AI-assisted unmanned aerial vehicle low-altitude flight obstacle avoidance method and system

Abstract

The invention relates to the technical field of unmanned aerial vehicle obstacle avoidance, in particular to an AI-assisted unmanned aerial vehicle low-altitude flight obstacle avoidance method and system. The method comprises the following steps: carrying out real-time unmanned aerial vehicle low-altitude flight environment perception based on a multi-modal perception network, and carrying out static obstacle identification and obstacle radiation range analysis to obtain a plurality of static obstacle radiation paths; carrying out dangerous potential energy field modeling based on the plurality of static barrier radiation paths, carrying out dangerous environment distribution fitting, and constructing a three-dimensional dangerous potential energy map; performing multi-path flight rehearsal according to the three-dimensional danger potential energy map, performing optimal flight path evaluation, and extracting an optimal flight path; and performing unmanned aerial vehicle flight processing based on the optimal flight path, performing dynamic obstacle visual identification and operation path prediction, and constructing a plurality of obstacle movement prediction paths. According to the invention, the task execution efficiency and safety of the unmanned aerial vehicle are improved through rapid and accurate obstacle avoidance decision making of the unmanned aerial vehicle.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle (UAV) obstacle avoidance technology, and in particular to an AI-assisted unmanned aerial vehicle (UAV) low-altitude flight obstacle avoidance method and system. Background Art

[0002] As drones are widely used for tasks such as logistics and transportation, emergency rescue, and environmental monitoring, low-altitude flight, as their normal operating mode, has gradually exposed a series of safety and efficiency challenges. Especially in densely populated urban areas, complex terrain, or wooded environments, drones face a large number of unpredictable obstacles during low-altitude flight, such as power lines, tree branches, building protrusions, and flying birds. These obstacles are diverse and frequently change, which can easily lead to collisions, resulting in equipment damage, mission failure, and even risks to personnel and property safety. Therefore, how to achieve precise obstacle avoidance for drones in complex and changing low-altitude environments has become one of the core challenges hindering their widespread application.

[0003] Traditional drone obstacle avoidance technologies primarily rely on hardware solutions such as visual cameras, which collect environmental information around the flight path to achieve basic obstacle avoidance. While these methods can be effective in static or well-structured environments, they often experience slow response and incomplete recognition when faced with the diverse, dynamic, and unstructured obstacles associated with low-altitude flight, making it difficult to meet the dual requirements of real-time performance and accuracy. Furthermore, existing obstacle avoidance mechanisms often lack environmental understanding and autonomous decision-making capabilities, relying excessively on preset rules and manual parameter adjustments. These mechanisms are unable to adapt to the changing conditions in complex scenarios, and safety performance is difficult to guarantee. Against this backdrop, there is an urgent need for a low-altitude flight obstacle avoidance method with intelligent perception, dynamic decision-making, and self-learning capabilities. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention proposes an AI-assisted low-altitude flight obstacle avoidance method and system for unmanned aerial vehicle (UAV) to solve at least one of the above technical problems.

[0005] To achieve the above objectives, the present invention provides an AI-assisted low-altitude flight obstacle avoidance method for unmanned aerial vehicles (UAVs). The UAV includes a multimodal perception network, which includes a visual camera, a laser radar, a millimeter-wave radar, and an UAV inertial measurement sensor. The method includes the following steps: Step S1: Based on the multimodal perception network, real-time UAV low-altitude flight environment perception is performed, and static obstacle recognition and obstacle radiation range analysis are performed, thereby multiple static obstacle radiation paths are obtained; Step S2: Modeling the hazard potential energy field based on the radiation paths of multiple static obstacles, fitting the hazard environment distribution, and constructing a three-dimensional hazard potential energy map; Step S3: Perform multi-path flight rehearsal based on the three-dimensional hazard potential energy map, evaluate the optimal flight path, and extract the optimal flight path; Step S4: executing UAV flight processing based on the optimal flight path, performing dynamic obstacle visual recognition and operation path prediction, and constructing multiple obstacle movement prediction paths; Step S5: Construct a three-layer obstacle avoidance decision architecture, select the optimal decision layer and make intelligent obstacle avoidance decisions based on the predicted movement paths of multiple obstacles, and build an intelligent adaptive obstacle avoidance strategy; Step S6: Upload to the shared obstacle avoidance cloud space according to the intelligent adaptive obstacle avoidance strategy, conduct collaborative learning obstacle avoidance and path optimization migration, and build a collaborative obstacle avoidance learning model.

[0006] In this specification, an AI-assisted UAV low-altitude flight obstacle avoidance system is provided, which is used to execute the above-mentioned AI-assisted UAV low-altitude flight obstacle avoidance method, including: The environmental perception module is used to perform real-time UAV low-altitude flight environment perception based on a multimodal perception network, and to identify static obstacles and analyze the obstacle radiation range, thereby identifying multiple static obstacle radiation paths; The hazard potential energy module is used to model the hazard potential energy field based on the radiation paths of multiple static obstacles, perform hazard environment distribution fitting, and construct a three-dimensional hazard potential energy map; The flight rehearsal module is used to conduct multi-path flight rehearsals based on the three-dimensional hazard potential energy map, evaluate the optimal flight path, and extract the optimal flight path; The movement prediction module is used to execute UAV flight processing based on the optimal flight path, perform dynamic obstacle visual recognition and operation path prediction, and construct multiple obstacle movement prediction paths; The adaptive obstacle avoidance module is used to build a three-layer obstacle avoidance decision-making architecture, select the optimal decision layer and make intelligent obstacle avoidance decisions based on the predicted movement paths of multiple obstacles, and build an intelligent adaptive obstacle avoidance strategy; The collaborative obstacle avoidance optimization module is used to upload to the shared obstacle avoidance cloud space based on the intelligent adaptive obstacle avoidance strategy, conduct collaborative learning obstacle avoidance and path optimization migration, and build a collaborative obstacle avoidance learning model.

[0007] The beneficial effects of this invention include: by integrating multiple modalities, including vision, radar, and depth sensing, the accuracy of identifying static obstacles (such as trees, power lines, and building edges) is significantly improved. A complete semantic understanding of the environment along the flight path is established, effectively avoiding blind spots and occlusions caused by single sensors. Analysis of obstacle "radiation paths" quantifies the impact range, allowing for pre-planned detours and providing physical boundary references for subsequent path planning. Spatially modeling the impact zones of multiple obstacles in the form of "hazard potential" ensures physical meaning and spatial continuity in obstacle avoidance. The three-dimensional hazard potential map is equivalent to the "high-risk terrain map" in navigation maps, enhancing the spatial intelligence of the path assessment system. This modeling approach is highly adaptable to sudden changes in terrain, densely built-up areas, or flight environments in forested areas, supporting safe modeling in complex terrain. Multiple possible paths are simulated in parallel, their safety and feasibility assessed, and the path with the lowest hazard potential is selected. The optimal path assessment not only considers safety but also incorporates multiple factors, such as flight time, energy consumption, and camera angle, for weighted optimization. Pre-determining the primary path through a pre-running process reduces the frequency of obstacle avoidance calculations during subsequent flight and improves processing efficiency. The system can identify dynamic obstacles such as pedestrians, vehicles, birds, and windblown branches, significantly improving its practicality at low altitudes. Instead of passive avoidance, it proactively adjusts its path based on the target's motion trends, increasing its success rate. This ensures flight safety in scenarios with frequent dynamic interference, such as densely trafficked areas, campuses, and construction sites. Its three-layer architecture (i.e., perception layer → decision layer → control layer) achieves a system-wide linkage of "fast response, precise strategy, and stable execution." When the primary obstacle avoidance strategy fails or is delayed, it quickly switches to the emergency strategy layer, enhancing system robustness. Different obstacle types trigger different strategy templates, demonstrating strategy adaptability and supporting differentiated response mechanisms. Obstacle avoidance experience from each drone is aggregated and uploaded to form a "group learning platform," rapidly improving the system's overall obstacle avoidance capabilities. Path optimization results can be shared with other drones flying in the same area, eliminating the need for repeated obstacle avoidance learning. The system features continuous updates and gradual optimization capabilities, becoming increasingly intelligent with use and continuously improving its ability to cope with complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 This is a schematic diagram of the steps of an AI-assisted UAV low-altitude flight obstacle avoidance method of the present invention; Figure 2 Detailed implementation flow chart of step S1; Figure 3 Detailed implementation flow chart of step S2; Figure 4 Schematic diagram of the detailed implementation steps of step S3. DETAILED DESCRIPTION

[0009] 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.

[0010] This application provides an AI-assisted low-altitude drone flight obstacle avoidance method and system. The execution entities of the AI-assisted low-altitude drone flight obstacle avoidance method and system include, but are not limited to, the following: mechanical equipment, data processing platforms, cloud server nodes, network upload devices, etc. equipped with the system, which can be regarded as general computing nodes of this application. The data processing platform includes, but is not limited to, at least one of an audio and image management system, an information management system, and a cloud data management system.

[0011] See also Figures 1 to 4 The present invention provides an AI-assisted UAV low-altitude flight obstacle avoidance method, comprising the following steps: Step S1: Based on the multimodal perception network, real-time UAV low-altitude flight environment perception is performed, and static obstacle recognition and obstacle radiation range analysis are performed, thereby multiple static obstacle radiation paths are obtained; Step S2: Modeling the hazard potential energy field based on the radiation paths of multiple static obstacles, fitting the hazard environment distribution, and constructing a three-dimensional hazard potential energy map; Step S3: Perform multi-path flight rehearsal based on the three-dimensional hazard potential energy map, evaluate the optimal flight path, and extract the optimal flight path; Step S4: executing UAV flight processing based on the optimal flight path, performing dynamic obstacle visual recognition and operation path prediction, and constructing multiple obstacle movement prediction paths; Step S5: Construct a three-layer obstacle avoidance decision architecture, select the optimal decision layer and make intelligent obstacle avoidance decisions based on the predicted movement paths of multiple obstacles, and build an intelligent adaptive obstacle avoidance strategy; Step S6: Upload to the shared obstacle avoidance cloud space according to the intelligent adaptive obstacle avoidance strategy, conduct collaborative learning obstacle avoidance and path optimization migration, and build a collaborative obstacle avoidance learning model.

