UAV vision-guided obstacle avoidance system and method based on polarization information

Abstract

The present invention discloses a vision-guided obstacle avoidance system and method for unmanned aerial vehicle (UAV) based on polarization information, and relates to the technical field of navigation and obstacle avoidance for UAVs. Polarization image information in flight scenes is collected in real time by a polarization sensor to enhance the perception of complex environments. Polarization parameters are extracted by a compressed observation method to improve the efficiency and accuracy of feature extraction. Combined with the visual attention mechanism, the obstacle background area is accurately identified, an adaptive decision fusion model is constructed, feature weights are optimized, and the reliability of target contour features is enhanced. A preliminary path search is performed based on the target contour features, and a centralized optimization sampling method is adopted to effectively improve the probability of generating the optimal flight path in a complex environment. The obstacle avoidance flight module smoothes the flight path to ensure the smoothness and controllability of the flight route, effectively improving the navigation and obstacle avoidance capabilities of the UAV in complex environments.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle (UAV) navigation and obstacle avoidance technology, and in particular to a UAV vision-guided obstacle avoidance system and method based on polarization information. Background Art

[0002] Currently, drone technology is widely used in various fields, such as agricultural monitoring, environmental monitoring, and express logistics. However, in complex environments, traditional vision-guided obstacle avoidance systems often rely on RGB images for target detection and path planning. While these systems can identify static obstacles, they often struggle to accurately determine the nature and location of obstacles in dynamic environments, especially those with complex lighting conditions and significant background noise. This can lead to obstacle avoidance failures or inappropriate path planning. Furthermore, existing technologies perform poorly in handling obstructions and reflected light sources, impacting the drone's autonomous flight capabilities and safety.

[0003] Most existing technical solutions acquire image information through a single visual sensor and process it using computer vision algorithms. However, traditional RGB images fail to fully utilize light polarization information, resulting in low target detection accuracy and slow response speed. Furthermore, these solutions are susceptible to the effects of reflected light and shadows when processing high dynamic range scenes, leading to misjudgments or missed detections. Therefore, there is an urgent need for a new vision-guided obstacle avoidance system for drones that can effectively integrate polarization information to improve the accuracy and reliability of obstacle detection, thereby enabling more efficient path planning and obstacle avoidance flight. Summary of the Invention

[0004] In response to the deficiencies of the prior art, the present invention provides a UAV vision-guided obstacle avoidance system and method based on polarization information, which solves the problems of the above-mentioned background technology.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a UAV vision-guided obstacle avoidance system based on polarization information, comprising the following modules: a polarization information acquisition module, a polarization information extraction module, an obstacle target detection module, a path planning module, and an obstacle avoidance flight module; the polarization information acquisition module is used to collect polarization image information in a flight scene through a polarization sensor and obtain images with different polarization directions; the polarization information extraction module is used to extract feature information from images with different polarization directions through compressed observation, and set an extraction termination condition to obtain polarization parameters; the obstacle target detection module is used to perform multi-level clustering on the polarization parameters and extract the global uniqueness feature of the degree of polarization layer by layer. The system uses the global feature distribution and spatial distribution characteristics to judge the obstacle target background area through the visual attention mechanism and mark it. According to the global feature distribution and the marked obstacle target background area, an adaptive decision fusion model is constructed to optimize the feature weights and output the target contour features. The path planning module is used to perform a preliminary path search based on the target contour features through the path planning algorithm, and make obstacle avoidance judgments according to the global feature distribution and the marked obstacle target background area. The preliminary path search is centrally optimized and sampled to output the optimal flight path. The obstacle avoidance flight module is used to smooth the optimal flight path, remove redundant inflection points, and parse the smoothed flight path into a waypoint sequence.

[0006] Furthermore, the specific process of extracting feature information from images with different polarization directions through compressed observation is as follows: the flight scene is divided into several scene blocks, and a Gaussian random matrix is ​​applied to each scene block as a compressed observation matrix to perform multiple measurements; the measurement value of each scene block is generated according to the compressed sensing theory, and only the measurement matrix of the calculated scene block is stored.

[0007] Furthermore, the specific process of setting the extraction termination condition is as follows: reconstruct the scene for each transmitted image block measurement value, calculate the inner product between the current reconstruction result and the previous reconstruction result, and when the inner product value gradually approaches 1, judge the accuracy of the reconstruction result; when the inner product reaches the threshold, it indicates that the scene reconstruction information is saturated, and the measurement value transmission is terminated.

