Low-altitude unmanned aerial vehicle sector risk grid autonomous obstacle avoidance control method

Abstract

The application discloses a kind of low-altitude unmanned aerial vehicle sector risk grid autonomous obstacle avoidance control methods, comprising the following steps: obtaining aircraft motion state and environmental perception data, pre-processing perception data, and marking as unknown area with no effective perception area;With body coordinate system as reference, according to azimuth, pitch angle, distance layer, divide several sectors, determine the nearest obstacle distance in each sector and unknown area identification;Tracking is implemented to time-sequential obstacles, and the subsequent sector occupation trend is predicted;Combine multi-dimensional parameter calculation to calculate the risk value of each sector, and the risk value corresponding to dynamic obstacles is determined according to the occupation trend;Generate candidate control instruction set under the constraint of flight dynamics;Calculate the comprehensive cost through trajectory deduction, and select the optimal execution instruction output;Execute the instruction, and return to the perception data acquisition and preprocessing step in the next control cycle;The application solves the problem of insufficient reliability of unmanned aerial vehicle autonomous obstacle avoidance in complex low-altitude environment.

Description

Technical Field

[0001] This invention relates to the technical field of autonomous control of unmanned aerial vehicles (UAVs), specifically to an autonomous obstacle avoidance control method for low-altitude UAVs using a sector risk grid. Background Technology

[0002] With the rapid development of the low-altitude economy, unmanned aerial vehicles (UAVs) are increasingly being used in urban patrol, logistics delivery, and security monitoring. The low-altitude flight environment is characterized by dense obstacles, varied terrain, and frequent movement of dynamic targets, placing extremely high demands on the autonomous obstacle avoidance capabilities of UAVs. The reliability of the autonomous obstacle avoidance control method directly determines the flight safety and mission execution efficiency of the UAV.

[0003] Currently, most common obstacle avoidance methods for unmanned aerial vehicles (UAVs) are based on single-point ranging results from distance sensors to trigger obstacle avoidance actions, or use simple two-dimensional grids for path planning. These approaches have the following shortcomings: First, they lack specialized risk marking and quantification for areas without effective perception data, such as sensor blind spots and over-range conditions, easily assuming unknown areas are passable, posing significant safety risks. Second, they only assess the risk of dynamic obstacles based on the current instantaneous position, lacking prediction and advance response to target movement trends, resulting in delayed obstacle avoidance response when encountering suddenly moving obstacles. Third, the risk assessment dimension is singular, relying solely on obstacle distance as the criterion, without incorporating multi-dimensional information such as relative approach speed, terrain height margin, perception confidence, and target hazard level, leading to insufficient accuracy in risk assessment in complex low-altitude scenarios. Fourth, the obstacle avoidance command generation stage does not fully incorporate flight dynamics constraints and mission trajectory tracking requirements, resulting in output control commands that easily exceed flight control execution capabilities, exhibit large command jitter, and deviate excessively from the mission path, making it difficult to balance flight safety, control smoothness, and mission execution efficiency.

[0004] In summary, existing obstacle avoidance technologies for low-altitude unmanned aerial vehicles are difficult to adapt to complex and dynamic low-altitude flight scenarios, and cannot simultaneously meet the multiple requirements of obstacle avoidance reliability, dynamic response capability, and engineering feasibility. Summary of the Invention

[0005] This invention provides an autonomous obstacle avoidance control method for low-altitude unmanned aerial vehicles (UAVs) based on sector risk grids to solve the aforementioned technical problems. Specifically, the technical solution is as follows: A method for autonomous obstacle avoidance control of a low-altitude unmanned aerial vehicle sector risk grid includes the following steps: S1. Acquire the motion state of the unmanned aerial vehicle and the environmental perception data of the surrounding environment, preprocess the environmental perception data and mark the areas without effective perception data as unknown areas; S2. Using the body coordinate system of the unmanned aerial vehicle as a reference, the surrounding space is divided into multiple sectors according to azimuth, pitch and range, and the distance to the nearest obstacle and the marking of unknown areas are determined according to the preprocessed environmental perception data. S3. Track obstacles with continuous temporal sequence and predict the occupancy trend of each obstacle in each sector in the future time period; S4. Calculate the risk value of each sector based on the distance to the nearest obstacle in each sector, the unknown area marker, the relative approach speed, the measurement confidence level, the terrain height margin, and the obstacle type. The risk value of the sector corresponding to the dynamic obstacle is determined based on the occupancy trend of the obstacle. S5. Generate a candidate control command set under the flight dynamics constraints of the unmanned aerial vehicle; S6. Perform trajectory deduction on the candidate control command set, calculate the comprehensive cost based on the risk value of the deduced trajectory in each sector and the deviation from the mission trajectory, and select the one with the best comprehensive cost as the execution command output. S7. Execute the execution instruction and return to S1 in the next control cycle.

[0006] Further, the preprocessing of the environmental perception data and the marking of areas without valid perception data as unknown areas includes: The environmental perception data is synchronized in time and transformed in spatial coordinates to be unified into the body coordinate system. Invalid points are removed and field of view validity is marked on the environmental perception data after unifying the coordinate system; Areas where no valid perception data falls after invalid point removal and field of view validity marking are marked as unknown areas.