[0012] In the embodiment of the present invention, see Figure 1 , is a schematic flow chart of the steps of an AI-assisted low-altitude flight obstacle avoidance method for a UAV according to the present invention. In this example, the steps of the AI-assisted low-altitude flight obstacle avoidance method for a UAV include: Step S1: Based on the multimodal perception network, real-time UAV low-altitude flight environment perception is performed, and static obstacle recognition and obstacle radiation range analysis are performed, thereby multiple static obstacle radiation paths are obtained; In this embodiment, real-time environmental perception is achieved by deploying a multimodal perception network that integrates four sensor modules: an RGB-D camera, a LiDAR (Lidar), a millimeter-wave radar, and an ultrasonic sensor. The RGB-D camera operates at a 30Hz frequency, with an effective detection range of 0.5-50 meters. It is primarily responsible for texture recognition and geometric outline extraction of static obstacles. The LiDAR uses a 360-degree scanning pattern, an angular resolution of 0.25 degrees, and a ranging accuracy of ±3cm, enabling precise measurement of the spatial position and size of obstacles. The millimeter-wave radar operates in the 77GHz frequency band, has a detection range of 200 meters, and an angular coverage of ±60 degrees. It primarily identifies metallic obstacles such as buildings and vehicles. The ultrasonic sensor array consists of eight sensors, operates at a 40kHz frequency, has a detection range of 0.02-4 meters, and is used for close-range obstacle detection. The system uses a Kalman filter algorithm to fuse multi-sensor data, achieving time synchronization accuracy within 1 millisecond. Static obstacle recognition uses an improved YOLO v5 algorithm, achieving 92.5% accuracy and a processing speed of 45fps. Obstacle radiation range analysis is modeled based on the obstacle's geometry, material properties, and flight altitude. The radiation range of buildings is set to 1.5 times their actual size, while the radiation range of tree-like obstacles is set to 2 times to account for wind sway. The radiation range of slender obstacles like utility poles is set to 3 times their diameter. Through this process, the system can obtain the precise location, geometric parameters, and safe radiation range of all static obstacles in the flight environment in real time, providing basic data support for subsequent hazardous potential energy field modeling.

[0013] Step S2: Modeling the hazard potential energy field based on the radiation paths of multiple static obstacles, fitting the hazard environment distribution, and constructing a three-dimensional hazard potential energy map; In this embodiment, a three-dimensional hazard potential energy field model is constructed based on static obstacle radiation path data. The model divides the flight space into grid cells of 1 meter × 1 meter × 0.5 meter, and each grid cell is assigned a corresponding hazard potential energy value. The hazard potential energy calculation uses an improved Gaussian attenuation function. The obstacle hazard coefficient K value in the formula is set according to the obstacle type: K = 1.0 for buildings, K = 1.5 for high-voltage power lines, K = 0.8 for moving vehicles, and K = 0.6 for vegetation. The distance attenuation parameter σ is dynamically adjusted according to the flight speed of the drone: σ = 2.0 for low-speed flight (≤5m / s), σ = 3.0 for medium-speed flight (5-10m / s), and σ = 4.0 for high-speed flight (>10m / s). The potential energy field update frequency is set to 10Hz to ensure real-time response to environmental changes. The hazard environment distribution is fitted using a radial basis function neural network with 200 hidden layer nodes, a learning rate of 0.01, 5000 training iterations, and a fitting accuracy requirement of a mean square error of less than 0.001. The three-dimensional hazard potential energy map is stored using an octree data structure, achieving an 85% compression rate and a query time complexity of O(log n). The potential energy field gradient is calculated using the central difference method, with a gradient threshold set to 0.1 to identify the boundaries of the hazard zone. The system also incorporates a time decay mechanism. For the historical trajectory of moving obstacles, the hazard potential energy decays exponentially with a time constant of τ = 3 seconds to avoid overly conservative path planning. Through precise potential energy field modeling, the system can quantify the hazard level of each spatial point in the flight environment, providing a reliable decision-making basis for the intelligent path planning algorithm.

[0014] Step S3: Perform multi-path flight rehearsal based on the three-dimensional hazard potential energy map, evaluate the optimal flight path, and extract the optimal flight path; In this embodiment, a three-dimensional hazard potential map is used for multi-path flight rehearsal, and an improved Rapid Search Random Tree (RRT*) algorithm is used to generate candidate flight paths. The algorithm parameter settings include: search radius r = 15 meters, maximum number of iterations N = 10,000, step size δ = 2 meters, and target offset tolerance ε = 3 meters. To improve search efficiency, a heuristic offset strategy is introduced, and the target guidance probability is set to 0.3, so that the path search is more inclined to the target point. Path evaluation uses a multi-objective optimization function, including four evaluation indicators: path length L, total hazard potential U, path smoothness S, and flight time T. The weight coefficients are set to w1 = 0.3, w2 = 0.4, w3 = 0.2, and w4 = 0.1, respectively. Path smoothness is quantified by calculating the angular change between adjacent waypoints, and the angle change threshold is set to 30 degrees. The system generates 8 candidate paths in parallel, and the calculation time for each path is controlled within 200 milliseconds. Path feasibility testing includes: dynamic constraint checking (maximum acceleration ≤ 3m / s², maximum angular velocity ≤ 60° / s), collision detection (safety distance ≥ 2 meters), and flight altitude limit (5-120 meters). The optimal path selection adopts an improved TOPSIS method, which ranks each candidate path by calculating the relative closeness to the ideal solution, and selects the path ranked first as the optimal flight path. Path optimization also takes into account the influence of the wind field, and fine-tunes the path based on real-time wind speed data (sampling frequency 20Hz), and the wind speed compensation coefficient is set to 0.15. The final output of the optimal flight path contains complete information such as waypoint coordinates, flight speed, flight altitude, and estimated arrival time, providing accurate navigation instructions for autonomous flight of the drone.

[0015] Step S4: executing UAV flight processing based on the optimal flight path, performing dynamic obstacle visual recognition and operation path prediction, and constructing multiple obstacle movement prediction paths; In this embodiment, drone flight control is performed based on the optimal flight path, and a real-time dynamic obstacle detection and prediction system is simultaneously activated. Flight control uses a cascade PID controller, with the position loop PID parameters set to Kp=2.5, Ki=0.1, and Kd=0.8, and the attitude loop PID parameters set to Kp=4.0, Ki=0.2, and Kd=1.2. The control frequency is 100Hz. Waypoint tracking accuracy requires a horizontal error of ≤1 meter, a vertical error of ≤0.5 meter, and a speed control accuracy of ≤0.2m / s. Dynamic obstacle detection uses a deep learning-based target detection algorithm, using an improved YOLOv7 model trained on a dataset of 20,000 aerial images, achieving a detection accuracy of 94.2%. Detection targets include dynamic obstacles such as birds, other drones, helicopters, and balloons, and the detection range is set to a sector-shaped area of ​​100 meters in front, 50 meters to the side, and 30 meters behind. The motion trajectory of dynamic obstacles is predicted using a long short-term memory network (LSTM). The network structure consists of two LSTM layers (with 128 and 64 hidden units, respectively) and one fully connected layer. The input historical trajectory length is 10 time steps, and the motion trajectory within the next 5 seconds is predicted. The mean absolute error (MAE) is used to evaluate the prediction accuracy, and the MAE is required to be ≤ 0.8 meters. The system establishes a dynamic obstacle motion model, considering three modes: constant speed linear motion, uniform acceleration motion, and curved motion, and determines the motion type through trajectory fitting. For biological targets such as birds, a random walk model is introduced, and the uncertainty coefficient is set to 0.3. The predicted paths of multiple obstacles are generated using the Monte Carlo method, generating 50 possible trajectories for each obstacle, with a confidence level set to 95%. The prediction results are updated in real time at a frequency of 5Hz to ensure rapid response to dynamic environmental changes.

[0016] Step S5: Construct a three-layer obstacle avoidance decision architecture, select the optimal decision layer and make intelligent obstacle avoidance decisions based on the predicted movement paths of multiple obstacles, and build an intelligent adaptive obstacle avoidance strategy; In this embodiment, a three - layer obstacle avoidance decision - making architecture is constructed, namely the global path planning layer, the local obstacle avoidance layer, and the emergency maneuver layer. The global path planning layer is responsible for long - term path planning. The planning time window is 30 - 60 seconds, the update frequency is 1Hz, and path replanning is performed based on static obstacle information and long - term prediction of dynamic obstacles. The local obstacle avoidance layer processes medium - and short - term obstacle avoidance decisions. The time window is 5 - 15 seconds, the update frequency is 10Hz, and it mainly deals with local environmental changes and short - term motion prediction of dynamic obstacles. The emergency maneuver layer executes immediate obstacle avoidance actions, with a response time ≤ 100 milliseconds and an update frequency of 50Hz, dealing with sudden dangerous situations. The decision - making layer selects based on the threat assessment index TI. The calculation formula considers four factors: obstacle distance d, relative speed v, collision probability P, and danger level H. The threshold is set as follows: when TI ≤ 0.3, the global layer is dominant; when 0.3 < TI ≤ 0.7, the local layer is dominant; when TI > 0.7, the emergency layer takes over. The intelligent obstacle avoidance decision uses a deep reinforcement learning algorithm, the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm. The network structure contains 3 hidden layers (the number of nodes is 256, 128, and 64 respectively), the learning rate is set to 0.001, and the discount factor γ = 0.99. The training environment contains 1000 different obstacle configuration scenarios, and each scenario is trained for 10000 time steps. The evaluation metrics for the obstacle avoidance strategy include success rate (≥ 95%), average obstacle avoidance time (≤ 3 seconds), path deviation (≤ 15%), and energy consumption efficiency (≤ 110% of the standard energy consumption). The system establishes an obstacle avoidance action library, which includes 12 basic actions such as ascending, descending, turning left, turning right, decelerating, and accelerating. The parameter range and execution time of each action are clearly defined. The intelligent adaptive obstacle avoidance strategy is continuously optimized through online learning. The learning rate adopts an annealing strategy, with an initial learning rate of 0.01, decaying 10% every 1000 steps, and a minimum learning rate of 0.0001.