[0008] Furthermore, multi-level clustering is performed on the polarization parameters to extract the global uniqueness and spatial distribution characteristics of the polarization degree layer by layer. The specific process is as follows: First-level clustering: The polarization degree is preliminarily clustered using a Gaussian mixture model to obtain the probability value of each pixel under each Gaussian component, the polarization degree of each pixel is calculated, a probability vector is generated, and the two elements with the largest probability are retained; Second-level clustering: Re-clustering is performed using the probability density function of the first layer to extract the feature distribution, calculate the similarity between different pixels, and generate a similarity matrix; the global uniqueness characteristics of the target are extracted by calculating the mean polarization degree and spatial distance of each region; the spatial distribution characteristics of the polarization degree are calculated, the spatial variance of each pixel is evaluated, and the target area is identified.

[0009] Furthermore, the specific process of judging the background area of ​​the obstacle target and marking it through the visual attention mechanism is as follows: Background area judgment: According to the probability vector and feature distribution information generated by the first layer of clustering, the boundary between the obstacle target and the background area in the image is identified; by analyzing the boundary connectivity of each area, the similarity between the target area and the background area is calculated, and their spatial distribution differences are judged; Marking process: For the identified background area, the visual attention mechanism is applied to mark the area through the set threshold.

[0010] Furthermore, the specific process of constructing an adaptive decision fusion model, optimizing feature weights, and outputting target contour features is as follows: constructing a multi-level decision framework through comprehensive analysis of the extracted polarization parameter features, background area features marked by the visual attention mechanism, and global uniqueness features; through weighted fusion, various features are comprehensively processed through weighted coefficients to construct an adaptive decision fusion model; dynamically adjusting feature weights through an adaptive algorithm to output target contour features.

[0011] Furthermore, based on the target contour features, the logical process of performing preliminary path search through the path planning algorithm is as follows: construct a tree structure with the starting point as the root node; in free space, add vertices by random sampling to expand the tree structure; reconnect the tree by minimizing the cost of the newly added vertices and nearby vertices; after finding a preliminary feasible solution, generate a local ellipsoid optimization area of ​​the solution.

[0012] Furthermore, obstacle avoidance judgment is made based on the global feature distribution and the marked obstacle target background area. The specific process of centralized optimization sampling of the preliminary path search is as follows: the cost of the preliminary path is calculated. When the cost is lower than the set threshold, the centralized optimization stage is entered; the ellipsoid area is sampled in the local space of the solution to further refine the path; the tree structure is updated, new low-cost vertices are connected, the path is optimized, the updated path is compared with the preliminary path, and the one with the lowest cost is selected as the optimal flight path.

[0013] A vision-guided obstacle avoidance method for a UAV based on polarization information comprises the following steps: S1. collecting polarization image information in a flight scene through a polarization sensor to obtain images with different polarization directions; S2. extracting feature information from images with different polarization directions through compressed observation, setting extraction termination conditions, and obtaining polarization parameters; S3. performing multi-level clustering on the polarization parameters, extracting the global uniqueness and spatial distribution features of the polarization degree layer by layer, judging the obstacle target background area through a visual attention mechanism, and marking it; constructing an adaptive decision fusion model based on the global feature distribution and the marked obstacle target background area, optimizing feature weights, and outputting target contour features; S4. performing a preliminary path search based on the target contour features through a path planning algorithm, performing obstacle avoidance judgment based on the global feature distribution and the marked obstacle target background area, performing centralized optimization sampling on the preliminary path search, and outputting the optimal flight path; S5. smoothing the optimal flight path, removing redundant inflection points, and parsing the smoothed flight path into a waypoint sequence.

[0014] The present invention has the following beneficial effects:

[0015] (1) The UAV visual guidance obstacle avoidance system based on polarization information collects polarization image information in the flight scene through a polarization sensor, obtains images with different polarization directions, provides multi-dimensional information of the environment, extracts feature information from images with different polarization directions through compressed observation, sets extraction termination conditions, obtains polarization parameters, and extracts key polarization features, thereby enhancing target recognition capabilities and processing efficiency. The polarization parameters are analyzed through multi-level clustering, and global uniqueness features and spatial distribution features are extracted layer by layer. The background area of ​​the obstacle target is judged and marked by combining the visual attention mechanism. The accuracy and speed of obstacle detection are improved, and potential obstacles in complex environments can be quickly and accurately identified, reducing flight risks.

[0016] (2) The UAV visual guidance obstacle avoidance method based on polarization information performs preliminary path search and obstacle avoidance judgment through the path planning algorithm of the target contour features. And through centralized optimization sampling, the success rate and efficiency of path planning are improved, and the global optimal path can be found quickly. It ensures that the UAV can flexibly respond to complex environments during flight and improves the accuracy and effectiveness of path planning. The optimal flight path is smoothed and redundant inflection points are removed to ensure the stability and safety of the UAV flight. This processing not only improves the comfort of flight, but also reduces the impact on the equipment, reduces the failure rate, and extends the service life of the UAV. At the same time, the smooth flight path improves flight efficiency and ensures that the UAV can fly smoothly in complex environments.

