, the Euclidean clustering fusion method is used for processing. Euclidean clustering is a commonly used point cloud segmentation algorithm, which classifies similar points into one category based on the Euclidean distance between the points.

[0054] The LiDAR point cloud data is fused with the ultrasonic data, and the contours of objects in the environment are identified through clustering methods to form structured point cloud data.

[0055] For multiple frames of continuous image data, they are arranged in timestamp order to construct an image sequence. This step generates continuous image frames, which provides a basis for subsequent motion estimation and obstacle tracking. Continuous image frames contain information about the environment changing over time, which helps the system understand dynamic scenes.

[0056] Through the above processing, the original environment data is converted into three types of structured data: fused image data, structured point cloud data and continuous image frames.

[0057] This embodiment significantly improves the UAV's ability to perceive the environment through multi-sensor data fusion and structured processing. The fusion of visible light and infrared images enhances the system's adaptability under different lighting conditions and improves the ability to capture environmental details and thermal radiation information. The Euclidean clustering fusion of lidar and ultrasonic data combines the advantages of long-distance precise measurement and close-range detection to generate more accurate three-dimensional environmental structure information. The time series processing of multiple frames of continuous images provides a basis for dynamic scene understanding. This comprehensive use of multi-source data greatly improves the comprehensiveness and accuracy of environmental perception, and provides a rich and reliable information basis for subsequent regional boundary identification and obstacle avoidance decisions. At the same time, the structured processing of data simplifies the complexity of subsequent algorithms and improves the real-time and efficiency of the system. This method not only enhances the adaptability of UAVs in complex environments, but also lays the foundation for more accurate navigation and safer flight.

[0058] In one embodiment, the image data is processed for boundary recognition to obtain corresponding boundary recognition results, and the point cloud data is processed for obstacle feature extraction to obtain corresponding obstacle features, including: Perform edge feature detection on image data to obtain corresponding edge feature information. Edge feature detection uses the Canny operator to detect edge pixels in the image by calculating the gradient amplitude and direction of the image to form edge feature information. Edge feature information includes the position, intensity and direction of edge pixels.

[0059] Image data is segmented based on edge feature information to obtain a regional image. Image segmentation uses the region growing method, taking edge pixels in edge feature information as seed points and expanding to the surrounding area until the preset threshold condition is reached, dividing the image into multiple regions to form a regional image. The preset threshold conditions include pixel gray value difference, regional area, etc.

[0060] The regional image is analyzed according to the edge feature information to obtain the boundary contour information. The contour analysis uses the contour tracking algorithm to track the boundary pixels of each region in the regional image, extract the contour line of the region, and form the boundary contour information. The boundary contour information includes the coordinate sequence, length, curvature and other features of the contour line.

[0061] The boundary contour information is subjected to curve fitting processing to obtain a preliminary boundary curve, and the preliminary boundary curve is subjected to discretization sampling to obtain a boundary sampling point set.

[0062] The edge feature information is weighted to obtain the edge feature weight, and the boundary sampling point set is weighted according to the edge feature weight to obtain the weighted boundary point set.

[0063] The weighted boundary point set is fitted by least squares to obtain the optimized boundary curve, and the optimized boundary curve is subjected to curvature analysis to obtain the boundary curvature characteristics.

[0064] The optimized boundary curve is segmented according to the boundary curvature characteristics to obtain the segmented boundary curve, and the segmented boundary curve is parameterized to obtain the boundary parameter equation.

[0065] The image data is marked with boundaries according to the boundary parameter equation to obtain a marked boundary image, and the marked boundary image is morphologically processed to obtain boundary morphological features.

[0066] The labeled boundary image is classified according to the boundary morphological features to obtain boundary type information; the boundary parameter equation and boundary type information are combined to obtain the boundary recognition result. The boundary recognition result includes information such as the location, shape and type of the boundary.

[0067] Perform vector extraction on the point cloud data to obtain the obstacle vector. The vector extraction uses the principal component analysis method to calculate the covariance matrix of the local point set in the point cloud data and extract the main direction vector as the obstacle vector. The obstacle vector includes information such as direction, length and starting point.