[0017] Step S6: Upload the intelligent adaptive obstacle avoidance strategy to the shared obstacle avoidance cloud space for collaborative learning of obstacle avoidance and path optimization migration, and construct a collaborative obstacle avoidance learning model.

[0018] In this embodiment, the intelligent adaptive obstacle avoidance strategy is uploaded to the shared obstacle avoidance cloud space to establish a distributed collaborative learning system. The cloud space adopts a distributed architecture and is deployed on 5 server nodes in different geographical locations. Each node is equipped with a 16-core CPU, 64GB of memory and 1TB of SSD storage. Data transmission adopts an encrypted communication protocol, and the upload bandwidth requirement is ≥10Mbps and the latency is ≤50 milliseconds. The obstacle avoidance strategy data format adopts a standardized JSON format, which contains fields such as environmental parameters, obstacle information, obstacle avoidance action sequence, execution results, etc. The size of a single strategy file is controlled within 5KB. Collaborative learning adopts a federated learning framework, with each drone as a client and the cloud server as a parameter server. The global model update cycle is set to 24 hours. Each time the top 20% clients with the best performance are selected to participate in aggregation, and the aggregation weight is weighted averaged according to the client data volume and model performance. Path optimization migration adopts a transfer learning method. The similarity between the source domain and the target domain is calculated by the cosine distance of the environmental feature vector, and the similarity threshold is set to 0.8. Knowledge distillation technology is used for model compression, compressing complex cloud models (with 5 million parameters) into lightweight edge models (with 500,000 parameters), with accuracy loss controlled within 2%. The collaborative obstacle avoidance learning model uses an ensemble learning approach, integrating obstacle avoidance experience from different geographical environments and different flight scenarios, and making decisions through a weighted voting mechanism. The system establishes an obstacle avoidance knowledge graph that includes nodes such as obstacle type, environmental characteristics, obstacle avoidance strategy, and success rate, and uses graph neural networks for knowledge reasoning and strategy recommendation. The model performance is evaluated using a cross-validation method and verified on a test set containing 50,000 obstacle avoidance scenarios, requiring an average success rate ≥ 96% and an average response time ≤ 200 milliseconds.

[0019] In this embodiment, refer to Figure 2 , is a flowchart of the detailed implementation steps of step S1. In this embodiment, the detailed implementation steps of step S1 include: Based on the multimodal perception network, real-time UAV low-altitude flight environment perception is carried out to collect multimodal environment perception data; Perform iterative detection of environmental parameter noise on multimodal environmental perception data and mark environmental noise points; Adaptively optimize the filtering based on the environmental noise points to construct the filtering environment perception data; Perform multi-dimensional identification of static obstacles based on filtered environmental perception data and mark environmental static obstacle nodes; Calculating the real-time position and three-dimensional morphological parameters of the static obstacle nodes in the environment; Obstacle radiation range analysis is performed based on the real-time position and three-dimensional morphological parameters, thereby determining radiation paths of multiple static obstacles.

[0020] In this embodiment, drones face numerous visual obstructions, lighting interference, and obstacles with varying structures in low-altitude environments. The perception capabilities of a single sensor are insufficient to meet flight safety requirements. A multimodal perception network is constructed to comprehensively collect environmental information. This network integrates multiple sensors, including an RGB visible light camera, an infrared thermal imager, a laser radar (LiDAR), and a millimeter-wave radar, covering information collection requirements across different bands and dimensions. The RGB camera primarily provides color and texture information, the infrared thermal imager supplements heat source identification in low-light environments, the LiDAR collects 3D point cloud data of spatial structures, and the millimeter-wave radar ensures stable operation in adverse weather conditions such as rain and fog. Temporal synchronization and spatial calibration between sensors ensure data coordination. In actual deployment, the RGB camera operates at 30 fps, the LiDAR scans at 10 Hz, and the millimeter-wave radar operates in the 77 GHz frequency band. The multimodal perception network is controlled by a deep neural network architecture. A Transformer-based attention mechanism fuses the outputs of each sensor to extract key environmental features. By performing spatiotemporal alignment of the data and adjusting modal weights, the system ultimately outputs a unified environmental representation tensor for processing by subsequent modules. Experiments show that the system can stably acquire environmental perception data covering 360° with a maximum radius of 100 meters under various weather and lighting conditions, meeting the real-time and accuracy requirements for low-altitude flight. An iterative noise detection algorithm based on spatiotemporal density clustering and modal consistency checking is employed. Initial clustering is performed on the point cloud data using an improved DBSCAN algorithm to remove isolated points and low-density areas. A sliding window technique (with a window size of 5 frames) is then used to analyze the stability of feature points in the temporal dimension and detect anomalous data with significant fluctuations. Finally, the perception of the same physical point by each modality is compared and analyzed. Points that appear in infrared images but cannot be confirmed in RGB images, or frequently disappear in millimeter-wave radar, are identified as potential noise points. Experiments used a typical complex urban block as the test scene, collecting approximately 50,000 3D points and 1280x720 resolution RGB images per second. After noise detection, an average of approximately 12% of invalid or distorted perception data was removed per frame. This method significantly improves the subsequent modeling and recognition accuracy, reduces the false obstacle recognition rate, and effectively supports high-reliability environmental understanding of drones.

[0021] After environmental noise points are marked, the system needs to further process these anomalies to ensure greater stability and consistency in the perceived information. This step employs an adaptive filtering strategy, building a fusion filtering model based on the characteristics of multimodal data. The filtering core utilizes an improved multimodal weighted Kalman filter (Multi-mod AI Weighted Kalman Filter). This method builds on the prediction-update mechanism of the traditional Kalman filter by incorporating a modal confidence factor and a dynamic temporal adjustment mechanism. Each modality has a different signal-to-noise ratio in different environments. For example, RGB images have a higher weight during the day, while infrared modalities dominate at night or in foggy environments. Therefore, modal weights are automatically assigned based on the current scenario. For example, in clear skies, the initial weights for RGB, LiDAR, and infrared are 0.5, 0.3, and 0.2, respectively; in low-visibility environments, they are adjusted to 0.2, 0.4, and 0.4. By modeling the temporal sequence of environmental conditions, the filter predicts and corrects the state of each sensing point, eliminating sudden and abrupt changes. Running this filter on an experimental platform maintained data latency below 30ms, improved the continuity of the environmental model by 17%, and raised point cloud accuracy to ±4cm, significantly reducing track fluctuations caused by sensor noise. The resulting filtered data is a high-quality representation of the environment with clear structure, low noise, and strong continuity, providing solid data support for obstacle recognition.

[0022] Static obstacle recognition is a key component of drone obstacle avoidance systems. Its goal is to identify physical structures that could potentially impact flight safety, such as building walls, utility poles, and trees, from filtered, high-precision environmental data. This step employs multimodal feature extraction and semantic fusion recognition methods, combining the semantic features of visual images with the geometric features of point clouds to perform multi-dimensional identification of static obstacles. The system employs a dual-channel neural network architecture: the image channel employs a modified Mask R-CNN model to perform instance segmentation on RGB images, outputting the 2D boundaries and categories of obstacles; the point cloud channel employs the PointNet++ architecture to perform semantic classification and cluster analysis on spatial geometric structures. Subsequently, coordinate mapping is used to project objects in image space onto a 3D point cloud, forming a complete 3D semantic model of the obstacle. The system further labels each identified obstacle node with a type (e.g., "tree," "pole," or "wall") and assigns a confidence score. In an experimental setting, the system processed approximately 60 frames of imagery and 60,000 point cloud points per second within a real-world neighborhood with a 150-meter radius. It consistently identified all major static structures with an error rate below 5%, achieving an average recognition accuracy of 93.6%. The marked static obstacle nodes contained information such as their center point coordinates, outline boundaries, and category, providing key input for subsequent spatial modeling and path analysis.

[0023] After static obstacle identification and labeling, the next step is to perform structured modeling on these obstacles, extracting their specific spatial location and geometric shape. This modeling process not only forms the foundation for the obstacle avoidance algorithm but also provides a core support for trajectory planning and boundary assessment. This step utilizes a combination of spatial fitting and geometric reconstruction to create a 3D model of the obstacle point cloud data. First, the system clusters the point cloud in the area surrounding each obstacle node and uses voxel grid technology to compress the raw point cloud to optimal accuracy (approximately 1cm voxel resolution). The system then selects an appropriate fitting algorithm based on the obstacle's shape: cylindrical least squares fitting is used for columnar structures, planar fitting is used for wall surfaces, and Alpha Shape is used to generate an envelope for complex shapes. During the fitting process, the system also extracts parameters such as the center point coordinates (X, Y, Z), dimensions (height, width, thickness), and orientation angle of each obstacle. In the experimental environment, each obstacle consisted of approximately 4,000 points on average, with a fitting error within ±5 cm. The entire modeling process takes an average of approximately 45 milliseconds, meeting real-time update requirements. The output 3D parametric model will be stored in the global map cache for obstacle path analysis and dynamic obstacle avoidance decision making.