[0017] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of the UAV vision-guided obstacle avoidance system based on polarization information of the present invention.

[0019] Figure 2 This is a flow chart of the UAV vision-guided obstacle avoidance method based on polarization information of the present invention. DETAILED DESCRIPTION

[0020] The present invention utilizes a polarization-based vision-guided obstacle avoidance system and method for drones, effectively addressing the difficulty drones face in accurately identifying obstacles and planning safe flight paths in complex environments. By collecting and processing multi-level polarization information, the system analyzes the surrounding environment in real time, improving obstacle detection accuracy and optimizing path planning, ensuring safe and stable flight in various lighting conditions.

[0021] The overall approach to the problems in the embodiments of this application is as follows:

[0022] Polarization image information in the flight scene is collected through the polarization sensor to obtain images in different polarization directions.

[0023] Feature information of images with different polarization directions is extracted through compressed observation, and the extraction termination condition is set to obtain polarization parameters.

[0024] Multi-level clustering is performed on the polarization parameters to extract the global uniqueness and spatial distribution characteristics of the polarization degree layer by layer. The obstacle target background area is determined and marked through the visual attention mechanism. Based on the global feature distribution and the marked obstacle target background area, an adaptive decision fusion model is constructed to optimize the feature weights and output the target contour features.

[0025] Based on the target contour features, a preliminary path search is performed through the path planning algorithm. Obstacle avoidance judgment is made based on the global feature distribution and the marked obstacle target background area. The preliminary path search is centrally optimized and sampled to output the optimal flight path.

[0026] The optimal flight path is smoothed to remove redundant inflection points, and the smoothed flight path is parsed into a waypoint sequence.

[0027] See also Figure 1The embodiment of the present invention provides a technical solution: a UAV vision-guided obstacle avoidance system based on polarization information, comprising the following modules: a polarization information acquisition module, a polarization information extraction module, an obstacle target detection module, a path planning module, and an obstacle avoidance flight module; the polarization information acquisition module is used to collect polarization image information in the flight scene through a polarization sensor and obtain images with different polarization directions; the polarization information extraction module is used to extract feature information from images with different polarization directions through compressed observation, and set an extraction termination condition to obtain polarization parameters; the obstacle target detection module is used to perform multi-level clustering on the polarization parameters, and extract the global uniqueness features and spatial distribution of polarization degrees layer by layer. The system uses the global feature distribution and the marked obstacle target background area to determine the obstacle target background area and mark it. It builds an adaptive decision fusion model based on the global feature distribution and the marked obstacle target background area, optimizes the feature weights, and outputs the target contour features. The path planning module is used to perform a preliminary path search based on the target contour features through the path planning algorithm, and makes obstacle avoidance judgments based on the global feature distribution and the marked obstacle target background area. It performs centralized optimization sampling on the preliminary path search and outputs the optimal flight path. The obstacle avoidance flight module is used to smooth the optimal flight path, remove redundant inflection points, and parse the smoothed flight path into a waypoint sequence.

[0028] In this embodiment, the polarization information acquisition module includes: Polarization information: refers to the characteristics of the vibration direction of light waves, typically obtained through a polarization sensor. Different surfaces or objects will have different effects on the polarization state of light waves, so polarization information can provide important clues about the surface characteristics of objects and environmental information. Polarization sensor: A sensor used to capture polarized light information, which can measure the polarization state of light to generate a polarization image. Polarization information extraction module: Compressed observation: A data processing technique used to reduce the amount of data while retaining important features. This method can improve processing efficiency and reduce computational burden. Feature information extraction: Refers to extracting useful feature information from the original image for subsequent analysis and processing. Obstacle target detection module: Multi-level clustering: A machine learning technique that identifies patterns by grouping data into different categories. In this case, it is used to analyze polarization parameters to identify obstacles. Globally unique features: Refers to features that are unique in a dataset and can effectively distinguish different objects or areas. Spatial distribution features: Features that describe the distribution of data in space and can reflect the shape, position, and mutual relationships of objects. Visual attention mechanism: a technology that imitates the human visual system, focusing on important areas in the image to improve the efficiency and accuracy of target recognition. Adaptive decision fusion model: an intelligent model that integrates information from different sources to make decisions and can dynamically adjust to optimize the results. Path planning module: path planning algorithm: an algorithm used to determine the optimal path from the starting point to the end point. In this system, the flight path of the drone is generated by analyzing environmental characteristics and obstacle information. Focused optimization sampling: refers to finer sampling within the local area of ​​the known solution to improve the efficiency and accuracy of path search. Obstacle avoidance flight module: smoothing: a post-processing technology used to optimize the path, remove unnecessary inflection points, and make the flight path more natural and efficient. Waypoint sequence: a flight path consisting of a series of predetermined position points for the drone to refer to and navigate during flight.

