Intelligent cruise trajectory planning method for anti-interference unmanned aerial vehicle
By employing multimodal perception and fusion technology and real-time 3D modeling, the problem of insufficient environmental adaptability and anti-interference capability in UAV inspection was solved, generating efficient and reliable cruise trajectories and improving the efficiency and accuracy of power line inspection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-07
- Publication Date
- 2026-03-13
AI Technical Summary
Existing drone inspection solutions lack environmental adaptability, cannot cope with sudden obstacles, have poor anti-interference capabilities, and have unstable image quality, which affects defect identification.
Employing multimodal perception and lightweight front-end fusion technology, combining visual images and inertial measurement unit data, and using cross-modal attention neural networks for feature fusion, a 3D voxel map is constructed in real time. A lightweight target detection model is used to identify key targets, and an anti-interference cruise trajectory is generated through a trajectory optimization algorithm.
It achieves high-precision pose estimation and target recognition for UAVs in complex environments, reduces false alarm rate, generates stable and reliable cruise trajectories, and improves inspection efficiency and diagnostic reliability.
Smart Images
Figure CN121657692A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, and in particular to an intelligent cruise trajectory planning method for anti-jamming UAVs. Background Technology
[0002] Power distribution network systems cover a wide area and have a large number of devices. Traditional manual inspection methods are not only inefficient, but also greatly affected by factors such as terrain and weather, making it difficult to achieve comprehensive and timely inspection. With its advantages of high flexibility, high efficiency and no terrain restrictions, drone technology has become an important means of power distribution network inspection.
[0003] However, existing drones generate fixed flight paths by manually setting waypoints before the mission, and the drones perform inspection tasks according to the preset routes. However, this approach has obvious limitations: it lacks environmental adaptability and cannot cope with sudden obstacles (such as temporary construction equipment, tree growth, etc.) that may occur during the inspection process; it also has poor anti-interference capabilities, and the image quality is unstable under environmental interference such as wind disturbance and light changes, which affects the subsequent defect identification. Therefore, an anti-interference drone intelligent cruise trajectory planning method is proposed to solve this problem. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides the following technical solution: An anti-jamming drone intelligent cruise trajectory planning method includes the following steps: S1 Multimodal Perception and Lightweight Front-End Fusion: Acquires visual image sequences and inertial measurement unit data from the UAV; constructs a lightweight cross-modal attention neural network that utilizes front-end visual features. For the query vector, with inertial features Given key and value vectors, feature fusion is performed using an attention mechanism to output fused features that are robust to dynamic environmental disturbances. ; S2 Real-time 3D Modeling: Based on the fused features obtained in step S1 Using visual image sequences and dense simultaneous localization and mapping (DSM-MA) algorithms, the UAV pose is estimated in real time, and a real-time 3D voxel map with Euclidean distance field information of the inspection target is incrementally constructed. ; S3 Target Recognition and Screening: A lightweight target detection model is used to identify potential targets in real-time images to obtain a set B of potential targets; subsequently, each target in the potential target set B is projected onto the real-time 3D voxel map generated in step S2. Spatial consistency verification is performed to filter out false alarm targets with inconsistent spatial locations; then, correlation filtering is performed based on a preset task list; finally, information gain is used as the basis for the final step. Sort the remaining targets and filter out the list of key target locations. ; S4 Adaptive trajectory planning: using the real-time 3D voxel map generated in step S2 As an environmental constraint, the list of key target locations obtained in step S3 As a task-oriented approach, a trajectory optimization problem with dynamic constraints and multi-objective costs is constructed and solved through numerical optimization algorithms to generate an anti-interference intelligent cruise trajectory in real time.
[0005] As an improvement to the above technical solution, the cross-modal attention neural network in the multimodal perception and front-end lightweight fusion step, the fused features Obtained through the following methods: ,in, The j-th component representing the inertial characteristic, with weighting coefficients. Based on visual features With inertial characteristics The similarity determines whether the following conditions are met. .
[0006] As an improvement to the above technical solution, in step S2, the camera pose is estimated in real time by minimizing the joint error function. and update depth map The joint error function for: Where Ω is the set of valid pixels in the image; , These are the image intensities of the current frame and the previous frame, respectively; These are pixel coordinates; For camera model and estimated pose The image distortion function; , λ represents the depth maps of the current frame and the previous frame, respectively; ∇ is the spatial gradient operator; λ is the regularization coefficient.