[0024] After obtaining the precise location and 3D structural parameters of an obstacle, the system further calculates the potential "flight exclusion zone" it creates in space—the obstacle's radiation radius. This analysis assesses the obstacle's potential impact on the drone's trajectory and constructs a path planning area that avoids its spatial impact. Radiation analysis first extends a safety buffer zone based on the obstacle's 3D boundaries. The buffer value is determined by the drone's dimensions (e.g., 0.6 meters wide, 0.4 meters high), flight speed (e.g., 6 m / s), and system response time (e.g., 200 ms). Typically, the system adds a safety radius of 1 to 1.5 meters to each obstacle, forming an obstacle "radiation volume." These volumes constitute a no-fly zone in 3D space. The system then maps these radiation volumes to occupancy states on a 3D grid map, representing the obstacle volume using voxels (Voxel Occupancy Map). Based on this occupancy map, the system runs a parallel path planning algorithm, including an A*-based search algorithm and a reinforcement learning path sampling method, to generate an obstacle avoidance route within the target area. In experiments, in an urban low-altitude flight zone measuring 80m x 80m x 30m, the system updated approximately 12 obstacle radiators per second and generated at least 15 feasible paths, with latency under 300ms. This mechanism significantly improved the safety and redundancy of flight paths, supporting stable flight control in complex environments.

[0025] In this embodiment, refer to Figure 3, is a flowchart of the detailed implementation steps of step S2. In this embodiment, the detailed implementation steps of step S2 include: Perform real-time dynamic rendering of the radiation paths of multiple static obstacles, visualize the radiation paths, and generate multiple radiation visualization paths; Identify the real-time motion parameters and flight planning paths of drones; Calculate the flight speed, altitude, acceleration and rate of change of the motion direction of the real-time motion parameters of the UAV to obtain the real-time multi-dimensional motion characteristics of the UAV; Perform flight modeling based on the real-time multi-dimensional motion characteristics of the UAV and the flight planning path to obtain a UAV flight twin model; Conduct multi-region flight hazard level assessment based on multiple radiation visualization paths to obtain multi-path regional hazard levels; According to the hazard level of the multi-path area, the hazard potential field of the UAV flight twin model is modeled, and the hazard environment distribution is fitted to construct a three-dimensional hazard potential map.

[0026] In this embodiment, the radial paths of multiple static obstacles are visualized in three dimensions, enabling real-time dynamic rendering of the obstacle paths. The rendering system uses a GPU-accelerated volume rendering algorithm, combined with 3D raster map data, to convert the radial range of each obstacle into spatial entities with visual attributes. Each obstacle is expanded into a set of colored blocks with varying transparency within the map, with different colors representing different safety levels (for example, red indicates a high-risk area, orange indicates a medium-risk area, and green indicates a flyable area). Simultaneously, the system performs a projection transformation on the radial paths based on the drone's flight direction, rendering the impact area that overlaps with the flight trajectory. To ensure real-time performance, the rendering module is deployed on the NVIDIA Jetson AGX Orin edge computing platform, with rendering latency kept to under 80ms. The experimental scenario consists of a simulated urban environment measuring 120m x 100m x 30m, containing an average of 45 static obstacles. The system generates clear and coherent radial visualization paths at a refresh rate of 15 frames per second. This visualization process not only enhances the system's spatial understanding of the flight safety boundary but also provides an observable dynamic input foundation for subsequent hazard level modeling and twin simulation. The system utilizes a high-precision IMU (Inertial Measurement Unit), GPS / RTK module, visual odometry (VIO), and internal flight control system feedback to identify current flight parameters. The IMU sampling frequency is set to 200Hz, outputting basic motion information such as the drone's acceleration, angular velocity, and heading angle. The GPS / RTK system provides centimeter-level position data at 10Hz per second, and the VIO module operates at 30fps to enhance the stability of position information in GPS-weak conditions. The system fuses these sensory information using a Kalman filter to output precise flight velocity (V), altitude (H), heading angle (θ), and position coordinates (X, Y, Z). Simultaneously, the flight control system outputs the planned flight path, including the waypoint sequence and corresponding flight constraints, such as speed restrictions and ascent and descent sections. Experimental tests in an outdoor urban environment with complex building occlusion demonstrated that the fused positioning system achieved an average error of less than 15cm, motion parameter identification latency of less than 50ms, and a planned path resolution accuracy of 98.5%. This module provides high-frequency, high-precision motion data stream input for subsequent flight modeling and obstacle avoidance reasoning, ensuring the model's comprehensive perception and response to flight status.

[0027] Using a time-series kinematic analysis method, the UAV's velocity (V), altitude (H), acceleration (A), and heading rate of change (Δθ) are continuously derived and statistically modeled. Specifically, flight velocity is calculated by differentially calculating three-dimensional coordinates, which are then decomposed into a velocity vector using timestamps. Acceleration is derived by superimposing the linear acceleration measured by the IMU with the flight velocity derivative. Altitude is determined by combining a barometer and a GPS / RTK dual-redundancy solution, achieving accuracy of ±0.2 meters. The heading rate of change represents the per-second trend in the UAV's heading angle, representing the complexity of its flight maneuvers. To further characterize the UAV's flight stability and risk trends, the system extracts the mean, extreme value, and coefficient of variation of these four types of motion features within a 0.5-second time window, constructing a dynamic multidimensional flight feature vector. In experiments, the system consistently outputs high-dimensional feature information under various complex flight missions (such as cornering, high-speed maneuvers, and sharp turn avoidance), processing an average of 60 sets of motion feature data per second with manageable error. Through this module, the system can accurately perceive instantaneous state changes during flight, providing a mathematical and quantitative basis for flight twin simulation and path risk modeling.

[0028] To achieve higher-level intelligent obstacle avoidance prediction and dynamic risk assessment, the system requires a real-time responsive "flight twin model." This model digitally maps the flight behavior of a real drone under specific mission conditions and can be used to predict path risks and deduce future flight behavior. Based on the drone's multi-dimensional motion characteristics and flight path information, a time-continuous flight trajectory prediction model is established by integrating dynamic flight data with behavioral pattern learning. This model utilizes a recurrent neural network (LSTM) architecture, taking the flight state at each moment as input to predict flight path deviation, speed change, and directional trends within the next 1-2 seconds. Each training round collects data including high-dimensional vectors such as flight speed, acceleration, attitude angle, and trajectory curvature, using a sliding window mechanism to maintain temporal consistency. The system updates its state 20 times per second, achieving a flight twin prediction accuracy of less than 0.4 meters within 10 frames. Experimental verification shows that under random wind disturbances, the twin model achieves a 95% or greater fit for predicted trajectories within the next 1 second. This model enables "future state" prediction and identification of abnormal trajectories, enabling proactive flight strategy adjustments. The establishment of a flight twin model is a prerequisite for constructing a hazardous potential energy field and assessing regional risks. Its high timeliness and accuracy directly determine the response speed and safety assurance capabilities of the obstacle avoidance system.

[0029] Building on the visualization of static obstacle radiation paths, the system needs to assess the hazard level of different spatial regions to further achieve quantitative flight safety management at the regional level. This step combines multiple radiation paths, the drone's current motion state, track density, obstacle density, and other factors using a multi-parameter risk aggregation model to assign a hazard level label to each spatial voxel. This is achieved through a scoring mechanism based on weighted risk indicators, taking into account factors such as obstacle density (ρ), radiation overlap coefficient (λ), flight path proximity (d), and the drone's current acceleration and maneuverability index (α). The system divides the entire flight space into 0.5m×0.5m×0.5m grid cells using a three-dimensional grid. For each cell, a hazard score R = f(ρ, λ, d, α) is calculated and classified into five hazard levels (very low, low, medium, high, and very high) based on thresholds. In experiments, this assessment mechanism was deployed in a simulated urban block scenario, achieving a hazard level classification accuracy of 92.7% and processing approximately 100,000 voxels per second. This process realizes regional-level dynamic risk management on the UAV flight path, providing a safety level reference for subsequent potential field modeling and autonomous path selection.

[0030] Based on multiple regional hazard levels, the system converts hazard information in the flight space into a "hazard potential field" with directional and gradient characteristics, and performs a three-dimensional fitting to construct a potential navigation map for drone routing decisions. The Risk PotentiAI Field is an abstract model that simulates natural physical fields. It transforms factors such as obstacles and high-risk areas into attractive / repulsive forces, guiding drones to avoid hazardous areas. During modeling, the system uses the predicted trajectory of the flight twin model as a "particle source" and superimposes multiple Gaussian distribution kernel functions in three-dimensional space based on hazard levels, generating a strong repulsive potential energy in high-risk areas. Low-risk areas are designated as zones with flat potential energy or attractive potential energy. The system uses B-spline surface fitting to interpolate and smooth the three-dimensional potential energy distribution map to ensure model continuity and physical consistency. In practice, constructing a 20m×20m×10m three-dimensional flight zone potential energy map takes approximately 150ms, with a spatial resolution of 0.2 meters. Experimental results show that the model effectively reflects the risky terrain characteristics in the flight space. Combined with dynamic path prediction, it can proactively identify potential collision zones and automatically adjust course. The three-dimensional hazard potential map is ultimately used to drive the path optimization algorithm, enabling the path to evolve in space toward low-risk areas. It is one of the core mechanisms for AI-assisted drones to achieve adaptive and safe flight.