[0007] Furthermore, the tracking of time-series continuous obstacles and the prediction of the occupancy trend of each obstacle in each sector over future time periods include: Obstacle clustering and inter-frame association are performed on the preprocessed environmental perception data to determine the temporal correspondence of the same obstacle; The process noise covariance of the Kalman filter is configured according to the motion properties of the obstacle; static obstacles are configured with a low noise level, and dynamic obstacles are configured with a high noise level. Using the time-series correspondence as the observation input, and the three-dimensional position and velocity of the obstacle in the body coordinate system as the state variables, recursive state estimation is performed through the configured Kalman filter. Based on the state estimation results at the current moment, the obstacle movement trajectory within the preset time domain is extrapolated forward, and the occupancy trend of each sector is determined according to the obstacle movement trajectory.

[0008] Further, the calculation of the risk value of each sector based on the nearest obstacle distance, the unknown area marker, the relative approach speed, the measurement confidence level, the terrain height margin, and the obstacle type includes: The collision time is determined based on the distance to the nearest obstacle and the relative approach speed. The measurement confidence level is determined based on the sensor echo signal-to-noise ratio and the effective field of view coverage. Calculate the terrain height margin based on the current flight altitude and the terrain height in the sector direction; The risk value of each sector is obtained by weighting and fusing the nearest obstacle distance, collision time, measurement confidence level, terrain height margin, unknown area marker, and obstacle type.

[0009] Furthermore, the sector risk value corresponding to the dynamic obstacle is determined based on the obstacle's occupancy trend, specifically including: For the dynamic obstacle, the risk value of each sector obtained by weighted fusion is used as the base risk value; Based on the occupancy trend, the base risk value of the sector predicted to be occupied by the dynamic obstacle in the future time period is modulated upwards; The base risk value of the sector currently occupied by the dynamic obstacle but no longer occupied in the future is down-modulated.

[0010] Furthermore, the generation of a candidate control command set under the flight dynamics constraints of the unmanned aerial vehicle includes: Discrete sampling is performed within the preset control command parameter space, and the sampled values ​​of each dimension are combined into multiple candidate command vectors; The candidate command vectors are verified one by one according to the flight dynamics constraints of the unmanned aerial vehicle; Candidate instruction vectors that do not meet any of the constraints are removed, and the remaining candidate instruction vectors are integrated into a candidate control instruction set.

[0011] Furthermore, the step of performing trajectory extrapolation on the candidate control command set, and calculating the comprehensive cost based on the risk value of the extrapolated trajectory in each sector and its deviation from the mission trajectory, includes: Using each candidate instruction in the candidate control instruction set as input, the corresponding predicted trajectory is recursively generated within a preset prediction time domain. By accumulating the risk value of the sectors traversed along each of the predicted paths, the risk cost corresponding to each candidate instruction is obtained. Calculate the deviation between the predicted end position of each trajectory and the corresponding target point on the current mission trajectory to obtain the trajectory deviation cost corresponding to each candidate command; The energy consumption cost is determined based on the degree of difference between each candidate command and the current flight state. The instruction smoothing cost is determined based on the degree of difference between each candidate instruction and the instruction executed in the previous cycle. The cost of traversing the unknown region is determined based on the proportion of steps that fall into the unknown region in the predicted flight path. The risk cost, the track deviation cost, the energy consumption cost, the command smoothness cost, and the unknown area crossing cost are weighted and summed to obtain the comprehensive cost corresponding to each candidate command.

[0012] Furthermore, the formula for calculating the comprehensive cost is as follows: ; Where b1 is the track deviation cost weighting coefficient, E track (u) represents the track deviation cost, b2 is the risk cost weighting coefficient, and E risk (u) represents the risk cost, b3 is the energy consumption cost weighting coefficient, and E energy (u) represents the energy consumption cost, b4 is the smoothness cost weighting coefficient, and E smooth (u) represents the smoothness cost of the instruction, b5 is the weighting coefficient for the cost of traversing unknown regions, and E unknown (u) represents the cost of traversing the unknown region.

[0013] Furthermore, when executing the execution instruction, a graded obstacle avoidance strategy is implemented according to the risk level: When the risk value reaches the first threshold, the target course is corrected by a preset angle. When the risk value reaches a second threshold that is greater than the first threshold, deceleration is performed and the lateral search weight is increased; When the risk value reaches a third threshold that is greater than the second threshold, if there is a low-risk passage to the side or above, a lateral detour or ascent shall be performed. If the risk values ​​of all sectors traversed by the candidate control instruction set are higher than the third threshold, then a security protection mode is executed.

[0014] Furthermore, the hierarchical obstacle avoidance strategy is configured with a hysteresis handling mechanism: There are separate action entry and action exit thresholds for each level of obstacle avoidance action; When executing the execution instruction, the obstacle avoidance action is triggered only when the number of control cycles in which the risk value continuously exceeds the corresponding risk level threshold reaches the action entry threshold. After an obstacle avoidance action is triggered, the current obstacle avoidance action is allowed to exit only when the number of control cycles in which the risk value is continuously lower than the preset release threshold reaches the action exit threshold. Wherein, the action exit threshold is greater than the action entry threshold.