[0029] Specifically, the specific process of extracting feature information from images with different polarization directions through compressed observation is as follows: the flight scene is divided into several scene blocks, and a Gaussian random matrix is ​​applied to each scene block as a compressed observation matrix to perform multiple measurements; the measurement value of each scene block is generated according to the compressed sensing theory, and only the measurement matrix of the calculated scene block is stored.

[0030] In this embodiment, the scene T is divided into n scene blocks, each of which is denoted as B. i (0≤i<n), each block size is z i (0≤i<n). Design measurement matrix λ={λ i}(0≤i<n,λ∈R j×z ) as the compressed measurement matrix of T, where the sub-measurement matrix λ of λi It is scene block B i The block measurement matrix, the measurement matrix λ is a Gaussian random matrix, each matrix element in the matrix obeys the independent normal distribution N(0,1), let the block measurement matrix λ i =λ[1:k i ,1:z i ], k i For B i Block measurement times, i.e. block B i According to the theory of compressed sensing, the measurement value I of scene T can be expressed as:

[0031] λ i Scene block B i Block measurement matrix of scene block B i The size of z i Sum block measurement times k i This directly determines the size of the intra-block measurement matrix. When calculating measurements, only the intra-block measurement matrix for each scene block needs to be stored, eliminating the need to store the measurement matrix λ for the entire target scene. This significantly reduces the storage and computational complexity of measurements during compressed observation. Furthermore, according to normal observation practices, when a scene block is located at the edge of the scene, it contains more background information than target information. At this time, the number of block measurements for this scene block is small, resulting in a smaller block measurement matrix and fewer measurement values. This largely eliminates useless background information within the scene block.

[0032] Specifically, the specific process of setting the extraction termination condition is as follows: reconstruct the scene for each transmitted image block measurement value, calculate the inner product between the current reconstruction result and the previous reconstruction result, and when the inner product value gradually approaches 1, judge the accuracy of the reconstruction result; when the inner product reaches the threshold, it indicates that the scene reconstruction information is saturated, and the measurement value transmission is terminated.

[0033] In this embodiment, the compressed measurement value I iImage reconstruction is performed by fusion. To accurately reconstruct an n-dimensional signal with a fixed sparsity S in the transform domain, the total number of measurements k for all scene blocks must satisfy the equation k = O(S·ln(n)). However, before observing scene information, the neural network cannot know the exact sparsity value of the scene to be observed in a certain transform domain. Furthermore, due to the variability of scene information and the adaptive nature of the neural network, the total number of measurements k is also uncertain. Therefore, the photoreceptors observing the information do not need to obtain the optimal number of transmissions for the measurement value; they only need to minimize the number of compressed observations while maximizing the quality of scene reconstruction. Assuming the scene reconstruction quality is Q, the objective function that the received signal value must satisfy is: {I} = argmax(Q)i∈V(1-3); where V represents the set of all scene blocks. Since the original image is unavailable, there is no standard for the size of the scene reconstruction quality Q. Therefore, each time an image block measurement value is transmitted, the scene is reconstructed based on the aggregated measurement value. The quality of the reconstructed scene is measured by calculating the inner product q between the current reconstruction result and the previous reconstruction result. The inner product calculation formula is as follows: j = <T j ,T j+1 >(1-4); T in formula (1-4) j and T j+1 They represent the jth and j+1th reconstruction results respectively. As the number of measurements increases, the reconstruction results become more and more accurate. The inner product q j It is also getting closer to 1. When q j =1, the reconstruction result information is saturated and the transmission of the measurement value is terminated.

[0034] Specifically, the polarization parameters are clustered at multiple levels to extract the global uniqueness and spatial distribution characteristics of the polarization degree layer by layer. The specific process is as follows: First-layer clustering: The polarization degree is preliminarily clustered using a Gaussian mixture model to obtain the probability value of each pixel under each Gaussian component, calculate the polarization degree of each pixel, generate a probability vector, and retain the two elements with the largest probability; Second-layer clustering: Re-clustering is performed using the probability density function of the first layer to extract the feature distribution, calculate the similarity between different pixels, and generate a similarity matrix; the global uniqueness characteristics of the target are extracted by calculating the mean polarization degree and spatial distance of each region; the spatial distribution characteristics of the polarization degree are calculated, the spatial variance of each pixel is evaluated, and the target area is identified.