[0031] In this embodiment, reference Figure 4 , is a flowchart of the detailed implementation steps of step S3. In this embodiment, the detailed implementation steps of step S3 include: Perform hazard attenuation calculations on three-dimensional hazard potential energy maps and identify environmental hazard attenuation indexes; Conduct multi-path flight rehearsals based on the environmental hazard attenuation index and extract multiple flight simulation paths; Calculating the obstacle collision probability of the flight simulation, and performing safe path screening to extract multiple candidate safe flight paths; Particle swarm exploration is performed on multiple candidate safe flight paths, and the optimal flight path is evaluated to extract the optimal flight path.

[0032] In this embodiment, after the three-dimensional hazard potential energy map is constructed, in order to achieve more accurate path dynamic evaluation and control response, it is necessary to perform "hazard attenuation" calculations on the potential energy values ​​of different areas in space. This step aims to identify the Risk Attenuation Index (RAI) of each spatial position in the potential energy map, which is used to reflect the rate at which the hazard field diffuses and disappears in space, thereby further supporting path preview and selection. In the specific implementation, the system adopts a Gaussian kernel diffusion model, regards each high-risk voxel as a source point, and calculates its potential energy diffusion intensity in space according to a certain time / distance attenuation law. The definition of the attenuation index refers to the physical attenuation model, and is in the form of , where β is the environmental attenuation coefficient (dependent on obstacle density and structural material), and d is the spatial distance from the obstacle center. To improve computational efficiency, the system employs a voxel traversal optimization algorithm. Leveraging a GPU parallel architecture, it can process attenuation value updates for over 150,000 voxel nodes per second. In experiments, five types of static obstacles (such as building walls, light poles, and trees) were deployed within a 30m x 30m x 10m flight zone. The system constructed an attenuation map within 120ms, achieving a 93.2% agreement between the average RAI distribution trend and the actual collision hotspot. By identifying the hazard attenuation index for each region, the system not only quantifies the spatial distribution trend of risk but also serves as an important input for subsequent flight path simulation and risk assessment.

[0033] After identifying hazard attenuation characteristics in space, the system uses this information to conduct multi-path flight rehearsals, simulating the UAV's dynamic flight behavior under different path selections and extracting a set of representative flight simulation paths. The goal of this step is to predict the potential flight outcomes in three-dimensional space under different flight strategies and attitude response conditions through scenario simulation, and to identify potential high-risk routes in advance. To implement this, the system first generates multiple sample paths within the navigation space near the current planned flight path. B-spline curves are used to connect the flight start and end points, while ensuring smoothness and dynamic feasibility. Each sample path is input as a "virtual flight trajectory" into the physical simulation module for rehearsal. The simulation engine integrates aerodynamic models, flight control models, and environmental disturbance models (such as wind speed and turbulence). During the simulation, the system reads the three-dimensional hazard potential energy map and the RAI index in real time, performs an integrated risk assessment on the spatial segments traversed by each path, and records the risk weight of each key point along the path. In the experimental platform, for a flight scenario involving 20 obstacles, the system can complete the simulation and analysis of no fewer than 30 paths within 500ms, outputting metrics such as flight time, risk score, and energy consumption. This module enables the system to proactively identify flight paths with significant risk concentrations, providing fundamental data support for selecting safer paths.

[0034] After completing dynamic rehearsals of multiple simulated paths, the system further analyzes the obstacle collision probability of each path to establish a quantitative assessment of path safety and extract a set of candidate safe flight paths for subsequent optimization. This step, based on a Bayesian risk model and spatial risk integration, comprehensively considers factors such as flight trajectory, spatial distribution risk, and flight control error to calculate the collision probability for each simulated path. The method divides the path into multiple time segments (each 200ms long). Within each segment, the position distribution of the drone's track within a 3D voxel and the corresponding potential energy value are calculated. Based on the drone's response error (e.g., trajectory offset σ = ±0.25 meters), a collision probability function P(collision) = ∑[V(x)·p(x|σ)] is established, where V(x) is the voxel risk value and p(x|σ) is the probability density function at that voxel after the offset. The system accumulates the total risk value for each path through simulation and sets a safety threshold (e.g., collision probability <5%). In experimental testing, path simulations were conducted in a complex, multi-obstacle environment. The system's collision assessment for each path averaged 30 milliseconds, resulting in the selection of 6 to 10 safe candidate paths (from an initial 30). These paths complement each other in spatial coverage, covering different altitudes and directional combinations, providing strategic redundancy for subsequent global path optimization. The generation of candidate paths ensures that the drone has multiple path options when faced with sudden environmental disturbances, significantly enhancing the system's robustness and obstacle avoidance flexibility.

[0035] To select a flight path with minimal risk, optimal energy efficiency, and smoothness from the candidate paths, the system employs the particle swarm optimization (PSO) algorithm to perform a global search and optimization to extract the optimal path. The PSO algorithm is an intelligent optimization method based on swarm collaboration and local optimal convergence, suitable for high-dimensional path planning problems. Each candidate path is encoded as an individual particle, and the particle's fitness function is designed as a multi-objective weighted function encompassing five core metrics: collision risk score, path length, flight energy consumption, flight time, and maneuverability cost (e.g., curvature change rate). During the initialization phase, the system randomly selects 510 paths from the candidate path set as the initial particle population. Subsequently, in each iteration, the path is fine-tuned based on the local optimal solution of the current particle and the global optimal solution of the swarm, simulating the response trends of the drone under different trajectory strategies in the same area. The PSO parameters set in the experiment are: population size N = 10, maximum number of iterations 30, inertia weight 0.6-0.9, and learning factor c1 = c2 = 1.8. In a test area containing 12 obstacles, the system completed path optimization in 1.2 seconds. The resulting optimal path, with a collision probability below 2%, reduced average energy consumption by 6.5%, shortened flight time by 3.1%, and improved path smoothness by 12%. This path not only geometrically avoids all high-risk areas but also exhibits strong stability and energy efficiency, providing a highly reliable flight execution path for the flight control system and serving as the final critical link in the closed-loop decision-making chain for AI-assisted obstacle avoidance.

[0036] In this embodiment, step S4 includes the following steps: Execute the flight of the UAV based on the optimal flight path and collect real-time flight monitoring video; Perform image enhancement and optimization on real-time flight monitoring videos to construct spatiotemporal consistency monitoring videos; Perform deep visual recognition of dynamic obstacles on spatiotemporal consistency monitoring videos, perform image frame segmentation, and extract obstacle image frames; Perform dynamic image frame optical flow tracking based on the obstacle image frame and extract multiple image frame optical flow trajectories; Perform obstacle motion feature mining on the optical flow trajectories of multiple image frames to generate obstacle motion feature parameters; Obstacle operation trend analysis is performed based on obstacle operation characteristic parameters, and operation path prediction is performed to construct multiple obstacle movement prediction paths.

[0037] In this embodiment, the drone is not only required to precisely fly along a predetermined path, but also to capture environmental video information during flight, providing basic image data for the subsequent dynamic obstacle perception and trend prediction modules. The flight execution module is led by a high-precision flight control system. After the path is imported, the flight attitude is controlled in real time based on flight control commands at 20 Hz per second, including roll, pitch, yaw, and thrust adjustments. To achieve high-frequency and stable sampling, the system is equipped with a high-definition surveillance camera (1920×1080 resolution, 60 fps) and an edge computing device (such as the NVIDIA Jetson Xavier NX) to simultaneously record flight footage during flight. The monitoring video data is cached in real time and uploaded to the image processing module in batches. A 1-minute circular buffer is also retained locally for fault recovery. In the experiment, an optimal path flight mission was executed in an open obstacle course measuring 80 m × 60 m × 20 m. The flight duration was 3 minutes and 20 seconds, with an average speed of 5.4 m / s. A total of 12 GB of video data was collected during the flight, and the frame loss rate was kept below 0.5%. This step ensures flight accuracy while providing high-quality visual raw materials for dynamic obstacle recognition, optical flow tracking, and path prediction, thus achieving a closed-loop connection between perception and control. The raw video images collected during flight are often affected by various interferences such as vibration, illumination changes, blur, and jitter, which are not conducive to the accuracy and stability of subsequent visual recognition. Therefore, the core goal of this step is to perform image enhancement and spatiotemporal consistency processing on real-time flight monitoring videos to improve image clarity and inter-frame continuity. The system first applies a temporal image stabilization algorithm to perform motion compensation on each frame, and eliminates frame skipping and displacement caused by flight jitter by estimating the translation and rotation vectors between images.

[0038] A lighting enhancement algorithm based on Retinex theory is used to compensate for brightness and enhance details in each image frame, making obstacle features clearer in both bright and dark areas. Furthermore, local gradient directional consistency detection within a three-frame temporal window is used to correct image edge drift caused by noise or blur. In the spatial domain, multi-scale high-frequency filtering combined with unsharp convolution is applied to enhance texture detail. In experimental testing, the optimization process improved the image structural similarity index (SSIM) by 16%, edge sharpness by 21%, and inter-frame variance by approximately 28% in 7200 frames of video captured during dynamic flight. This spatiotemporally consistent video output serves as a crucial visual foundation for subsequent dynamic obstacle depth recognition and motion prediction, significantly improving the system's visual robustness and recognition accuracy under rapid motion. Using the YOLOv8 model, which incorporates depth information, and combined with video streams captured by binocular vision or RGB-D sensors, obstacle detection and depth estimation are performed on each image frame. The system first detects objects, outputting the category, confidence score, and image bounding box coordinates for each obstacle. It then maps the 2D bounding boxes into 3D space through inter-frame matching of the point cloud data in the depth channel, estimating the depth and spatial scale of each obstacle's center point. To improve detection stability and avoid duplicate recognition, the system uses a temporal tracking mechanism (based on Kalman filtering and Hungarian matching) to maintain inter-frame consistency of the detection boxes. Experiments were conducted in a low-altitude street environment with dynamic elements such as pedestrians, bicycles, and vehicles. The YOLOv8 model achieved an average detection frame rate of 38 fps, an mAP@0.5 index of 87.4%, and an average depth estimation error of ±0.25 meters. During the 5-minute flight test, over 2,400 image frames were detected, effectively identifying all dynamic obstacles and generating corresponding image frames, providing accurate spatial input for subsequent optical flow tracking and trajectory modeling.