[0007] As an improvement to the above technical solution, the trajectory optimization problem constructed in step S4 has an objective function It consists of the following cost function integrals: ; in, For parameterized drone trajectories; , These are the weighting coefficients for jerk and acceleration, respectively. Based on the real-time 3D voxel map The obstacle repulsion term calculated from the Euclidean notated distance field; Based on the list of key target locations The generated task attraction item.
[0008] As an improvement to the above technical solution, the obstacle repulsion term for: ; in, For discretized trajectory points; Given the Euclidean sign distance field function, query the distance from the point to the nearest obstacle; κ represents the preset absolute safety distance; κ is the weighting coefficient of the exclusion term.
[0009] As an improvement to the above technical solution, the task attraction item for: ; in, List of key target locations The k-th target point in the sequence; The estimated time for the trajectory to approach the target point; The attraction weight for the target point is calculated in step S3, and its value is the same as the information gain. Positive correlation, that is This makes targets with higher information gain more attractive to trajectories.
[0010] As an improvement to the above technical solution, the information gain in step S3 for: ; in, The feature vector extracted from target b, The prior feature vector of the normal target.
[0011] The beneficial effects of this invention are: By employing a lightweight cross-modal attention mechanism, visual and inertial data are dynamically fused, automatically increasing the reliance on IMU motion information when visual interference occurs (such as wind-induced blurring or sudden changes in lighting); weighting coefficients It is adaptively generated based on visual-inertial feature similarity, enabling "on-demand extraction" of key motion information, effectively suppressing noise, and outputting stable and reliable fused features. ; Based on a dense SLAM system driven by fusion features, a real-time 3D voxel map with Euclidean signed distance field (ESDF) is constructed. The joint error function simultaneously optimizes the robust photometric error and the depth map smoothing regularization term, effectively addressing illumination changes and dynamic object interference, ensuring high accuracy in pose estimation and map reconstruction; it does not rely on prior maps, enabling online incremental modeling and adapting to complex and dynamic inspection environments (such as power poles and wind interference scenarios). A four-level screening process of "2D detection + 3D spatial verification + task relevance filtering + information gain ranking" is adopted to significantly reduce the false alarm rate. An information gain Ig(b) based on cosine similarity anomaly is introduced to measure the degree of target anomaly by feature direction difference, which is robust to changes in illumination and scale. High information gain targets (such as suspected damaged insulators) are prioritized for inclusion in the task list. This enables "perception as decision-making," improving the targeting of inspections and the defect detection rate. A multi-objective optimization framework is constructed, taking into account dynamic smoothness, hard safety constraints, and task value orientation, to achieve a closed-loop intelligent response of "perceiving anomalies → increasing task weight → trajectory actively approaching". The generated trajectory can not only avoid interference and obstacles, but also actively approach high-value targets, significantly improving inspection efficiency and diagnostic reliability. Detailed Implementation
[0012] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0013] An anti-jamming drone intelligent cruise trajectory planning method includes the following steps: S1 Multimodal Perception and Lightweight Front-End Fusion: Acquires visual image sequences and inertial measurement unit data from the UAV; constructs a lightweight cross-modal attention neural network that utilizes front-end visual features. For the query vector, with inertial features Given key and value vectors, feature fusion is performed using an attention mechanism to output fused features that are robust to dynamic environmental disturbances. ; ,in, The j-th component representing the inertial characteristic, with weighting coefficients. Based on visual features With inertial characteristics The similarity determines whether the following conditions are met. ; Visual features As a query signal, it is used in conjunction with inertial features. Calculate the similarity of each component to generate attention weights. These weights determine the proportion of information extracted from inertial data. This process allows the model to dynamically and selectively extract motion information most relevant to the current visual scene (such as displacement information to compensate for image blur) from the inertial data, while suppressing irrelevant noise. Finally, a robust fusion feature is output through weighted summation. ; In other words, the drone captures image sequences and IMU data in real time, and the lightweight cross-modal attention network in the onboard computing unit starts working. When the visual image becomes blurred due to wind interference, the network uses a formula... Automatically adjust weights This gives greater importance to the inertial features that indicate camera shake, thereby outputting stable fusion features; S2 Real-time 3D Modeling: Based on the fused features obtained in step S1 Using visual image sequences and dense simultaneous localization and mapping (DSM-MA) algorithms, the UAV pose is estimated in real time, and a real-time 3D voxel map with Euclidean distance field information of the inspection target is incrementally constructed. Camera pose is estimated in real time by minimizing the joint error function. and update depth map The joint error function for: Where Ω is the set of valid pixels in the image; , These are the image intensities of the current frame and the previous frame, respectively; These are pixel coordinates; For camera model and estimated pose