[0015] The advantages of this invention lie in its use of a three-dimensional sector risk grid architecture in a body coordinate system. This architecture combines multiple parameters—including nearest obstacle distance, collision time, measurement confidence level, terrain height margin, unknown area marking, and obstacle type—to quantify sector risk, significantly improving the comprehensiveness and accuracy of risk assessment in complex low-altitude scenarios. Furthermore, by using obstacle time-series tracking and motion trend prediction, it enables targeted modulation of sector risks corresponding to dynamic obstacles, achieving early response to moving targets and effectively improving the problem of delayed obstacle avoidance response in dynamic scenarios. Finally, it marks perception blind spots and over-range areas. By incorporating unknown areas into risk calculations and trajectory cost constraints, the system avoids misjudging unperceived areas as safe passages, thus enhancing flight safety redundancy in unknown environments. Candidate control commands are generated under flight dynamics constraints, ensuring that all output commands remain within the flight control's executable range and mitigating the difficulty of implementing commands during rapid acceleration / deceleration or super-maneuvering maneuvers. Finally, through trajectory extrapolation and multi-objective comprehensive cost optimization, the system balances mission trajectory tracking accuracy, command smoothness, and energy consumption control while ensuring flight safety, thereby comprehensively improving the reliability and engineering adaptability of autonomous obstacle avoidance for low-altitude unmanned aerial vehicles. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the autonomous obstacle avoidance control method for low-altitude unmanned aerial vehicles sector risk grid according to the present invention; Detailed Implementation

[0018] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0019] like Figure 1 The diagram shows the autonomous obstacle avoidance control method for low-altitude unmanned aerial vehicles (UAVs) based on a sector risk grid, comprising the following steps: S1. Acquire the motion state of the unmanned aerial vehicle and the environmental perception data of the surrounding environment, preprocess the environmental perception data and mark the areas without effective perception data as unknown areas; S2. Using the body coordinate system of the unmanned aerial vehicle as a reference, the surrounding space is divided into multiple sectors according to azimuth, pitch and range, and the distance to the nearest obstacle and the marking of unknown areas are determined according to the preprocessed environmental perception data. S3. Track obstacles with continuous temporal sequence and predict the occupancy trend of each obstacle in each sector in the future time period; S4. Calculate the risk value of each sector based on the distance to the nearest obstacle in each sector, the unknown area marker, the relative approach speed, the measurement confidence level, the terrain height margin, and the obstacle type. The risk value of the sector corresponding to the dynamic obstacle is determined based on the occupancy trend of the obstacle. S5. Generate a candidate control command set under the flight dynamics constraints of the unmanned aerial vehicle; S6. Perform trajectory deduction on the candidate control command set, calculate the comprehensive cost based on the risk value of the deduced trajectory in each sector and the deviation from the mission trajectory, and select the one with the best comprehensive cost as the execution command output. S7. Execute the execution instruction and return to S1 in the next control cycle.

[0020] This invention provides an autonomous obstacle avoidance control method for low-altitude unmanned aerial vehicles (UAVs) using sector risk grids. It can be applied to various low-altitude UAVs, including multi-rotor and fixed-wing models, and is suitable for low-altitude flight scenarios with complex obstacles and terrain undulations, such as urban edges, industrial parks, mountainous forests, and power transmission corridors. This method uses the UAV's body coordinate system as the core reference, discretizing the surrounding space into standardized sector risk grids. It uniformly integrates multi-dimensional risk factors such as obstacle distance, collision time, sensor confidence, terrain height margin, unknown areas, and obstacle types. Combined with flight dynamics constraints, it generates executable obstacle avoidance control commands, forming a complete autonomous obstacle avoidance control chain encompassing perception preprocessing, grid construction, obstacle prediction, risk quantification, command selection, cost optimization, and a rolling closed loop. Simultaneously, it incorporates a hierarchical obstacle avoidance strategy and a hysteresis anti-shake mechanism, improving the executability of obstacle avoidance commands and the smoothness of flight paths while ensuring low-altitude flight safety. The steps described above are detailed below.

[0021] The aforementioned flight dynamics constraints refer to the physical boundaries that an aircraft can execute at the flight control level, specifically including: maximum horizontal speed (typically 15 m / s² for rotorcraft and 30 m / s² for fixed-wing aircraft) and maximum horizontal acceleration (typically 3-6 m / s²). 2The maximum climb rate (2-5 m / s), maximum descent rate (2-4 m / s), maximum roll / pitch rate (30-90° / s), and maximum yaw rate (45-90° / s) are all specified. These limits are determined by the aircraft parameters and the flight control safety envelope, and can be set in the parameter configuration file. A planner that does not incorporate these constraints may output sharp turns or climbs that are physically impossible for the flight control to execute within a single control cycle.