[0068] Obstacle vectors are feature classified to obtain obstacle features. Feature classification uses the support vector machine algorithm to classify obstacles based on the direction and length of obstacle vectors, and obtains feature information such as obstacle type, size, and position to form obstacle features.

[0069] This embodiment achieves accurate identification of the flight area boundary by performing multi-level processing on the image data, including edge feature detection, image segmentation, contour analysis and boundary fitting identification, and improves the accuracy and robustness of boundary identification. The edge and region information in the image is effectively extracted by using advanced algorithms such as the Canny operator and the region growing method, laying the foundation for subsequent boundary identification. Vector extraction and feature classification are performed on the point cloud data to achieve effective identification and feature extraction of obstacles. By extracting obstacle vectors using the principal component analysis method and combining feature classification with the support vector machine algorithm, key information such as the type, size and position of obstacles can be accurately identified, providing reliable environmental perception capabilities for the obstacle avoidance control of the drone. By combining the processing results of image data and point cloud data, this method achieves comprehensive perception of the flight area boundary and obstacles, provides rich environmental information for the autonomous navigation and obstacle avoidance decision-making of the drone, and significantly improves the flight safety and mission execution capabilities of the drone in complex environments.

[0070] In one embodiment, a preset Transformer model is used to predict long-range dependent features of continuous image frames to obtain corresponding motion prediction information, including: The Transformer model in this embodiment is a deep learning model based on the attention mechanism, which is used to predict long-range dependent features of consecutive image frames. The model includes main layers such as input layer, encoder, decoder and output layer.

[0071] The continuous image frames are input to the input layer of the Transformer model for processing. The input layer extracts features from the image through methods such as convolutional neural networks, obtains feature vectors that represent the visual features of each frame of the image, and forms an image feature sequence.

[0072] Perform dimensional analysis on the image feature sequence to obtain feature dimension information, generate a position index sequence based on the feature dimension information, and obtain the original position index.

[0073] The original position index is standardized to obtain a standard position index, and the sine function value of the standard position index is calculated according to the sine and cosine functions to obtain a sine position coding matrix.

[0074] The sine position coding matrix is ​​dimensionally expanded to obtain the extended position coding, and the extended position coding is element-wise added to the image feature sequence to obtain the preliminary feature sequence coding. The preliminary feature sequence coding is layer-normalized to obtain the normalized feature sequence coding.

[0075] The normalized feature sequence encoding is input into the residual connection layer to obtain the residual feature sequence, and the residual feature sequence is processed by the activation function to obtain the activated feature sequence. The activated feature sequence is input into the feedforward neural network layer to obtain the final feature sequence encoding.

[0076] The purpose of position encoding is to add position information to each element in the sequence, so that the model can distinguish features at different time steps. Sine position encoding uses sine and cosine functions to generate waveforms of different frequencies to encode position information, resulting in a sinusoidal position encoding matrix.

[0077] The position-encoded feature sequence is input into the Transformer encoder. The encoder contains multiple stacked self-attention layers and feed-forward neural network layers. In the self-attention layer, the model calculates the correlation between each element in the sequence and all other elements, thereby capturing the long-range dependencies within the sequence. The calculation of self-attention is based on three vectors: query, key, and value. The attention output is obtained by calculating the similarity between the query and the key and weighted summing it with the value.

[0078] The output of the self-attention layer is transformed nonlinearly through the feedforward neural network layer to further extract features. The feedforward network layer consists of two linear transformations and a nonlinear activation function. The transformed features are added to the original input through residual connections to form an encoded feature representation. Residual connections help alleviate the gradient vanishing problem of deep networks.

[0079] The encoded feature representation is input into the global self-attention mechanism layer to calculate the correlation between features at different time steps. The global self-attention layer can capture dependencies over a longer time span and obtain feature representations with long-range dependency information.

[0080] The Transformer decoder receives the output of the encoder as input. The decoder also contains multiple stacked attention layers and feed-forward network layers. The criss-cross attention layer in the decoder calculates the attention between the decoded sequence and the encoded sequence, incorporating the encoder's information into the decoding process. The criss-cross attention is calculated in a similar way to the self-attention, but the query comes from the decoded sequence, and the key and value come from the encoded sequence.