[0035] In this embodiment, in the first layer of feature extraction, the polarization degree is clustered and calculated using the Gaussian mixture model. The polarization degree of each pixel I x It is represented by k Gaussian components linearly weighted, I x Belongs to the kth i The probability of (i=1,2,...,N) components is calculated using the following formula: Where, ω k Indicates the kth i The weighting coefficients of the Gaussian components, and are the mean polarization degree and covariance matrix of the Gaussian component respectively. Get model k i The global feature F(k i ) in, Represents component k i and k j The Euclidean distance of the mean polarization degree, D(k i ,k j ) represents the spatial distance of the component center. According to formula (2-1), the polarization probability vector P=[p1,p2,...,p k ], each probability vector P is a Gaussian component, retain the two elements with the largest probability in the vector, and set the others to zero. The message passing clustering algorithm is used to cluster the Gaussian components retained for each pixel. The similarity matrix in the message passing clustering algorithm is given by the spatial correlation matrix between the Gaussian components in the upper layer. The Gaussian component k i and k j The spatial correlation matrix is ​​expressed as: Then after the second layer clustering calculation, the polarization degree of each pixel I x The total probability of belonging to model K can be expressed as: Then the calculation formula for the spatial distribution of polarization degree of model K is: in, Represents pixel I x The distance from the center of all pixels. The spatial distribution characteristic of the polarization degree is the sum of the weighted spatial variances. The smaller the spatial variance of the polarization degree, the more likely the area where the polarization degree is located belongs to the target.

[0036] Specifically, the specific process of judging the obstacle target background area through the visual attention mechanism and marking it is as follows: Background area judgment: According to the probability vector and feature distribution information generated by the first layer of clustering, the boundary between the obstacle target and the background area in the image is identified; by analyzing the boundary connectivity of each area, the similarity between the target area and the background area is calculated, and their spatial distribution differences are judged; Marking process: For the identified background area, the visual attention mechanism is applied to mark the area through the set threshold.

[0037] In this embodiment, background region determination: Probability vector generated by first-level clustering: During the multi-level clustering process, the first-level clustering generates a probability vector by analyzing the feature information of the polarization image. This vector represents the probability that different regions in the image belong to obstacles or background. Feature distribution information: Feature distribution information includes visual features related to obstacle targets (such as color, shape, texture, etc.) and visual features of the background. By comparing these features, the system can determine which regions are obstacles and which are background. Boundary identification: Based on the probability vector and feature distribution information, the system analyzes features in the image to identify the boundaries between obstacle targets and background regions. This process may involve image processing techniques such as edge detection to define the boundaries between different regions. Boundary connectivity analysis: Analyzing the boundary connectivity of each region is an important step in determining the integrity and coherence of the background and target regions. Connectivity analysis can help determine the presence of obstacles and their coherence. Similarity calculation: The similarity between the target and background regions is calculated to determine the differences in their spatial distribution. This process may use various similarity metrics (such as cosine similarity and Euclidean distance) to quantify the degree of similarity between regions. Labeling process: Visual attention mechanism: The visual attention mechanism is a technology that mimics human visual processing and can focus on important parts of the image. By analyzing the background area, the system will be able to distinguish which areas require further processing. Setting a threshold: In order to determine whether to mark an area as background, the system needs to set a threshold. This threshold can be determined based on previous training or experience data to determine whether the feature similarity meets the labeling requirements. Region labeling: Once the similarity of a region exceeds the set threshold, the region is marked as background. This labeling process will help the system distinguish between obstacle targets and background in subsequent processing, thereby making more effective obstacle avoidance decisions.

[0038] Specifically, the specific process of constructing an adaptive decision fusion model, optimizing feature weights, and outputting target contour features is as follows: constructing a multi-level decision framework through comprehensive analysis of the extracted polarization parameter features, background area features marked by the visual attention mechanism, and global uniqueness features; through weighted fusion, various features are comprehensively processed through weighted coefficients to construct an adaptive decision fusion model; dynamically adjusting feature weights through an adaptive algorithm to output target contour features.