[0022] This invention uses the aircraft's airframe as a reference, dividing the surrounding space into multiple sectors, each sector containing azimuth, pitch, and range layers. The system converts the results of lidar, depth camera, millimeter-wave radar, ultrasonic, or visual recognition into a unified sector risk grid, and then calculates a risk value (Risk) for each sector. m (k) and, in conjunction with the current flight speed, target trajectory, flight control capability boundaries, and sensor confidence, generate candidate speed / heading / altitude commands, selecting the optimal command that satisfies the safe distance and dynamic constraints. Specifically: For step S1, acquiring the motion state of the unmanned aerial vehicle and environmental perception data of the surrounding environment, preprocessing the environmental perception data and marking areas without effective perception data as unknown areas, includes: performing time synchronization and spatial coordinate transformation on the environmental perception data to unify it to the body coordinate system; removing invalid points and marking the field of view validity of the environmental perception data after unifying the coordinate system; marking areas where no effective perception data falls after removing invalid points and marking the field of view validity as unknown areas.

[0023] The motion state data of the unmanned aerial vehicle (UAV) is jointly acquired by the onboard flight control system, inertial measurement unit, and satellite navigation module. Specifically, it includes the UAV's three-dimensional position, three-dimensional velocity, attitude angles, absolute flight altitude, heading angle, vertical velocity, and current flight control operating mode. This state data provides a complete state reference for subsequent coordinate transformations, relative approach velocity calculations, and flight dynamics constraint verification. Environmental perception data can be provided by one or more combinations of lidar, depth cameras, millimeter-wave radar, ultrasonic ranging devices, or visual target recognition modules. The corresponding data formats cover obstacle distance values ​​in the forward, lateral, and downward directions, 3D point clouds, depth maps, visual target detection boxes, and millimeter-wave radar target lists. Multi-source perception data complement each other, jointly covering the detection needs of different orientations, pitch angles, and distances around the UAV. This adapts to diverse obstacle perception scenarios in complex low-altitude environments, effectively compensating for the limitations of single sensors in terms of limited field of view and insufficient scene adaptability, thus improving the comprehensiveness and scene adaptability of environmental perception.

[0024] During preprocessing, the first step is to synchronize the time and transform the spatial coordinates of the multi-source sensing data, unifying them into the aircraft's coordinate system. The time synchronization stage uses the flight control system's master clock as a unified reference. Through linear timestamp interpolation, sensor data from different sampling frequencies are aligned to the same control cycle time, eliminating spatial position deviations caused by inconsistent sampling times of different sensors. This ensures accurate temporal correspondence of multi-source data, avoids distortion in obstacle position calculations caused by timing deviations, and guarantees the temporal consistency of subsequent data fusion. Based on this, a spatial coordinate transformation is performed. Using the aircraft's current attitude angles and position information, a homogeneous coordinate transformation matrix is ​​used to transform the obstacle data collected by each sensor from its respective sensor local coordinate system to the UAV's body coordinate system. The body coordinate system has the aircraft's center of mass as its origin, with the forward direction of the body as the positive X-axis, the right side of the body as the positive Y-axis, and the vertical downward direction of the body as the positive Z-axis. During coordinate transformation, motion compensation is completed simultaneously. Based on the aircraft's velocity and angular velocity increments within the current control cycle, motion distortion correction is performed on the obstacle ranging results. This eliminates the offset errors caused by the aircraft's own translation and rotation in the spatial position calculation of obstacles. It can unify the spatial reference benchmark of all sensing data and offset the point cloud motion distortion caused by the aircraft's own maneuvering, ensuring the accuracy of obstacle orientation and distance calculations and providing reliable spatial data support for the subsequent construction of sector risk grids.

[0025] After coordinate unification, invalid point removal and field-of-view validity marking are performed on the sensing data. Invalid point removal is divided into two levels: The first level is statistical outlier removal, which performs neighborhood statistical filtering on the 3D point cloud or ranging data, calculates the average distance between each measurement point and its neighbors, and identifies and removes measurement points whose average distance exceeds a preset statistical threshold as noise points. This effectively filters out false measurement points generated by sensor random noise and environmental clutter, reduces the interference of invalid data on subsequent obstacle identification and risk calculation, and improves the signal-to-noise ratio of the sensing data. The second level is background filtering, which combines the aircraft's current attitude, flight altitude information, and point cloud normal features to identify and filter non-obstacle measurement points belonging to ground terrain and sky areas. This can eliminate false detections from non-obstacle backgrounds such as ground and sky, reduce false alarms caused by ground objects and terrain in low-altitude environments, avoid meaningless deceleration or detour maneuvers, and improve the pertinence of obstacle avoidance decisions and the continuity of flight missions. After invalid points are removed, the effective field of view of each sensor is marked. Based on the azimuth detection angle range, pitch detection angle range, and maximum detection range of each sensor, its effective sensing area in the body coordinate system is defined. Measurement points within the effective field of view are marked as effective sensing data, and areas beyond the field of view boundary or maximum detection range are marked as areas without sensing coverage. This clearly defines the sensing boundary and effective data range of each sensor, providing a clear and quantifiable basis for subsequent determination of unknown areas.