[0081] The output of the decoder passes through the feedforward network layer and the residual connection to obtain a decoded feature sequence. The decoded feature sequence is time-series predicted to predict the features of several future time steps. The prediction method can use a recurrent neural network for prediction.

[0082] The predicted feature sequence is input into the fully connected layer for motion state prediction. The fully connected layer maps the features to specific motion parameter spaces, such as position, speed, acceleration, etc., to obtain preliminary motion prediction results.

[0083] Conduct uncertainty analysis on the preliminary prediction results to evaluate the reliability of the prediction. Uncertainty analysis can be based on methods such as Bayesian neural networks or ensemble learning to output confidence intervals or probability distributions of the prediction results.

[0084] The prediction confidence and preliminary prediction results are input into the output layer of the Transformer model. The output layer modifies the prediction results according to the confidence to obtain the final motion prediction information. The modification process may include smoothing or discarding low-confidence predictions.

[0085] This embodiment uses the Transformer model to predict long-range dependent features of continuous image frames, which can effectively capture the spatiotemporal correlation in the image sequence and achieve accurate prediction of the UAV's motion state. The introduction of the multi-head attention mechanism and the global self-attention layer enables the model to capture dependencies at different time scales and improve the robustness of the prediction. The application of the cross-attention mechanism realizes information fusion between the encoder and the decoder, and enhances the feature extraction capability of the model. By introducing uncertainty analysis and prediction correction, the reliability and accuracy of the prediction results are improved. This method does not require manual feature design, can adaptively learn motion patterns in different scenarios, and has strong generalization capabilities.

[0086] In one embodiment, three-dimensional construction is performed based on boundary recognition results, obstacle features, and motion prediction information to obtain a three-dimensional environment map, including: Boundary point extraction is performed on the boundary recognition results to obtain a boundary point set. Boundary point extraction refers to extracting a set of discrete points representing the boundary from the boundary recognition results. These boundary points are usually key points or feature points on the boundary contour. The boundary point set is a set of multiple two-dimensional or three-dimensional coordinate points, each of which represents a position on the boundary.

[0087] Based on the obtained boundary point set, the boundary is delineated to obtain the corresponding boundary curve. Boundary delineation refers to fitting a continuous boundary curve using the boundary point set. This boundary curve more accurately represents the shape and contour of the actual boundary. Common boundary delineation methods include polynomial fitting, spline interpolation, etc.

[0088] Obstacle features are analyzed in three dimensions to obtain obstacle 3D data. Obstacle features refer to information such as the shape, size, and position of obstacles. Three-dimensional analysis refers to the use of sensor data or image processing technology to extract the 3D geometric information of obstacles. Obstacle 3D data usually includes information such as the 3D coordinates and size of obstacles.

[0089] The obstacle 3D model is obtained by performing 3D modeling based on the obstacle characteristics and obstacle 3D data. 3D modeling refers to constructing a digital 3D representation of the obstacle using the obstacle's geometric information. The obstacle 3D model can be in the form of a point cloud model, a mesh model or a parametric model.

[0090] The motion prediction information is transformed to obtain the motion trajectory in the global coordinate system. The motion prediction information refers to the predicted future motion state of the UAV. Coordinate transformation refers to converting the motion prediction information in the local coordinate system to the global coordinate system. The motion trajectory in the global coordinate system represents the predicted motion path of the UAV in the entire environment.

[0091] The initial 3D environment map is constructed based on the boundary curve, obstacle 3D model and motion trajectory. The process of constructing the initial 3D environment map includes the following steps: The boundary plane mesh is generated according to the boundary curve to obtain the boundary mesh model. The boundary plane mesh generation refers to discretizing the continuous boundary curve into a plane mesh composed of multiple mesh units. The boundary mesh model is a discretized representation of the boundary, which is convenient for subsequent processing and calculation.

[0092] The position and attitude information of the drone at each time point is calculated based on the motion trajectory. The position and attitude information includes the position and attitude of the drone, which is usually represented by three-dimensional coordinates and Euler angles or quaternions.