[0039] In this embodiment, extracted polarization parameter features: Features obtained from the polarization information extraction module typically include information such as polarization degree and polarization direction. These features provide the reflective properties of the obstacle surface, helping to distinguish obstacles from the background. Background area features marked by the visual attention mechanism: Background area features marked by the visual attention mechanism reflect the recognition of the target under specific background conditions. These markings provide boundary information between the obstacle and the background, further helping to distinguish the target from the surrounding environment. Globally unique features: Globally unique features refer to the overall feature extraction of a specific target within the entire image, typically involving information such as the target's shape, size, and color distribution, ensuring accurate target recognition in multiple environments. Multi-level decision framework: Features from the above different sources are analyzed at multiple levels to construct a decision framework consisting of multiple levels. Each level of the framework can process different levels of information, such as local features, global features, and contextual information, making the decision more comprehensive. Weighted fusion: Different features are weighted using weighting coefficients to form a comprehensive feature representation. This process may use mathematical methods such as weighted average and weighted sum to ensure that each feature has its corresponding influence in the decision. Adaptive decision fusion model: This model can dynamically adjust the weights of features based on feedback and new data in different environments and conditions. Through training data or real-time feedback, the model can learn which features are more important under specific conditions, thereby optimizing the decision-making process. Adaptive algorithm: The adaptive algorithm dynamically adjusts feature weights by monitoring environmental changes and system performance in real time. This mechanism allows the model to respond quickly to new information and optimize decision quality. For example, in some cases, the degree of polarization may better reflect the characteristics of obstacles, while in other cases, global uniqueness features may be more important. Target contour features: After the above comprehensive analysis and weighted processing, the target contour features finally output by the model contain the boundary information and shape characteristics of the obstacle. These contour features provide important input for the subsequent path planning module, helping the drone to make efficient obstacle avoidance decisions in complex environments.

[0040] Specifically, based on the target contour features, the logical process of preliminary path search through the path planning algorithm is as follows: construct a tree structure with the starting point as the root node; in free space, add vertices by random sampling to expand the tree structure; reconnect the tree by minimizing the cost of the newly added vertices and nearby vertices; after finding a preliminary feasible solution, generate a local ellipsoid optimization area of ​​the solution.

[0041] In this implementation, the RRT algorithm based on concentrated search in the ellipsoid region proceeds as follows: a tree T = (V, E) is gradually constructed in the planning space, where V is a set of vertices and E is the edges connecting these vertices. The tree structure is initialized with the starting point as the root node. New vertices are added to the tree through random sampling in the free space. This process expands the tree structure by generating random states, gradually covering the planning space. Tree reconnection: For each newly added vertex, the tree is reconnected by minimizing the cost between the newly added vertex and nearby vertices. This process ensures that the cost of adjacent vertices in the tree is minimized, thereby enhancing the path optimization capability. Centralized Optimization Phase: Once a preliminary feasible solution is found (i.e., path p is not empty), the centralized sampling RRT* algorithm will perform optimized sampling of the ellipsoid region within the local solution space by calculating the cost value of the solution. Ellipsoid Region Sampling: During the centralized optimization phase, the ellipsoid region is sampled based on the cost value of the solution to further refine the path. This centralized sampling method in the local space can increase the probability of quickly finding the local optimal solution in complex environments. Update the tree structure: Connect new low-cost vertices to the tree and update the tree structure to optimize the path. During this process, the cost of the updated path is compared with the cost of the preliminary path, and the path with the lowest cost is selected as the optimal flight path.

[0042] Specifically, obstacle avoidance judgment is made based on the global feature distribution and the marked obstacle target background area, and the specific process of centralized optimization sampling of the preliminary path search is as follows: the cost of the preliminary path is calculated, and when the cost is lower than the set threshold, the centralized optimization stage is entered; the ellipsoid area is sampled in the local space of the solution to further refine the path; the tree structure is updated, new low-cost vertices are connected, the path is optimized, the updated path is compared with the preliminary path, and the one with the lowest cost is selected as the optimal flight path.

[0043] In this implementation scheme, obstacle avoidance judgment: obstacle avoidance judgment is performed based on the global feature distribution and the marked obstacle target background area. On this basis, the cost value of the preliminary path is calculated. Cost judgment and centralized optimization: when the cost is lower than the set threshold, the centralized optimization stage is entered, and the sampling operation of the ellipsoid area is performed. This step further refines the path, improves the planning success rate and convergence speed. Path optimization selection: during the optimization process, the tree structure is updated, and the newly generated low-cost vertices are connected to the tree and compared with the preliminary path. Finally, the path with the lowest cost is selected as the optimal flight path. The improved RRT* algorithm effectively improves the path planning capability of the UAV in complex environments, ensuring that while quickly finding the local optimal solution, the success rate and efficiency of planning are significantly improved.