[0026] After invalid point removal and field-of-view validity marking, the space surrounding the aircraft is verified one by one. Areas where no valid sensing data falls are uniformly marked as unknown areas. These areas include blind spots in the sensor's physical field of view, distant areas beyond the sensor's maximum detection range, and perception shadow areas formed by obstacles. Unknown areas cannot be directly determined as safe and passable areas. They must be included in the risk assessment system through a special unknown area penalty item in the subsequent sector risk calculation stage. This can clearly distinguish between detected safe areas and undetected unknown areas, avoiding the default assumption that perception blind spots and obstructed areas are passable spaces. This avoids the safety hazards caused by blind spots from the source of perception and strengthens the safety redundancy of low-altitude obstacle avoidance.

[0027] For step S2, using the body coordinate system of the unmanned aerial vehicle as a reference, the surrounding space is divided into multiple sectors according to azimuth, pitch and range, and the distance to the nearest obstacle and the unknown area are determined according to the preprocessed environmental perception data.

[0028] Specifically, the sector division process uses the UAV's body coordinate system as the spatial reference, with the forward direction of the UAV as the zero-azimuth reference point. The space surrounding the UAV is uniformly discretized along three dimensions: azimuth, pitch, and distance, constructing a three-dimensional sector grid system. This transforms the continuous surrounding space into quantifiable, discrete grid units. The unified grid framework enables the integration of multi-source, heterogeneous sensing data into a single spatial evaluation system, achieving refined differentiation of obstacle risks based on azimuth, pitch, and distance. This provides a standardized spatial reference for subsequent multi-dimensional risk quantification and obstacle avoidance command evaluation.

[0029] The aforementioned azimuth angle dimensions cover an omnidirectional range of ±180° around the aircraft, achieving full-circumferential risk coverage around the aircraft. The angular resolution can be configured to 22.5°, 30°, or 45° depending on sensor accuracy and computing power, corresponding to 16, 12, or 8 azimuth sectors per circle. This embodiment preferably uses a 22.5° resolution. This resolution level strikes a balance between azimuth resolution accuracy and computational efficiency, accurately distinguishing obstacles from different directions without placing an excessive burden on onboard computing power due to an excessive number of grids, thus adapting to the onboard computing resource constraints of low-altitude aircraft.

[0030] The aforementioned pitch angle dimension covers a detection range of -30° to +60°, covering low-altitude terrain and near-ground obstacles downwards, and overhead obstacles and climb avoidance space upwards. The pitch resolution is set to 15° to 30°, divided into 3 to 6 pitch sectors. This pitch range is tailored to the characteristics of obstacle avoidance scenarios in low-altitude flight, focusing on covering forward-looking and near-ground high-risk areas, while also taking into account the risk assessment of upward climb avoidance channels, thus matching the actual detection needs of low-altitude obstacle avoidance.

[0031] The aforementioned distance layers are divided into three levels along the radial direction of the aircraft: near distance, medium distance, and far distance, corresponding to ranges of 0-5m, 5-20m, and 20-100m, respectively. The distance range of each layer can be linearly scaled according to the maximum range of the sensor. This three-level distance layering enables graded risk management of obstacles at different distances: the near distance layer corresponds to emergency collision risk, the medium distance layer corresponds to the obstacle avoidance decision zone, and the far distance layer corresponds to the early warning and prediction zone, which aligns with the risk response logic of low-altitude obstacle avoidance from far to near.

[0032] After being divided according to the above rules, the total number of sectors is the number of azimuth sectors × the number of pitch layers × the number of distance layers. In this embodiment, the typical configuration is 16×3×3, a total of 144 sectors. The standardized grid size is compatible with the onboard computing capabilities of most low-altitude aircraft and can ensure the real-time performance of obstacle avoidance control.

[0033] After sector division, the preprocessed valid sensing data is mapped to the corresponding sectors according to spatial location, and the core attribute parameters of each sector are determined one by one. This is based on the nearest obstacle distance d. m (k) Extract the distance values ​​of all valid distance measurement points within the sector, and take the minimum value as the nearest obstacle distance in the sector. The effective distance measurement range is limited to 0.5m to 100m, and invalid distance measurement values ​​exceeding the range are discarded. Using the nearest obstacle in the sector to represent the maximum collision threat in that direction ensures that the risk assessment is always biased towards safety, avoiding the omission of high-risk obstacles at close range due to taking average distances or random point distances, and preserving sufficient safety redundancy for obstacle avoidance decisions. If there is no valid measurement data in the sector, it is directly marked as an unknown sector.

[0034] Mark u for unknown regions m (k) Maps the marked unknown areas to corresponding sectors according to their spatial range. If no valid sensing data falls into a sector, the unknown area flag for that sector is set to 1, corresponding to a sensor blind zone, over-range area, or occluded shadow area; if valid sensing data exists in a sector, the unknown area flag is set to 0. Sectors marked as unknown are not directly determined as safe and passable areas. Their uncertainty is reflected in subsequent risk calculations through a special penalty term. This allows the unknown attributes from the sensing preprocessing stage to be fully transferred to the risk grid system, maintaining consistency in risk assessment and avoiding safety misjudgments caused by sensing blind zones at the spatial grid level.