[0093] Based on the pose information, the boundary mesh model and the obstacle 3D model are projected into the same coordinate system and spatially registered to obtain the aligned environment model. Spatial registration refers to aligning data from different sources into the same coordinate system to ensure spatial consistency between the data. The aligned environment model contains comprehensive information about boundaries and obstacles.

[0094] The environment model is triangulated to obtain a triangular mesh model. Triangulation is the process of decomposing complex geometric shapes into a mesh structure composed of multiple triangles. The triangular mesh model is a discretized representation of the environment, which facilitates subsequent surface reconstruction and rendering.

[0095] The triangular mesh model is reconstructed to obtain the initial environment map. Surface reconstruction refers to reconstructing a continuous surface model based on a discrete triangular mesh model. The initial environment map is a rough three-dimensional representation of the environment, which contains basic geometric information of boundaries and obstacles.

[0096] The initial environment map is processed into a three-dimensional grid to obtain a three-dimensional environment map. Three-dimensional grid processing refers to discretizing a continuous surface model into a three-dimensional grid structure composed of multiple voxels. The three-dimensional environment map is a refined three-dimensional representation of the environment and can be used for subsequent path planning and obstacle avoidance decisions.

[0097] This embodiment achieves high-precision representation of regional boundaries by accurately extracting points and drawing curves from boundary recognition results, providing a reliable basis for boundary recognition of drones. Combined with three-dimensional analysis and modeling of obstacle features and coordinate transformation of motion prediction information, a comprehensive three-dimensional environmental map containing boundaries, obstacles and motion trajectories is constructed. This method not only improves the accuracy of environmental perception, but also enhances the adaptability to dynamic scenes. Through a multi-step three-dimensional construction process, including boundary mesh generation, spatial alignment, triangulation and surface reconstruction, the conversion from discrete data to continuous models is achieved, improving the integrity and precision of environmental representation. The final three-dimensional gridding process optimizes the structure of the environmental map, provides high-quality data support for subsequent path planning and obstacle avoidance decisions, and significantly improves the autonomous navigation and obstacle avoidance capabilities of drones in complex environments.

[0098] In one embodiment, boundary tracking is performed based on the three-dimensional environment map to obtain corresponding dynamic boundary information, and obstacle avoidance path planning is performed on the three-dimensional environment map according to the dynamic boundary information to obtain a corresponding obstacle avoidance path, including: After obtaining the 3D environment map, the dynamic boundary calculation is performed on the 3D environment map to obtain the corresponding dynamic area edge contour. The dynamic boundary calculation uses edge detection algorithms, such as Canny edge detection or Sobel operator, to extract the edge contours of objects in the environment. The dynamic area edge contour refers to the contours of dynamically changing objects in the environment, such as moving vehicles, pedestrians, etc.

[0099] The contour point curvature is calculated for the edge contour of the dynamic area to obtain the corresponding dynamic contour point. The contour point curvature calculation adopts the differential geometry method to calculate the curvature value of each point on the contour line. The dynamic contour point refers to the contour point with a larger curvature value, which represents the inflection point or sharp corner of the contour.

[0100] Based on the dynamic contour points combined with the edge contour of the dynamic area, the boundary analysis is performed to obtain the dynamic boundary information. The boundary analysis combines adjacent dynamic contour points into boundary segments through a clustering algorithm to obtain a complete dynamic boundary. The dynamic boundary information includes the location, shape, and movement direction of the boundary.

[0101] Obstacle analysis is performed on the three-dimensional environment map based on the dynamic boundary information to obtain the corresponding dynamic obstacle information, and a time-varying constraint field is constructed to obtain the corresponding dynamic obstacle avoidance constraint conditions. Obstacle analysis obtains the two-dimensional outline of the obstacle by projecting the dynamic boundary onto the map plane. The time-varying constraint field is constructed by predicting the movement of the obstacle to represent the no-fly zone in the future. The dynamic obstacle avoidance constraint conditions include the scope and duration of the no-fly zone.

[0102] According to the dynamic obstacle avoidance constraints, the path planning is performed on the three-dimensional environment map to obtain the initial path. The path planning uses the A* algorithm to search for the optimal path in the three-dimensional space while satisfying the dynamic obstacle avoidance constraints. The initial path is the shortest path that satisfies the constraints.