The image distortion function; , These are the depth maps of the current frame and the previous frame, respectively; ∇ is the spatial gradient operator; λ is the regularization coefficient; The error function contains two key terms. The robust photometric error term ensures that the pose estimation is insensitive to changes in illumination and dynamic objects; The depth map smoothing regularization term constrains the depth map gradient, ensuring the continuity and rationality of the reconstructed 3D geometric model. By minimizing this joint error, the system can simultaneously obtain accurate pose estimation and a consistent environment model. This method reduces the dependence on prior models and responds to environmental changes through online reconstruction. The generated ESDF map provides a real-time and accurate environmental representation for safety planning. That is, by using the output of S1, the system minimizes the joint error function. Real-time estimation of UAV pose and construction of 3D ESDF map of tower. Even under varying lighting conditions, the robust photometric error term ensures the accuracy of pose estimation; S3 Target Recognition and Screening: A lightweight target detection model is used to identify potential targets in real-time images to obtain a set B of potential targets; subsequently, each target in the potential target set B is projected onto the real-time 3D voxel map generated in step S2. Spatial consistency verification is performed to filter out false alarm targets with inconsistent spatial locations; then, correlation filtering is performed based on a preset task list; finally, information gain is used as the basis for the final step. Sort the remaining targets and filter out the list of key target locations. ; Information gain in step S3 for: ; in, The feature vector extracted from target b, The prior feature vector of the normal target; The cosine similarity anomaly is used to quantify the degree of anomalousness of a target. Its core lies in comparing the difference between the current target features and the normal prior features in terms of "direction" rather than absolute numerical difference. This makes it more robust to changes in illumination, scale, etc. The greater the difference, the higher the information gain, indicating that the target is more worthy of observation. This mechanism transforms the raw perception data into quantitative indicators with clear task value, enabling the system to intelligently focus on the most critical targets and effectively filter out false alarms and interference from irrelevant targets. That is, the target detection model identifies potential insulators, and after filtering out false alarms through a 3D map verification, The information gain of each insulator is calculated. A target with a characteristic orientation that differs significantly from that of a normal insulator is identified as a suspected defect. Its value was significantly higher than other targets, and it was given priority for inclusion. ; S4 Adaptive trajectory planning: using the real-time 3D voxel map generated in step S2 As an environmental constraint, the list of key target locations obtained in step S3 As a task-oriented approach, a trajectory optimization problem with dynamic constraints and multi-objective costs is constructed and solved through numerical optimization algorithms to generate an anti-interference intelligent cruise trajectory in real time. The constructed trajectory optimization problem has an objective function. for: ; in, For parameterized drone trajectories; , These are the weighting coefficients for jerk and acceleration, respectively. Based on the real-time 3D voxel map The obstacle repulsion term calculated from the Euclidean notated distance field; Based on the list of key target locations The generated task attraction item.
[0014] The obstacle rejection term for: ; in, For discretized trajectory points; Given the Euclidean sign distance field function, query the distance from the point to the nearest obstacle; κ represents the preset absolute safety distance; κ is the weighting coefficient of the exclusion term.
[0015] The task attraction for: ; in, List of key target locations The k-th target point in the sequence; The estimated time for the trajectory to approach the target point; The attraction weight for the target point is calculated in step S3, and its value is the same as the information gain. Positive correlation, that is This makes targets with higher information gain more attractive to trajectories; Smoothing term: By minimizing the higher-order derivatives of the trajectory (jerk and acceleration), the generated trajectory is guaranteed to conform to UAV dynamics, ensuring smooth flight. Obstacle repulsion term. A cubic penalty is applied when the distance between the trajectory point and the obstacle is less than the safe distance. At that time, the penalty value will rise sharply, creating a strong repulsive force to ensure hard safety constraints and task attractiveness. Information gain The attraction that transforms into a trajectory A closed loop of "perception-decision-planning" has been established. The higher the information gain of a target, the stronger its "attraction" to the UAV, guiding it to generate a trajectory that is more conducive to mission execution. This multi-objective optimization framework can balance safety, smoothness and mission efficiency in real time. The generated trajectory can dynamically respond to changes in the environment and mission requirements, achieving comprehensive anti-interference capabilities. That is, a set of potential targets B is obtained using a lightweight target detection model, and then projected onto the 3D map Mt generated by S2 for spatial consistency verification, filtering out false alarms, and constructing an objective function. Obstacle Rejection Term Ensure the trajectory maintains a safe distance from the pole; task attractiveness In the middle, the suspected damaged insulators were due to their high... Value to obtain high weight This generates a strong attraction, and by solving this problem, an optimal trajectory is ultimately generated that can both safely avoid obstacles and closely observe suspected defect points.