[0035] In addition to the core parameters mentioned above, each sector also synchronously stores auxiliary attributes related to risk calculation, providing complete data support for risk quantification in subsequent steps: measuring confidence level c m (k) is determined by the echo signal-to-noise ratio of the sensing data within the corresponding sector and the field-of-view coverage flag, reflecting the reliability of the sensing data in that sector; obstacle type S m(k) Assign values ​​based on the output of the target classification algorithm to distinguish obstacle types with different risk levels, such as static obstacles, dynamic obstacles, and personnel targets; relative approach speed v rel,m (k) is calculated from the difference between the projection of the aircraft's velocity vector in the direction of the center of the sector and the radial velocity of the obstacle; a positive value indicates that the aircraft and the obstacle are in a state of mutual approach; terrain height margin H m (k) is calculated from the difference between the current flight altitude and the minimum safe altitude of the terrain in the direction of the sector. The terrain altitude data can be obtained from a digital elevation model or lidar terrain scanning results. By synchronously storing the above multi-dimensional attributes, information such as distance, perception reliability, obstacle risk level, relative motion trend, and terrain safety margin can be integrated, providing complete data support for subsequent obstacle tracking prediction and multi-factor risk fusion calculation.

[0036] By constructing a three-dimensional discretized sector grid, this step transforms the continuous low-altitude surrounding space into standardized risk assessment units, realizing a unified mapping of multi-source sensing data from scattered measurement points to structured grids. This not only accurately locates obstacles and threats in different spatial orientations, but also integrates multi-dimensional risk information through a unified grid data structure.

[0037] For step S3, tracking obstacles with continuous temporal sequence and predicting the occupancy trend of each obstacle in the sector in the future includes: performing obstacle clustering and inter-frame association on the preprocessed environmental perception data to determine the temporal correspondence of the same obstacle; configuring the process noise covariance of Kalman filtering according to the motion attributes of the obstacle, configuring low noise level for static obstacles and high noise level for dynamic obstacles; using the temporal correspondence as the observation input and the three-dimensional position and velocity of the obstacle in the body coordinate system as the state variables, performing recursive state estimation through the configured Kalman filter; based on the state estimation result at the current moment, forward extrapolating the obstacle motion trajectory in the preset time domain, and determining its occupancy trend in the sector according to the obstacle motion trajectory.

[0038] Specifically, obstacle clustering and inter-frame association are first performed on the preprocessed environmental perception data to construct a continuous target tracking link. For 3D point clouds, ranging point sets, or target detection results, clustering is performed based on spatial distance, size contour, and reflection intensity features. Measurement points corresponding to the same obstacle are aggregated into independent target individuals, obtaining the cluster center position and shape contour of each obstacle. Then, by matching the spatial overlap, motion continuity, and size similarity of targets between adjacent frames, the temporal correspondence of the same obstacle in consecutive frames is determined, providing a continuous and stable observation sequence for subsequent motion state estimation, avoiding the confusion of observation data from different obstacles, and ensuring the accuracy of motion trend prediction.

[0039] Based on this, Kalman filtering is used to recursively estimate the motion state of each obstacle, and the state variable x of the obstacle is obtained. obs Set the three-dimensional position and three-dimensional velocity in the body coordinate system, that is: , Where, p x p y p z The following are the position components of the obstacle in the X, Y, and Z directions in the body coordinate system, respectively. x v y v z The values ​​are, in order, the velocity components of the obstacle in the X, Y, and Z directions in the machine's coordinate system.

[0040] State transition vector: x obs (k+1)=F·x obs (k)+w(k), Where w(k) is the process noise vector (6-dimensional) at time k, and F is the state transition matrix (6×6) of the uniform motion model. Specifically, the state transition adopts the uniform motion model, and the corresponding state transition matrix F is: , Where I3 is a 3-order identity matrix, 03 is a 3-order zero matrix, and T s The control period is set to 0.02 to 0.05 seconds. In this matrix, the upper left identity matrix transmits the position information from the previous moment, the upper right block adds the product of velocity and control period to the position update, and the lower right identity matrix maintains the assumption of uniform velocity components. This allows for the recursive calculation of obstacle motion state with low computational cost, adapting to the real-time calculation requirements of airborne platforms.

[0041] The process noise covariance of the Kalman filter is configured differently based on the motion properties of the obstacle: a lower process noise level is configured for static obstacles, typically 0.01 m / s. 2 After convergence, a stable position estimation result can be obtained; the dynamic obstacle configuration has a high process noise level, typically ranging from 0.5 to 2.0 m / s. 2 This allows for greater adjustment range in velocity estimates, adapting to changes in the velocity of moving targets. Differentiated noise configurations balance the estimation accuracy of static obstacles with the tracking flexibility of dynamic obstacles, avoiding the problems of static target jitter or dynamic target tracking lag caused by a single noise parameter.