[0103] The initial path is optimized and verified based on the dynamic obstacle information to obtain the obstacle avoidance path. The optimization verification simulates the UAV flight process to check whether it collides with dynamic obstacles and makes local adjustments to the path if necessary. The obstacle avoidance path is the final safe flight path.

[0104] This embodiment can accurately identify the dynamic boundary information in the environment by performing dynamic boundary calculation and contour point curvature analysis on the three-dimensional environment map, providing a reliable data basis for drone obstacle avoidance. By constructing a time-varying constraint field based on dynamic boundary information, effective prediction and constraint of dynamic obstacles are achieved, greatly improving the real-time and accuracy of obstacle avoidance. The A* algorithm is used in combination with dynamic obstacle avoidance constraints for path planning, which not only ensures the optimality of the path, but also meets the dynamic obstacle avoidance requirements. By optimizing and verifying the initial path, the safety and reliability of the obstacle avoidance path are further improved. The entire method realizes adaptive obstacle avoidance in complex dynamic environments, significantly improves the autonomous flight capability of drones in changing environments, and lays the foundation for the application of drones in complex scenes such as cities.

[0105] In one embodiment, a decision analysis is performed on the obstacle avoidance path, dynamic boundary information, and preset mission objectives to generate an automatic obstacle avoidance decision, including: The obstacle avoidance path is segmented and the overall obstacle avoidance path is divided into multiple shorter path sub-segments. The path sub-segment division can be based on factors such as preset distance thresholds or terrain features. Risk assessment is performed on each path sub-segment, taking into account factors such as terrain complexity and obstacle density, and the corresponding risk coefficient is calculated. The risk coefficient reflects the degree of danger of the path sub-segment.

[0106] The path sub-segments are prioritized according to the risk coefficient. The path sub-segments with lower risk coefficients have higher priorities, and the initial sorting results are obtained. Subsequently, the initial sorting results are adjusted according to the dynamic boundary information to obtain the optimized sorting results. The dynamic boundary information includes dynamic factors that may affect flight safety, such as real-time meteorological data and temporary no-fly zones.

[0107] Decompose the preset task goals and split the overall task goals into multiple specific executable subtask goals. Associate and match the optimization sorting results with the subtask goals to obtain multiple task path association combinations. Each task path association combination contains an optimized sorted path and the corresponding subtask goal.

[0108] Perform feasibility analysis on each mission path association combination, consider factors such as drone performance parameters and mission time limits, and calculate the corresponding feasibility score. The feasibility score reflects the execution difficulty and success probability of the combination. Filter the feasibility scores according to the preset feasibility screening rules to obtain a feasibility score set. The feasibility screening rules include conditions such as the minimum feasibility score threshold.

[0109] According to the feasibility score set, the corresponding task path association combination is selected for decision analysis, and the initial obstacle avoidance decision is obtained by comprehensively considering factors such as path safety and task completion. The initial obstacle avoidance decision is simulated to simulate the flight trajectory and task execution of the drone under the decision to obtain the simulation results. According to the simulation results, the initial obstacle avoidance decision is optimized, the flight parameters or path planning are adjusted, and finally the automatic obstacle avoidance decision is obtained.

[0110] This embodiment realizes a refined analysis of complex environments and improves the accuracy and safety of path planning by segmenting and risk-assessing the obstacle avoidance path and optimizing and sorting it in combination with dynamic boundary information. The task objectives are decomposed and matched with the path association so that the obstacle avoidance decision is more in line with the actual task requirements and the task completion efficiency is improved. Through feasibility analysis and screening, it is ensured that the generated obstacle avoidance decision has a high possibility of execution, reducing the probability of the drone encountering difficulties in actual flight. The simulation link is introduced to optimize the initial obstacle avoidance decision, further improving the reliability and adaptability of the decision. The overall method takes into account many factors, realizes a comprehensive analysis and optimization of the obstacle avoidance path of the drone, is suitable for autonomous flight in complex environments, and effectively improves the safety of the drone and the efficiency of task execution.