[0016] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it.
Claims
1. An anti-interference unmanned aerial vehicle (UAV) intelligent cruise trajectory planning method, characterized in that, Includes the following steps: S1 Multimodal Perception and Lightweight Front-End Fusion: Acquires visual image sequences and inertial measurement unit data from the UAV; constructs a lightweight cross-modal attention neural network that utilizes front-end visual features. For the query vector, with inertial features Given key and value vectors, feature fusion is performed using an attention mechanism to output fused features that are robust to dynamic environmental disturbances. ; S2 Real-time 3D Modeling: Based on the fused features obtained in step S1 Using visual image sequences and dense simultaneous localization and mapping (DSM-MA) algorithms, the UAV pose is estimated in real time, and a real-time 3D voxel map with Euclidean distance field information of the inspection target is incrementally constructed. ; S3 Target Recognition and Screening: A lightweight target detection model is used to identify potential targets in real-time images to obtain a set B of potential targets; subsequently, each target in the potential target set B is projected onto the real-time 3D voxel map generated in step S2. Spatial consistency verification is performed to filter out false alarm targets with inconsistent spatial locations; then, correlation filtering is performed based on a preset task list; finally, information gain is used as the basis for the final step. Sort the remaining targets and filter out the list of key target locations. ; S4 Adaptive trajectory planning: using the real-time 3D voxel map generated in step S2 As an environmental constraint, the list of key target locations obtained in step S3 As a task-oriented approach, a trajectory optimization problem with dynamic constraints and multi-objective costs is constructed and solved through numerical optimization algorithms to generate an anti-interference intelligent cruise trajectory in real time.
2. The anti-interference UAV intelligent cruise trajectory planning method according to claim 1, characterized in that: The cross-modal attention neural network in the multimodal perception and front-end lightweight fusion step, the fused features Obtained through the following methods: ,in, The j-th component representing the inertial characteristic, with weighting coefficients. Based on visual features With inertial characteristics The similarity determines whether the following conditions are met. .
3. The anti-interference UAV intelligent cruise trajectory planning method according to claim 2, characterized in that: In step S2, the camera pose is estimated in real time by minimizing the joint error function. and update depth map The joint error function for: Where Ω is the set of valid pixels in the image; , These are the image intensities of the current frame and the previous frame, respectively; These are pixel coordinates; For camera model and estimated pose The image distortion function; , λ represents the depth maps of the current frame and the previous frame, respectively; ∇ is the spatial gradient operator; λ is the regularization coefficient.
4. The anti-interference UAV intelligent cruise trajectory planning method according to claim 1, characterized in that: The trajectory optimization problem constructed in step S4 has an objective function. It consists of the following cost function integrals: ; in, For parameterized drone trajectories; , These are the weighting coefficients for jerk and acceleration, respectively. Based on the real-time 3D voxel map The obstacle repulsion term calculated from the Euclidean notated distance field; Based on the list of key target locations The generated task attraction item.
5. The anti-interference UAV intelligent cruise trajectory planning method according to claim 4, characterized in that: The obstacle rejection term for: ； in, For discretized trajectory points; Given the Euclidean sign distance field function, query the distance from the point to the nearest obstacle; κ represents the preset absolute safety distance; κ is the weighting coefficient of the exclusion term.
6. The anti-interference UAV intelligent cruise trajectory planning method according to claim 4, characterized in that: The task attraction for: ； in, List of key target locations The k-th target point in the sequence; The estimated time for the trajectory to approach the target point; The attraction weight for the target point is calculated in step S3, and its value is the same as the information gain. Positive correlation, that is This makes targets with higher information gain more attractive to trajectories.
7. The anti-interference UAV intelligent cruise trajectory planning method according to claim 4, characterized in that: Information gain in step S3 for: ； in, The feature vector extracted from target b, The prior feature vector of the normal target.