[0042] The measurement process updates the state estimate based on real-time observations of the obstacle. The measurement equation is as follows: , Among them, z obs(k) is the actual observation vector of the obstacle cluster center at time k, including the three-dimensional position measurement in the body coordinate system, obtained by the current frame observation of the obstacle cluster center by sensors such as lidar and depth camera. H is the measurement matrix with a value of [I3,03]. Only the first three-dimensional position components in the state vector are extracted and correspond to the observation values. The velocity component does not directly participate in the observation update. v(k) is the measurement noise vector at time k, which follows a Gaussian distribution with a mean of zero. Its covariance matrix is ​​determined by the ranging accuracy of the sensor. The typical lidar configuration is diag[0.02]. 2 0.02 2 0.02 2 ]m 2 A typical configuration for a depth camera is diag[0.05]. 2 0.05 2 0.10 2 ]m 2 By matching the measurement error characteristics of different sensors, the reliability of state estimation results can be improved.

[0043] After completing the state estimation at the current moment, in the preset prediction time domain T p Inside, based on the uniform velocity model, the trajectory of the obstacle is extrapolated forward to predict the time domain T. p The typical value ranges from 0.5s to 2.0s, and in this embodiment, 1.0s is preferred, corresponding to the number of prediction steps N. pred =T p / T s The number of steps ranges from 20 to 40. The obstacle positions obtained by extrapolation at each step are mapped to the sector risk grid constructed in step S2. The sector occupied by the obstacle at each prediction time is determined one by one, and finally the occupancy trend of each obstacle in each sector in the future prediction time domain is obtained.

[0044] Through the aforementioned time-series tracking and motion prediction processing, the instantaneous position information of obstacles can be extended into continuous motion trend prediction, breaking through the limitation of judging risk based solely on the current single-point distance. This allows for the early identification of flight areas that dynamic obstacles may enter in the future, providing a forward-looking basis for the risk modulation of subsequent dynamic obstacles. It effectively addresses the problem of delayed obstacle avoidance response caused by the sudden approach of moving targets in low-altitude environments, and enhances the adaptability of the entire obstacle avoidance system to dynamic and complex scenarios.

[0045] For step S4, the risk value of each sector is calculated based on the distance to the nearest obstacle in each sector, the unknown area marker, the relative approach speed, the measurement confidence level, the terrain height margin, and the obstacle type. The sector risk value corresponding to a dynamic obstacle is determined based on the occupancy trend of the obstacle.

[0046] In this step, calculating the risk value of each sector based on the nearest obstacle distance, the unknown area marker, the relative approach speed, the measurement confidence level, the terrain height margin, and the obstacle type includes: determining the collision time based on the nearest obstacle distance and the relative approach speed; determining the measurement confidence level based on the sensor echo signal-to-noise ratio and the effective field of view coverage; calculating the terrain height margin based on the current flight altitude and the terrain height in the sector direction; and weightedly fusing the nearest obstacle distance, the collision time, the measurement confidence level, the terrain height margin, the unknown area marker, and the obstacle type to obtain the risk value of each sector. The sector risk value corresponding to the dynamic obstacle is determined based on the obstacle's occupancy trend, specifically including: for the dynamic obstacle, using the risk value of each sector obtained by the weighted fusion as the base risk value; based on the occupancy trend, increasing the base risk value of the sectors predicted to be occupied by the dynamic obstacle in the future time period; and decreasing the base risk value of the sectors currently occupied by the dynamic obstacle but no longer occupied in the future.

[0047] First, the collision time for each sector is calculated to quantify the urgency of the collision risk from a time perspective. The collision time (TTC) for the m-th sector is... m (k) is calculated using the following formula: , , Where, d m (k) is the distance to the nearest obstacle, d safe To preset a safe distance, ε is a small value used to prevent division by zero, ranging from 0.05 to 0.10 m / s. This avoids numerical overflow when the relative velocity approaches zero, ensuring calculation stability. cruise T is the current cruise forward speed of the aircraft. brake The maximum braking time of the aircraft, ranging from 0.5 to 2.0 seconds, d min The minimum physical safety distance ranges from 0.5 to 3.0 meters, such as: v cruise =5m / s, T brake =1.0s, d min =1.0m, then d safe =6.0m. The safety distance, combined with the aircraft's braking capacity, ensures that the aircraft can decelerate and brake within a safe distance after obstacle avoidance is triggered. When the relative approach speed v... rel,m When (k)≤0, it indicates that the aircraft is relatively far away from or stationary from the obstacles in this sector, with no tendency to collide. At this time, TTC will be... m (k) is set to the saturation maximum value T. inf=999s, corresponding to a negligible contribution to the collision risk factor. Compared to relying solely on obstacle distance to assess risk, combining collision time with relative speed can more accurately reflect the actual urgency of a collision, distinguish the risk differences at different flight speeds at the same distance, and avoid the problems of excessive obstacle avoidance in low-speed, close-range scenarios and untimely response in high-speed, long-range scenarios.