[0111] Reference Figure 2 As shown, the present invention also provides a drone area boundary automatic recognition and obstacle avoidance system, which is applied to any of the above-mentioned drone area boundary automatic recognition and obstacle avoidance methods, including: The acquisition module is used to obtain the environmental perception data of the drone sensor, classify the environmental perception data into types, and obtain the corresponding image data, point cloud data and continuous image frames; The analysis module is used to perform boundary recognition processing on the image data to obtain the corresponding boundary recognition results, and to extract obstacle features from the point cloud data to obtain the corresponding obstacle features; Prediction module, which is used to predict the long-range dependency features of continuous image frames through a preset Transformer model to obtain corresponding motion prediction information; The association module is used to perform three-dimensional construction based on boundary recognition results, obstacle characteristics and motion prediction information to obtain a three-dimensional environment map; A processing module, the processing module is used to perform boundary tracking based on the three-dimensional environment map to obtain corresponding dynamic boundary information, and perform obstacle avoidance path planning on the three-dimensional environment map according to the dynamic boundary information to obtain a corresponding obstacle avoidance path; The control module is used to perform decision analysis on the obstacle avoidance path, dynamic boundary information and preset mission objectives to generate automatic obstacle avoidance decisions.

[0112] The present invention provides an automatic identification and obstacle avoidance system for unmanned aerial vehicle area boundaries. By comprehensively utilizing image data, point cloud data and continuous image frames, the system realizes multi-dimensional perception of the environment, improves the accuracy of boundary identification and obstacle detection, and overcomes the limitations that may be caused by a single data source. By fusing and three-dimensionally constructing multi-source data, a more accurate and comprehensive three-dimensional environmental map is generated, which provides reliable basic information for the navigation and obstacle avoidance of the unmanned aerial vehicle. Based on the real-time tracking and updating mechanism of dynamic boundary information, the system can respond to environmental changes in a timely manner, and improves the adaptability of the unmanned aerial vehicle in complex and rapidly changing environments. By combining dynamic boundary information and three-dimensional environmental maps for obstacle avoidance path planning, a more intelligent and flexible obstacle avoidance strategy is realized, which improves the safety and mission execution efficiency of the unmanned aerial vehicle. By combining obstacle avoidance paths, dynamic boundary information and preset mission targets, automatic obstacle avoidance decisions are generated, which realizes the autonomous decision-making ability of the unmanned aerial vehicle in the process of executing tasks and improves the overall intelligence level of the system.

[0113] Reference Figure 3 As shown, the present invention also provides a device for automatically identifying and avoiding obstacles of a drone area boundary, comprising: Memory, used to store programs; The processor is used to execute the program to implement each step of any one of the above-mentioned methods for automatic identification of drone area boundaries and obstacle avoidance.

[0114] In this embodiment, the processor and the memory may be connected via a bus or other means. The memory may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a read-only memory, a flash memory, a hard disk, or a solid-state drive. The processor may be a general-purpose processor, such as a central processing unit, a digital signal processor, an application-specific integrated circuit, or one or more integrated circuits configured to implement the embodiments of the present invention.

[0115] The present invention also provides a storage medium storing computer instructions, wherein the computer instructions are used to enable a computer to execute any of the above methods.

[0116] It should be noted that technicians in the relevant technical field can clearly understand that for the convenience and conciseness of description, the specific working process of the system and each module described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0117] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for automatic identification and obstacle avoidance of drone area boundaries, characterized in that: include: Acquire environmental perception data from the drone sensor, classify the environmental perception data into types, and obtain corresponding image data, point cloud data, and continuous image frames; Performing boundary recognition processing on the image data to obtain corresponding boundary recognition results, and performing obstacle feature extraction on the point cloud data to obtain corresponding obstacle features; Predicting the long-range dependent features of the continuous image frames through a preset Transformer model to obtain corresponding motion prediction information; Performing three-dimensional construction based on the boundary recognition result, the obstacle characteristics and the motion prediction information to obtain a three-dimensional environment map; Perform boundary tracking based on the three-dimensional environment map to obtain corresponding dynamic boundary information, and perform obstacle avoidance path planning on the three-dimensional environment map according to the dynamic boundary information to obtain a corresponding obstacle avoidance path; A decision analysis is performed on the obstacle avoidance path, the dynamic boundary information and the preset task goal to generate an automatic obstacle avoidance decision.