[0039] Based on the PyramidAI Lucas-Kanade (PyrLK) optical flow algorithm, sub-pixel motion vector estimation is achieved for key points within an image frame. Multiple sets of feature points (such as corners or edges) are automatically selected within each image frame and their displacement vectors are estimated across consecutive frames to derive the overall motion trajectory of the image frame. To prevent tracking failures or feature drift, the system introduces a motion consistency filter to remove feature points with abnormal motion (such as drift, jumps, overlapping, or occlusion). To achieve stable and high-frequency tracking, the image stream is processed at 30 fps, with a minimum of eight valid feature points per frame used for modeling, and the standard deviation of tracking error is maintained within 0.5 pixels. The system also combines depth differences between image frames to perform 3D reconstruction of the optical flow trajectory, mapping the 2D motion trajectory to a 3D trajectory. In experiments, the system tracked 112 dynamic image frames in a flight test scenario involving high-speed pedestrians and vehicles, with an average tracking time of 4.6 seconds and an error of within 0.3 meters. This module outputs a sequence of spatial trajectories for each obstacle in consecutive frames, a key step in achieving motion feature analysis and trend prediction. After completing continuous optical flow tracking of the image frame, the system further extracts the behavioral characteristics of the dynamic obstacle from each trajectory, providing a parameter basis for trend determination and path prediction. This step employs motion trajectory modeling and feature extraction methods to perform time series analysis and physical feature induction on the optical flow trajectories of the image frame. First, the system calculates multiple physical parameters for each trajectory, including velocity vector, acceleration change, path curvature, and average directional stability. It then introduces a motion classification mechanism, using a decision tree classifier to determine whether the obstacle is in linear motion (e.g., a straight pedestrian or a slow-moving vehicle) or nonlinear motion (e.g., random crossing or reversing). In terms of feature dimensions, the system selects eight key metrics, including maximum velocity (Vmax), average acceleration (Aavg), directional standard deviation (σθ), and average curvature (κavg), to construct the obstacle motion descriptor. To improve the robustness of the feature parameters, the system uses a five-frame sliding window for dynamic feature filtering to remove occasional abnormal changes. In experiments, 78 trajectories from four types of obstacles (pedestrians, electric vehicles, delivery trucks, and small unmanned vehicles) were modeled. The system extracted an average of 48 sets of features per second, achieving a recognition accuracy of 91.6%. This operational feature parameter set quantitatively represents obstacle behavior trends and provides highly available input features for subsequent path prediction models.

[0040] The system combines sequence modeling with trajectory prediction, integrating time series learning with physical modeling strategies. The system first uses a long short-term memory (LSTM) network as the primary trajectory prediction model. It inputs each obstacle's historical operational feature vector (such as speed, direction, and acceleration sequences) and trains it to output a sequence of predicted spatial positions within the next 1-3 seconds. To enhance the model's adaptability to complex behaviors, the system incorporates behavioral pattern labels (such as sudden stops, detours, and crossings) as auxiliary variables. Prediction results are output in time steps, constructing a spatial sequence of multiple possible future paths, forming a set of predicted obstacle movement paths. Furthermore, to ensure the confidence of the predicted paths, the system assigns a prediction confidence score (e.g., a probability distribution weight) to each path and uses Bayesian confidence bands to visualize the range of possible path variations. In experiments, the system used 300 real-world tracking trajectories as a training set, achieving an average prediction error within ±0.45 meters and a prediction latency of no more than 70ms. In tests conducted in a street intersection scenario, the system accurately predicted the movement trends of over 85% of obstacles. This step enables the system to know the possible movement direction and area of ​​dynamic obstacles in advance, realize dynamic path replanning and early risk avoidance, and is a key module for realizing the transformation of AI intelligent obstacle avoidance from "reactive" to "predictive".

[0041] In this embodiment, the specific steps of performing image enhancement optimization on the real-time flight monitoring video to construct the spatiotemporal consistency monitoring video are as follows: Decompose the global brightness optimization monitoring video into video frames, calculate the pixel brightness values ​​frame by frame, and generate pixel brightness values ​​for multiple image frames; Perform global video brightness average calculation based on the pixel brightness values ​​of multiple image frames to obtain the average video brightness; Perform brightness distribution analysis of the latest image frame based on the average brightness of the video to obtain the brightness distribution characteristics of the latest image frame; Perform adaptive brightness optimization on multiple regions of the image based on the brightness distribution characteristics of the latest image frame to obtain an adaptive brightness optimized video; The global video frame spatiotemporal consistency of the adaptive brightness optimized video is optimized to construct the spatiotemporal consistency monitoring video.

[0042] In this embodiment, the video decoding module decodes the input global brightness optimization monitoring video into frame-by-frame image data in chronological order. Each image frame maintains the original video resolution, such as 1920×1080 pixels, and a frame rate of 30 fps, ensuring temporal continuity and spatial detail integrity. The decoded single-frame image undergoes grayscale conversion, converting the color image into a single-channel brightness matrix with the brightness values ​​normalized between 0 and 255. This matrix is ​​then traversed pixel by pixel, and the brightness values ​​of all pixels are counted to form a corresponding two-dimensional brightness array. In an experimental environment, this process uses GPU acceleration to implement frame decomposition and grayscale conversion, ensuring that more than 30 frames per second can be processed with a latency of less than 30 milliseconds for real-time flight monitoring video acquisition. The resulting multi-frame pixel brightness data forms the critical raw data foundation for subsequent video brightness statistics and adaptive optimization. The system employs a sliding window statistical method, typically using a 5-second video frame sequence as the computation unit, or approximately 150 frames. The brightness of all pixels in each frame within the window is summed and averaged to obtain the average brightness value for that frame. The average brightness values ​​of all frames are then averaged again to obtain the global mean brightness value for the current video segment. This calculation smooths the effects of occasional illumination fluctuations and reflects long-term brightness trends. In experiments, the system analyzed video data collected by a low-altitude drone flight and found that the global mean brightness fluctuated within ±7% in stable lighting conditions, but could fluctuate up to ±15% in cloudy or partially shaded environments. This global mean brightness provides a robust global illumination benchmark for subsequent image brightness distribution analysis and adaptive optimization, ensuring the system's adaptability to complex natural lighting conditions.

[0043] The image frame is divided into multiple fixed-size regions, typically 8×8 or 16×16 grids, each containing a number of pixels. For each region, the mean, variance, and brightness histogram of the pixel brightness within it are calculated to reflect the illumination intensity and distribution characteristics of that region. Simultaneously, the system calculates a brightness histogram for the entire image and compares it with the global brightness mean calculated in the previous step to analyze brightness deviations in the current frame. This analysis helps identify dark, bright, and intermediate brightness regions, providing a basis for subsequent adaptive adjustments. In experiments, the system accurately identified up to 90% of abnormal brightness regions in various complex environments (such as changing shadows and sudden strong light), achieving rapid response to local illumination. The image frame brightness distribution feature vector typically has a dimension between 64 and 256, enabling a detailed description of the spatial structure of illumination. By partitioning the image into several subregions, an adaptive brightness adjustment factor is calculated for each subregion. Common methods include local histogram equalization (CLAHE) or regional brightness compensation based on the Retinex model. For each area, the system determines the enhancement or suppression coefficient based on the deviation between its brightness mean and the global mean. The contrast and brightness of low-light areas are enhanced, and overexposed areas are smoothed and suppressed. The optimization process adopts a weighted fusion strategy to avoid obvious separation between regional boundaries and ensure visual continuity. Experimental parameters show that the optimized video frames have an average contrast index improvement of more than 20%, and the brightness of local dark areas has increased by an average of 35%, while avoiding noise amplification. This step enables the monitoring video to more clearly display environmental details, especially in flight environments with alternating low light or strong light, greatly improving the reliability of dynamic obstacle detection.

[0044] To avoid inter-frame flickering and jumps introduced during the brightness optimization process, the system optimizes the spatiotemporal consistency of global video frames, ensuring video coherence and stability in both time and space. This step employs a smoothing algorithm based on temporal filtering. By analyzing the changing trends of the brightness adjustment factors between adjacent frames, a weighted average or recursive filter is applied for smoothing to reduce inter-frame brightness jumps. In the spatial dimension, a multi-scale edge-preserving filter is used to protect detailed edges and prevent blurring while smoothing discontinuous areas. Videos optimized for spatiotemporal consistency effectively suppress the "flicker effect" and improve viewing comfort. In experimental tests, this module reduced the standard deviation of inter-frame brightness differences by approximately 30%, and subjective evaluations showed a significant improvement in the visual stability of continuous flight videos. The resulting spatiotemporal consistency monitoring video provides stable and high-quality visual input for subsequent obstacle detection, tracking, and flight path planning, making it a key component in ensuring safe low-altitude drone flight monitoring.

[0045] In this embodiment, step S5 includes the following steps: Define the global path planning layer, local obstacle avoidance layer, and emergency maneuvering layer to build a three-layer obstacle avoidance decision-making architecture; Based on the three-layer obstacle avoidance decision architecture, the real-time flight status and danger level of multiple obstacle movement prediction paths are evaluated, and the optimal decision layer is selected to extract the optimal obstacle avoidance decision layer; Make intelligent obstacle avoidance decisions based on the optimal obstacle avoidance decision layer and build an intelligent adaptive obstacle avoidance strategy.