[0044] See also Figure 2The method for avoiding obstacles by visual guidance of UAV based on polarization information includes the following steps: S1. collecting polarization image information in the flight scene through a polarization sensor to obtain images with different polarization directions; S2. extracting feature information of images with different polarization directions through compressed observation, setting extraction termination conditions, and obtaining polarization parameters; S3. performing multi-level clustering on the polarization parameters, extracting the global uniqueness feature and spatial distribution feature of the polarization degree layer by layer, judging the obstacle target background area through the visual attention mechanism, and marking it; constructing an adaptive decision fusion model according to the global feature distribution and the marked obstacle target background area, optimizing the feature weight, and outputting the target contour feature; S4. performing preliminary path search based on the target contour feature through the path planning algorithm, and performing obstacle avoidance judgment according to the global feature distribution and the marked obstacle target background area, performing centralized optimization sampling on the preliminary path search, and outputting the optimal flight path; S5. smoothing the optimal flight path, removing redundant inflection points, and parsing the smoothed flight path into a waypoint sequence.

[0045] In this embodiment, many redundant inflection points usually appear in the flight path generated by the path planning algorithm. These inflection points may be caused by the complexity of the environment or the characteristics of the path planning algorithm. In order to improve the smoothness and smoothness of the flight path, the path needs to be smoothed. The smoothing process mainly includes removing unnecessary inflection points to reduce sharp turns in the flight route and ensure that the drone can fly in a more natural way. Once the path is smoothed, the smoothed flight path can be parsed into a specific waypoint sequence. Each waypoint will contain corresponding coordinate location information, which is usually expressed in the form of longitude and latitude or three-dimensional space coordinates. The waypoint sequence provides clear flight instructions for the drone, ensuring that the drone can accurately navigate along the predetermined route during flight. This parsing not only helps the flight control system understand and execute flight missions, but also improves the safety and efficiency of the entire flight process.

[0046] In summary, this application has at least the following effects:

[0047] This polarization-based vision-guided obstacle avoidance system and method for unmanned aerial vehicles (UAVs) uses a polarization sensor to collect polarization image information from the flight scene, obtaining images with different polarization directions and providing multi-dimensional information about the environment. Feature information is extracted from these images using compressed observation, and termination conditions are set to obtain polarization parameters and extract key polarization features, thereby enhancing target recognition and processing efficiency. Polarization parameters are analyzed using multi-level clustering to extract global uniqueness and spatial distribution features layer by layer. A visual attention mechanism is then incorporated to identify and mark obstacle, target, and background regions. This improves the accuracy and speed of obstacle detection, enabling rapid and accurate identification of potential obstacles in complex environments and reducing flight risks. A path planning algorithm based on target contour features performs preliminary path search and obstacle avoidance decisions. Centralized optimization sampling improves the success rate and efficiency of path planning, enabling the rapid discovery of the globally optimal path. This ensures the UAV can flexibly adapt to complex environments during flight and enhances the accuracy and effectiveness of path planning. The optimal flight path is smoothed to remove redundant inflection points, ensuring stable and safe flight. This treatment not only improves flight comfort, but also reduces impact on the equipment, reduces failure rate, and extends the service life of the drone. At the same time, the smooth flight path improves flight efficiency and ensures smooth flight in complex environments.

[0048] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0049] The present invention is described with reference to flowcharts and / or block diagrams of systems, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0050] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0051] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0052] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0053] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. The UAV vision guidance obstacle avoidance system based on polarization information is characterized by: It includes the following modules: polarization information acquisition module, polarization information extraction module, obstacle target detection module, path planning module, and obstacle avoidance flight module; The polarization information acquisition module is used to collect polarization image information in the flight scene through a polarization sensor to obtain images with different polarization directions; The polarization information extraction module is used to extract feature information of images with different polarization directions through compressed observation, set extraction termination conditions, and obtain polarization parameters; The obstacle target detection module is used to perform multi-level clustering on polarization parameters and extract the global uniqueness and spatial distribution characteristics of polarization degree layer by layer. The specific process is as follows: First-layer clustering: Preliminary clustering of polarization degree using Gaussian mixture model, obtaining the probability value of each pixel under each Gaussian component, calculating the polarization degree of each pixel, generating a probability vector, and retaining the two elements with the highest probability; Second-layer clustering: Re-clustering using the probability density function of the first layer, extracting feature distribution, calculating the similarity between different pixels, and generating a similarity matrix; By calculating the mean polarization degree and spatial distance of each region, the global uniqueness of the target is extracted; The spatial distribution characteristics of the polarization degree are calculated, the spatial variance of each pixel is evaluated, and the target area is identified. The obstacle target detection module is also used to determine the obstacle target background area through the visual attention mechanism and mark it. Based on the global feature distribution and the marked obstacle target background area, an adaptive decision fusion model is constructed to optimize the feature weights and output the target contour features. The path planning module is used to perform a preliminary path search based on the target contour features through a path planning algorithm, and make obstacle avoidance judgments based on the global feature distribution and the marked obstacle target background area, and perform centralized optimization sampling on the preliminary path search to output the optimal flight path; The obstacle avoidance flight module is used to smooth the optimal flight path, remove redundant inflection points, and parse the smoothed flight path into a waypoint sequence; The specific process of setting the extraction termination condition is as follows: The scene is reconstructed for each transmitted image block measurement value, and the inner product between the current reconstruction result and the previous reconstruction result is calculated. When the inner product value gradually approaches 1, the accuracy of the reconstruction result is judged; When the inner product reaches the threshold, it indicates that the scene reconstruction information is saturated and the measurement value transmission is terminated.