2. The method for automatic identification and obstacle avoidance of drone area boundaries according to claim 1 is characterized in that: The step of acquiring environmental perception data from the drone sensor, classifying the environmental perception data into different types, and obtaining corresponding image data, point cloud data, and continuous image frames includes: Acquire the original environmental data collected by the drone sensor, the original environmental data including single-frame image data, multiple-frame continuous image data, original infrared image data, original laser radar data and original ultrasonic data, Performing time stamp alignment and image fusion processing on the single-frame image data and the original infrared image data to obtain the fused image data; Performing Euclidean clustering fusion on the original laser radar data and the original ultrasonic data to obtain the structured point cloud data; The multiple frames of continuous image data are arranged in order of timestamps, and an image sequence is constructed to obtain the corresponding continuous image frames.

3. The method for automatic identification and obstacle avoidance of drone area boundaries according to claim 1, characterized in that: The performing boundary recognition processing on the image data to obtain a corresponding boundary recognition result, and performing obstacle feature extraction on the point cloud data to obtain a corresponding obstacle feature, comprises: Performing edge feature detection on the image data to obtain corresponding edge feature information; Performing image segmentation on the image data according to the edge feature information to obtain a regional image; Performing contour analysis on the regional image according to the edge feature information to obtain boundary contour information; Performing boundary fitting and recognition on the image data according to the boundary contour information and the edge feature information to obtain the boundary recognition result; Performing vector extraction on the point cloud data to obtain an obstacle vector; Feature classification is performed on the obstacle vector to obtain the obstacle feature.

4. The method for automatic identification and obstacle avoidance of drone area boundaries according to claim 1, characterized in that: The performing long-range dependency feature prediction on the continuous image frames through a preset Transformer model to obtain corresponding motion prediction information includes: Extracting features from the continuous image frames through the input layer of the Transformer model to obtain an image feature sequence; Performing sinusoidal position coding on the image feature sequence to obtain feature sequence coding; Input the feature sequence encoding into a Transformer encoder, perform self-attention calculation through a multi-head self-attention mechanism layer of the Transformer encoder, and obtain a self-attention output; The self-attention output is encoded and transformed by an encoding feedforward neural network layer, and input into an encoding residual connection layer for stacking processing to obtain an encoding feature representation; Inputting the encoded feature representation into the global self-attention mechanism layer to calculate the correlation between different time steps to obtain long-range dependency features; Inputting the long-range dependency feature into a Transformer decoder, performing cross-attention calculation on the long-range dependency feature and the encoding feature representation through a multi-head cross-attention mechanism layer of the Transformer decoder, and obtaining a cross-attention output; The cross attention output is encoded and transformed by a decoding feedforward neural network layer, and input into an encoding residual connection layer for decoding processing to obtain a decoding feature sequence; Performing time series prediction on the decoded feature sequence to obtain a predicted feature sequence; Inputting the predicted feature sequence into the fully connected layer to perform motion state prediction to obtain a preliminary motion prediction result; Performing uncertainty analysis on the preliminary motion prediction result to obtain prediction confidence; The prediction confidence and the preliminary motion prediction result are input into the output layer of the Transformer model for prediction correction, and the motion prediction information is output.

5. The method for automatic identification and obstacle avoidance of drone area boundaries according to claim 1, characterized in that: The three-dimensional construction is performed based on the boundary recognition result, the obstacle characteristics and the motion prediction information to obtain a three-dimensional environment map, including: Extracting boundary points from the boundary recognition result to obtain a boundary point set; Delineating the boundary according to the boundary point set to obtain a corresponding boundary curve; Performing three-dimensional analysis on the obstacle features to obtain three-dimensional obstacle data; Performing three-dimensional modeling based on the obstacle characteristics and the three-dimensional obstacle data to obtain a three-dimensional model of the obstacle; Performing coordinate transformation on the motion prediction information to obtain a motion trajectory in a global coordinate system; An initial three-dimensional environment map is constructed according to the boundary curve, the obstacle three-dimensional model and the motion trajectory, wherein the process of constructing the initial three-dimensional environment map includes: Generate a boundary plane mesh according to the boundary curve to obtain a boundary mesh model; Calculating the position and posture information of the UAV at each time point according to the motion trajectory; Projecting the boundary grid model and the obstacle three-dimensional model into the same coordinate system according to the posture information, and performing spatial registration to obtain an aligned environment model; Triangulating the environment model to obtain a triangular mesh model; Reconstructing the surface of the triangular mesh model to obtain an initial environment map; The initial environment map is subjected to three-dimensional gridding processing to obtain the three-dimensional environment map.