[0046] In this embodiment, to ensure flexible, real-time, and safe UAV obstacle avoidance in complex and volatile low-altitude environments, the system constructs a hierarchical obstacle avoidance decision-making architecture, divided into three functional modules: a global path planning layer, a local obstacle avoidance layer, and an emergency maneuvering layer. The global path planning layer primarily constructs a macro-path plan based on mission objectives and static environmental data (such as fixed obstacles like buildings and trees). It uses the A* algorithm, RRT (Rapid Random Tree), or Bezier curves to generate paths, ensuring the shortest overall flight distance and optimal energy consumption. The local obstacle avoidance layer focuses on dynamic factors within short and medium distances, such as pedestrians and moving vehicles, and senses and adjusts local flight direction and speed in real time. Common algorithms include the dynamic windowing algorithm (DWA) and the artificial potential field (APF). The emergency maneuvering layer handles sudden dangerous situations, such as approaching UAVs at high speed or unexpected obstacles, triggering evasive maneuvers such as emergency stops, sideways maneuvers, or ascents with millisecond-level response speeds. It is typically driven by state machine control logic and integrates sensors such as IMUs, vision sensors, and time-of-flight radars for coordinated control. The system adopts a multi-threaded architecture with module decoupling. Each layer independently processes perception and decision-making tasks, and is uniformly scheduled through the main control module.

[0047] Based on a three-layer obstacle avoidance architecture, the system dynamically determines the risk level of the current flight state based on the predicted paths of multiple obstacles encountered during flight, and selects the optimal response level. The specific process involves: First, the system analyzes the intersection of the predicted obstacle path and the drone's current and future states (position, velocity, and acceleration). It calculates the nearest flight intersection point within the next 1-3 seconds and further estimates the collision probability at each intersection point (based on a Gaussian error model and spatiotemporal overlap). Then, based on risk classification criteria, each assessment result is categorized into three levels: low risk (<30%), medium risk (30%-70%), and high risk (>70%). If the assessment result is low risk, the system continues to use the global path planning layer; if the assessment result is medium risk, the local obstacle avoidance layer is activated for fine-tuning the path; if the assessment result is high risk, the emergency maneuvering layer is immediately activated to implement evasive maneuvers such as hovering, sideways flight, or return home. This multi-layered risk response mechanism incorporates a Bayesian decision network to implement logical inference for hierarchical judgments. It also uses a convolutional neural network (CNN) to identify potential occluded obstacles in image frames, ensuring comprehensive risk assessment. In experiments, the system performed real-time assessments of the predicted paths of 32 moving obstacles in a dynamic cross-flight test scenario. The average hierarchical switching decision time was 48ms, and the misjudgment rate was kept to 2.7%, significantly improving the drone's safety and fault tolerance during flight.

[0048] After determining the optimal obstacle avoidance decision layer, the system executes the corresponding obstacle avoidance maneuvers. Its core goal is to intelligently adapt to changing flight environments and achieve real-time, safe obstacle avoidance while ensuring mission completion. Differentiated obstacle avoidance strategies are designed for each layer. In the global path planning layer, the strategy prioritizes graph-optimized path reconstruction algorithms, such as those based on the Dijkstra weight update model, to make large-scale adjustments to the flight path. In the local obstacle avoidance layer, the system uses lidar, time-of-flight depth sensors, and optical flow cameras in real time to obtain a local spatial model. It then uses artificial potential field methods and Bessel interpolation to predict the path direction, enabling variable-speed and adjustable-angle flight fine-tuning maneuvers. In the emergency maneuvering layer, the system utilizes a fast-response state machine (FSM) design with nine preset avoidance modes, including upward jumps, lateral sidesteps, and fixed-point emergency stops. These modes are prioritized based on the current available space and kinetic energy. For example, if a collision probability of greater than 90% is detected within 0.5 seconds, the system selects a combined upward jump and emergency stop, with attitude recovery occurring within 1 second. To achieve adaptive features, the system incorporates reinforcement learning modules (such as DQN) to provide real-time feedback training on the consequences of each obstacle avoidance action, optimizing the parameters of the obstacle avoidance strategy. Experiments have shown that in complex occlusion and dynamic intersection environments, this intelligent adaptive obstacle avoidance strategy can improve the average obstacle avoidance success rate to 96.5%, with a flight path deviation rate of less than 7.2%, effectively achieving a dynamic balance between path efficiency and flight safety.

[0049] In this embodiment, step S6 includes the following steps: Perform low-altitude obstacle avoidance operations based on intelligent adaptive obstacle avoidance strategies; Calculate the success rate of the low-altitude obstacle avoidance operation and analyze obstacle semantics, path selection and decision-making layer selection strategy; Upload the success rate, obstacle semantics, path selection, and decision-making layer selection strategy to the shared obstacle avoidance cloud space; Based on the shared obstacle avoidance cloud space, collaborative learning obstacle avoidance and path optimization migration are carried out to build a collaborative obstacle avoidance learning model.

[0050] In this embodiment, a previously developed multi-layer obstacle avoidance strategy is integrated with dynamic environmental perception results in real time to achieve precise avoidance of both static and dynamic obstacles during flight. During flight, the drone continuously synchronizes its multi-source sensors, including vision, lidar, IMU, and GNSS, to ensure continuous responsiveness to dynamic changes in obstacles. Whenever the flight path conflicts with the predicted obstacle trajectory, the system invokes the strategy selection module to determine the appropriate obstacle avoidance maneuver (such as turning, deceleration, or jumping), and uses the flight control unit to precisely control attitude and speed. The flight control system is set to a high-frequency response rate of 20Hz, with an obstacle avoidance decision frequency of 50ms per second, ensuring real-time and stable flight. In actual experiments, the drone completed a total of 80 obstacle avoidance missions in a typical urban park environment, encompassing both static obstacles (such as trees and utility poles) and dynamic obstacles (such as pedestrians and bicycles). Leveraging adaptive strategies, the system effectively avoided 78 potential collisions, demonstrating excellent online responsiveness and environmental adaptability. This represents a highly integrated implementation of AI-enabled intelligent decision-making for low-altitude operations. After completing an obstacle avoidance operation, the system evaluates the effectiveness of the entire process, focusing on quantifying the obstacle avoidance success rate and analyzing the semantic characteristics of obstacles, path selection behavior, and decision-making decisions. The obstacle avoidance success rate is calculated based on the criteria of "avoiding potential collisions while not triggering failure mechanisms." During this round of flight operations, the system successfully completed 78 effective avoidance maneuvers. In two unusual cases, the path failed due to sudden obstruction by obstacles, resulting in a 97.5% overall obstacle avoidance success rate. The system also semantically categorizes obstacles involved in the avoidance process into static (such as buildings and walls) and dynamic (such as electric vehicles and pedestrians), recording the recognition rate, response latency, and avoidance maneuver used for each obstacle type. Furthermore, based on path node change data, the system analyzes the frequency of path reconstruction and the degree of path deviation before and after obstacle avoidance to form a path selection preference model. In terms of decision-making layer analysis, the system records the triggering level (global, local, and emergency) of each obstacle avoidance event. It found that the local obstacle avoidance layer was activated most frequently, accounting for 63% of the time. The emergency maneuvering layer was triggered 9 times in high-risk environments, with an accuracy rate exceeding 95%. This multi-dimensional evaluation system not only reflects the effectiveness of the current strategy but also provides a detailed decision-making reference for subsequent model adjustments and path optimization.

[0051] To enable intelligent sharing and evolutionary learning of obstacle avoidance across missions and platforms, the system uploads obstacle avoidance assessment data generated during flight missions to a "shared obstacle avoidance cloud space." Built on a converged architecture of edge computing and cloud services, this space uploads local obstacle avoidance data (including success rates, obstacle semantic classifications, path node offsets, and decision-making records) to a cloud database via a 5G communication module. Data packets are formatted in JSON, supporting structured indexing and horizontal data fusion for efficient searchability. Before upload, the system desensitizes the data, retaining essential decision features and scene labels while removing location-specific privacy information to ensure data security. The upload mechanism utilizes a mission-cycle synchronization mechanism: uploads are performed after each complete obstacle avoidance mission, along with the current mission identifier, timestamp, and scene number for task-level clustering analysis in the cloud. During experimental flight tests on simulated urban streets, over 6,500 data nodes were uploaded, covering 47 obstacle semantic labels and 120 path response feature sets. A shared cloud platform aggregates obstacle avoidance data from multiple drones to form a generalizable knowledge base, supporting subsequent collaborative obstacle avoidance learning and scenario reconstruction, effectively advancing the system's evolution from isolated intelligence to swarm intelligence. After data upload, the system trains and transfers collaborative learning models within the shared obstacle avoidance cloud, enabling continuous evolution and improved adaptability of path planning capabilities. This collaborative learning model utilizes a combination of federated learning and transfer learning to ensure data privacy while enabling model knowledge sharing. On the cloud platform, obstacle avoidance experience uploaded by different drones is aggregated, and horizontal modeling is used to extract common features, such as avoidance strategies for specific semantic obstacles and path reconstruction logic for typical scenarios. The system utilizes reinforcement learning algorithms (such as PPO or TD3) to train the integrated obstacle avoidance policy network, while also incorporating a meta-learning mechanism to enable rapid adaptation to specific environmental changes in new tasks. During the path optimization migration phase, the system transfers parameters to new path planning tasks based on existing high-success policy models. Fine-tuning is then used to rapidly adjust to new environmental data, ensuring model transferability across different cities or flight missions. Experimental tests show that the collaborative learning model-based path planning strategy reduces average planning time by 22% in new obstacle avoidance scenarios compared to the original single-strategy model, increases the obstacle avoidance success rate to 98.3%, and effectively improves the overall robustness and adaptability of the system. This model marks a new stage in the AI ​​obstacle avoidance system's transition from local intelligence to collective cognition, forming the foundational capability architecture for future low-altitude digital transportation networks.