2. The UAV vision-guided obstacle avoidance system based on polarization information according to claim 1, characterized in that: The specific process of extracting feature information from images with different polarization directions through compressed observation is as follows: The flight scene is divided into several scene blocks, and a Gaussian random matrix is ​​applied to each scene block as a compressed observation matrix to perform multiple measurements; The measurement value of each scene block is generated according to the compressed sensing theory, and only the measurement matrix of the calculated scene block is stored.

3. The UAV vision-guided obstacle avoidance system based on polarization information according to claim 2, characterized in that: The specific process of judging the obstacle target background area and marking it through the visual attention mechanism is as follows: Background area judgment: Based on the probability vector and feature distribution information generated by the first layer of clustering, the boundary between the obstacle target and the background area in the image is identified; By analyzing the boundary connectivity of each area, the similarity between the target area and the background area is calculated to determine their spatial distribution differences; Marking process: For the identified background area, the visual attention mechanism is applied to mark the area through the set threshold.

4. The UAV vision-guided obstacle avoidance system based on polarization information according to claim 3, characterized in that: The specific process of building an adaptive decision fusion model, optimizing feature weights, and outputting target contour features is as follows: A multi-level decision framework is constructed by comprehensively analyzing the extracted polarization parameter features, background area features marked by the visual attention mechanism, and global uniqueness features. Through weighted fusion, various features are comprehensively processed through weighted coefficients to build an adaptive decision fusion model; feature weights are dynamically adjusted through adaptive algorithms to output target contour features.

5. The UAV vision-guided obstacle avoidance system based on polarization information according to claim 4, characterized in that: Based on the target contour features, the logical process of performing preliminary path search through the path planning algorithm is as follows: Build a tree structure with the starting point as the root node; In free space, vertices are added by random sampling to expand the tree structure; Reconnect the tree by minimizing the cost of the newly added vertex and the nearby vertices; After finding a preliminary feasible solution, a local ellipsoid optimization region of the solution is generated.

6. The UAV vision-guided obstacle avoidance system based on polarization information according to claim 5, characterized in that: The specific process of performing obstacle avoidance judgment and centralized optimization sampling for preliminary path search based on the global feature distribution and the marked obstacle target background area is as follows: Calculate the cost of the preliminary path. When the cost is lower than the set threshold, enter the centralized optimization stage. The local space of the solution is sampled in the ellipsoidal region to further refine the path; Update the tree structure, connect new low-cost vertices, optimize the path, compare the updated path with the preliminary path, and choose the one with the lowest cost as the optimal flight path.

7. A method for obstacle avoidance by visual guidance of a UAV based on polarization information, applied to the system for obstacle avoidance by visual guidance of a UAV based on polarization information according to any one of claims 1 to 6, characterized in that: The following steps are involved: S1. Collect polarization image information in the flight scene through a polarization sensor to obtain images with different polarization directions; S2. Extract feature information from images with different polarization directions through compressed observation, set extraction termination conditions, and obtain polarization parameters; S3. Perform multi-level clustering on the polarization parameters, extracting the globally unique and spatially distributed characteristics of the polarization degree layer by layer. Use the visual attention mechanism to identify and label the obstacle, target, and background regions. Based on the global feature distribution and the labeled obstacle, target, and background regions, construct an adaptive decision fusion model, optimize feature weights, and output target contour features. S4. Based on the target profile features, a path planning algorithm is used to perform a preliminary path search. Obstacle avoidance is determined based on the global feature distribution and the marked obstacle target background area. The preliminary path search is then sampled and optimized, resulting in the output of the optimal flight path. S5. Smooth the optimal flight path, remove redundant inflection points, and parse the smoothed flight path into a waypoint sequence.

Images