6. The method for automatic identification and obstacle avoidance of drone area boundaries according to claim 1, characterized in that: The performing boundary tracking based on the three-dimensional environment map to obtain corresponding dynamic boundary information, and performing obstacle avoidance path planning on the three-dimensional environment map according to the dynamic boundary information to obtain a corresponding obstacle avoidance path, includes: Performing dynamic boundary calculation on the three-dimensional environment map to obtain a corresponding dynamic area edge contour; Calculating the curvature of contour points on the edge contour of the dynamic area to obtain corresponding dynamic contour points; Performing boundary analysis based on the dynamic contour points combined with the dynamic area edge contour to obtain the dynamic boundary information; Performing obstacle analysis on the three-dimensional environment map based on the dynamic boundary information to obtain corresponding dynamic obstacle information, and constructing a time-varying constraint field to obtain corresponding dynamic obstacle avoidance constraint conditions; Performing path planning on the three-dimensional environment map according to the dynamic obstacle avoidance constraint condition to obtain an initial path; The initial path is optimized and verified according to the dynamic obstacle information to obtain the obstacle avoidance path.

7. The method for automatic identification and obstacle avoidance of drone area boundaries according to claim 1, characterized in that: The performing decision analysis on the obstacle avoidance path, the dynamic boundary information and the preset task goal to generate an automatic obstacle avoidance decision includes: Splitting the obstacle avoidance path into sections to obtain a plurality of path sub-segments; Performing risk assessment on each of the path sub-segments to obtain a corresponding risk coefficient; Prioritizing the path sub-segments according to the risk coefficients to obtain an initial ranking result; Adjusting the initial sorting result according to the dynamic boundary information to obtain an optimized sorting result; Decomposing the preset task goal to obtain multiple sub-task goals; Associating and matching the optimization sorting results with the subtask objectives to obtain multiple task path association combinations; Performing feasibility analysis on each of the task path association combinations to obtain a corresponding feasibility score; The feasibility scores are screened according to a preset feasibility screening rule to obtain a feasibility score set; Selecting the corresponding task path association combination according to the feasibility score set to perform decision analysis to obtain an initial obstacle avoidance decision; Performing simulation on the initial obstacle avoidance decision to obtain a simulation result; The initial obstacle avoidance decision is optimized according to the simulation result to obtain the automatic obstacle avoidance decision.

8. An automatic identification and obstacle avoidance device for drone area boundaries, characterized in that: The method for automatically identifying and avoiding obstacles of a drone region boundary as described in any one of claims 1 to 7 above comprises: An acquisition module, which is used to obtain environmental perception data from drone sensors, classify the environmental perception data into types, and obtain corresponding image data, point cloud data, and continuous image frames; An analysis module, the analysis module is used to perform boundary recognition processing on the image data to obtain corresponding boundary recognition results, and to extract obstacle features from the point cloud data to obtain corresponding obstacle features; A prediction module, the prediction module is used to perform long-range dependency feature prediction on the continuous image frames through a preset Transformer model to obtain corresponding motion prediction information; An association module, the association module is used to perform three-dimensional construction based on the boundary recognition result, the obstacle characteristics and the motion prediction information to obtain a three-dimensional environment map; A processing module, the processing module is used to perform boundary tracking based on the three-dimensional environment map to obtain corresponding dynamic boundary information, and perform obstacle avoidance path planning on the three-dimensional environment map according to the dynamic boundary information to obtain a corresponding obstacle avoidance path; A control module is used to perform decision analysis on the obstacle avoidance path, the dynamic boundary information and the preset task objectives to generate an automatic obstacle avoidance decision.

Images