[0052] In this embodiment, an AI-assisted UAV low-altitude flight obstacle avoidance system is provided, which is used to execute the above-mentioned AI-assisted UAV low-altitude flight obstacle avoidance method, including: The environmental perception module is used to perform real-time UAV low-altitude flight environment perception based on a multimodal perception network, and to identify static obstacles and analyze the obstacle radiation range, thereby identifying multiple static obstacle radiation paths; The hazard potential energy module is used to model the hazard potential energy field based on the radiation paths of multiple static obstacles, perform hazard environment distribution fitting, and construct a three-dimensional hazard potential energy map; The flight rehearsal module is used to conduct multi-path flight rehearsals based on the three-dimensional hazard potential energy map, evaluate the optimal flight path, and extract the optimal flight path; The movement prediction module is used to execute UAV flight processing based on the optimal flight path, perform dynamic obstacle visual recognition and operation path prediction, and construct multiple obstacle movement prediction paths; The adaptive obstacle avoidance module is used to build a three-layer obstacle avoidance decision-making architecture, select the optimal decision layer and make intelligent obstacle avoidance decisions based on the predicted movement paths of multiple obstacles, and build an intelligent adaptive obstacle avoidance strategy; The collaborative obstacle avoidance optimization module is used to upload to the shared obstacle avoidance cloud space based on the intelligent adaptive obstacle avoidance strategy, conduct collaborative learning obstacle avoidance and path optimization migration, and build a collaborative obstacle avoidance learning model.

[0053] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0054] The foregoing description is intended only to provide specific embodiments of the present invention, which are intended to enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest manner consistent with the principles and novel features disclosed herein.

Claims

1. An AI-assisted UAV low-altitude flight obstacle avoidance method, characterized in that: The UAV includes a multimodal perception network, which includes a visual camera, a laser radar, a millimeter-wave radar, and an UAV inertial measurement sensor, and includes the following steps: Step S1: Based on the multimodal perception network, real-time UAV low-altitude flight environment perception is performed, and static obstacle recognition and obstacle radiation range analysis are performed, thereby multiple static obstacle radiation paths are obtained; Step S2: Modeling the hazard potential energy field based on the radiation paths of multiple static obstacles, fitting the hazard environment distribution, and constructing a three-dimensional hazard potential energy map; Step S3: Perform multi-path flight rehearsal based on the three-dimensional hazard potential energy map, evaluate the optimal flight path, and extract the optimal flight path; Step S4: executing UAV flight processing based on the optimal flight path, performing dynamic obstacle visual recognition and operation path prediction, and constructing multiple obstacle movement prediction paths; Step S5: Construct a three-layer obstacle avoidance decision architecture, select the optimal decision layer and make intelligent obstacle avoidance decisions based on the predicted movement paths of multiple obstacles, and build an intelligent adaptive obstacle avoidance strategy; Step S6: Upload to the shared obstacle avoidance cloud space according to the intelligent adaptive obstacle avoidance strategy, conduct collaborative learning obstacle avoidance and path optimization migration, and build a collaborative obstacle avoidance learning model.

2. The AI-assisted UAV low-altitude flight obstacle avoidance method according to claim 1 is characterized in that: The specific steps of step S1 are: Based on the multimodal perception network, real-time UAV low-altitude flight environment perception is carried out to collect multimodal environment perception data; Perform iterative detection of environmental parameter noise on multimodal environmental perception data and mark environmental noise points; Adaptively optimize the filtering based on the environmental noise points to construct the filtering environment perception data; Perform multi-dimensional identification of static obstacles based on filtered environmental perception data and mark environmental static obstacle nodes; Calculating the real-time position and three-dimensional morphological parameters of the static obstacle nodes in the environment; Obstacle radiation range analysis is performed based on the real-time position and three-dimensional morphological parameters, thereby determining radiation paths of multiple static obstacles.

3. The AI-assisted UAV low-altitude flight obstacle avoidance method according to claim 1 is characterized in that: The specific steps of step S2 are: Perform real-time dynamic rendering of the radiation paths of multiple static obstacles, visualize the radiation paths, and generate multiple radiation visualization paths; Identify the real-time motion parameters and flight planning paths of drones; Calculate the flight speed, altitude, acceleration and rate of change of the motion direction of the real-time motion parameters of the UAV to obtain the real-time multi-dimensional motion characteristics of the UAV; Perform flight modeling based on the real-time multi-dimensional motion characteristics of the UAV and the flight planning path to obtain a UAV flight twin model; Conduct multi-region flight hazard level assessment based on multiple radiation visualization paths to obtain multi-path regional hazard levels; According to the hazard level of the multi-path area, the hazard potential field of the UAV flight twin model is modeled, and the hazard environment distribution is fitted to construct a three-dimensional hazard potential map.

4. The AI-assisted obstacle avoidance method for low-altitude UAV flight according to claim 1 is characterized in that: The specific steps of step S3 are: Perform hazard attenuation calculations on three-dimensional hazard potential energy maps and identify environmental hazard attenuation indexes; Conduct multi-path flight rehearsals based on the environmental hazard attenuation index and extract multiple flight simulation paths; Calculating the obstacle collision probability of the flight simulation, and performing safe path screening to extract multiple candidate safe flight paths; Particle swarm exploration is performed on multiple candidate safe flight paths, and the optimal flight path is evaluated to extract the optimal flight path.

5. The AI-assisted UAV low-altitude flight obstacle avoidance method according to claim 1 is characterized in that: The specific steps of step S4 are: Execute the flight of the UAV based on the optimal flight path and collect real-time flight monitoring video; Perform image enhancement and optimization on real-time flight monitoring videos to construct spatiotemporal consistency monitoring videos; Perform deep visual recognition of dynamic obstacles on spatiotemporal consistency monitoring videos, perform image frame segmentation, and extract obstacle image frames; Perform dynamic image frame optical flow tracking based on the obstacle image frame and extract multiple image frame optical flow trajectories; Perform obstacle motion feature mining on the optical flow trajectories of multiple image frames to generate obstacle motion feature parameters; Obstacle operation trend analysis is performed based on obstacle operation characteristic parameters, and operation path prediction is performed to construct multiple obstacle movement prediction paths.

6. The AI-assisted UAV low-altitude flight obstacle avoidance method according to claim 5 is characterized in that: The specific steps of performing image enhancement optimization on the real-time flight monitoring video and constructing the spatiotemporal consistency monitoring video are as follows: Decompose the global brightness optimization monitoring video into video frames, calculate the pixel brightness values ​​frame by frame, and generate pixel brightness values ​​for multiple image frames; Perform global video brightness average calculation based on the pixel brightness values ​​of multiple image frames to obtain the average video brightness; Perform brightness distribution analysis of the latest image frame based on the average brightness of the video to obtain the brightness distribution characteristics of the latest image frame; Perform adaptive brightness optimization on multiple regions of the image based on the brightness distribution characteristics of the latest image frame to obtain an adaptive brightness optimized video; The global video frame spatiotemporal consistency of the adaptive brightness optimized video is optimized to construct the spatiotemporal consistency monitoring video.

7. The AI-assisted UAV low-altitude flight obstacle avoidance method according to claim 1 is characterized in that: The specific steps of step S5 are: Define the global path planning layer, local obstacle avoidance layer, and emergency maneuvering layer to build a three-layer obstacle avoidance decision-making architecture; Based on the three-layer obstacle avoidance decision architecture, the real-time flight status and danger level of multiple obstacle movement prediction paths are evaluated, and the optimal decision layer is selected to extract the optimal obstacle avoidance decision layer; Make intelligent obstacle avoidance decisions based on the optimal obstacle avoidance decision layer and build an intelligent adaptive obstacle avoidance strategy.

8. The AI-assisted UAV low-altitude flight obstacle avoidance method according to claim 1 is characterized in that: The specific steps of step S6 are: Perform low-altitude obstacle avoidance operations based on intelligent adaptive obstacle avoidance strategies; Calculate the success rate of the low-altitude obstacle avoidance operation and analyze obstacle semantics, path selection and decision-making layer selection strategy; Upload the success rate, obstacle semantics, path selection, and decision-making layer selection strategy to the shared obstacle avoidance cloud space; Based on the shared obstacle avoidance cloud space, collaborative learning obstacle avoidance and path optimization migration are carried out to build a collaborative obstacle avoidance learning model.

9. An AI-assisted UAV low-altitude flight obstacle avoidance system, characterized in that: The method for executing the AI-assisted low-altitude flight obstacle avoidance method of a UAV according to claim 1 comprises: The environmental perception module is used to perform real-time UAV low-altitude flight environment perception based on a multimodal perception network, and to identify static obstacles and analyze the obstacle radiation range, thereby identifying multiple static obstacle radiation paths; The hazard potential energy module is used to model the hazard potential energy field based on the radiation paths of multiple static obstacles, perform hazard environment distribution fitting, and construct a three-dimensional hazard potential energy map; The flight rehearsal module is used to conduct multi-path flight rehearsals based on the three-dimensional hazard potential energy map, evaluate the optimal flight path, and extract the optimal flight path; The movement prediction module is used to execute UAV flight processing based on the optimal flight path, perform dynamic obstacle visual recognition and operation path prediction, and construct multiple obstacle movement prediction paths; The adaptive obstacle avoidance module is used to build a three-layer obstacle avoidance decision-making architecture, select the optimal decision layer and make intelligent obstacle avoidance decisions based on the predicted movement paths of multiple obstacles, and build an intelligent adaptive obstacle avoidance strategy; The collaborative obstacle avoidance optimization module is used to upload to the shared obstacle avoidance cloud space based on the intelligent adaptive obstacle avoidance strategy, conduct collaborative learning obstacle avoidance and path optimization migration, and build a collaborative obstacle avoidance learning model.

Images