[0048] Next, the measurement confidence level for each sector is calculated, and the risk assessment results are revised from the perspective of perceived data quality. Measurement confidence level c m (k) is determined by both the sensor echo signal-to-noise ratio and the effective field-of-view coverage, and is calculated using the following formula: , Among them, SNR m Let SNR be the signal-to-noise ratio of the sensor echo in the m-th sector. ref field_flag is the preset reference signal-to-noise ratio. m The effective field-of-view coverage indicator for the sensor is 0 or 1, with 1 for data within the field of view and 0 for data outside the field of view. The final measurement confidence score is normalized to [0,1], with higher values ​​indicating stronger reliability of the sensing data for that sector. Incorporating the measurement confidence score into the risk calculation system can impose risk penalties on low-quality sensing data with low signal-to-noise ratios and at the edge of the field of view, avoiding the use of unreliable measurement results as a basis for safety judgments and reducing the risk of safety misjudgments caused by sensing noise and weak echo areas.

[0049] Then, the terrain height margin for each sector is calculated to compensate for low-altitude terrain collision risk in the vertical dimension. The penalty term H corresponding to the terrain height margin is... m (k) is calculated using the following formula: , Among them, h safe_margin To provide a preset safety margin, the value ranges from 3 to 8 m, h uav h represents the current flight altitude of the unmanned aerial vehicle. terrain,m The terrain height in the direction of the m-th sector can be obtained through digital elevation modeling or lidar terrain scanning. When the calculated result is greater than 0, it indicates that the terrain height margin in that sector is insufficient, and the larger the value, the higher the risk of ground collision; when the calculated result is 0, it indicates that the terrain height margin is sufficient and there is no risk of terrain collision. The terrain height margin term is specifically adapted to low-altitude complex terrain flight scenarios, making up for the shortcomings of traditional obstacle avoidance schemes that only focus on horizontal obstacles and ignore the threat of terrain undulations, and realizing an integrated evaluation of horizontal and vertical risks.

[0050] After completing the quantitative calculation of each risk sub-item, the nearest obstacle distance, collision time, measurement confidence level, terrain height margin, unknown area markers, and obstacle type are weighted and fused to obtain the risk value (Risk) for each sector. m(k), the calculation formula is: , Among them, w d w is the distance term weighting coefficient, with values ​​ranging from [0.3, 0.5]. t The collision time term weighting coefficient has a value of [0.2, 04]. c , where w is the confidence level weight coefficient, with values ​​ranging from [0.05, 0.15]. h , where w is the weighting coefficient for the terrain surplus term, with values ​​ranging from [0.05, 0.15]. u The weighting coefficient for the unknown region term takes values ​​in the range [0.05, 0.15]. s The weight coefficient for the obstacle type item is [0.05, 0.1], and the sum of the six weights is 1, σ d This is the distance attenuation coefficient, ranging from 1 to 5m, used to control the rate attenuation of distance risk as the distance to the obstacle increases. σ t This is the collision time decay coefficient, ranging from 0.5 to 3.0 s, used to control the rate attenuation of time risk as the collision time increases. Obstacle type S m (k) Based on the target detection and classification results, differentiated risk coefficients are set for different types of obstacles: static obstacles are assigned a value of 1.0, dynamic obstacles are assigned a value of 1.2 to 1.5, and personnel targets are assigned a value of 2.0. The risk weight is increased for highly sensitive targets such as personnel, aligning with the safety priority requirements for low-altitude flight. m (k) is a marker for unknown regions, assigned by the preprocessing module. When valid sensing data exists in this sector, u... m =0 indicates effective sensing. When the sector is in the sensor blind zone or exceeds the range, u m =1 indicates a blind zone.

[0051] The multi-factor weighted fusion risk calculation method breaks through the limitations of traditional schemes that judge risk based solely on distance. It can comprehensively consider multiple dimensions of information, including spatial distance, collision urgency, perception reliability, terrain constraints, and target type, to output risk assessment results that are more consistent with actual low-altitude scenarios. Specifically, calibration can be performed by traversing typical obstacle scenarios in a simulation environment, targeting obstacle avoidance success rate and false alarm rate, and optimizing the weight combination through grid search or genetic algorithms.

[0052] For dynamic obstacles, the sector risk value is adjusted directionally based on the obstacle's occupancy trend. Specifically, the risk values ​​of each sector obtained through the weighted fusion are used as the base risk value, and adjusted according to the future sector occupancy trend output in step S3. Based on the occupancy trend, the base risk value of the sector predicted to be occupied by the dynamic obstacle in the future is modulated upwards, raising the risk level in the corresponding direction in advance. This allows the obstacle avoidance system to respond to impending collision threats in advance, solving the problem of delayed obstacle avoidance response when a dynamic obstacle suddenly approaches. The base risk value of the sector currently occupied by the dynamic obstacle but no longer occupied in the future is modulated downwards, restoring the true risk level in that direction. This avoids continuously triggering unnecessary obstacle avoidance actions as the obstacle is about to move out of the flight path, balancing flight safety and mission execution efficiency. Dynamic risk modulation based on motion prediction extends risk assessment from the current instantaneous state to future trend prediction, significantly improving the system's adaptability to complex low-altitude dynamic scenarios.

[0053] For step S5, a candidate control instruction set is generated under the flight dynamics constraints of the unmanned aerial vehicle.

[0054] In this step, generating a candidate control command set under the flight dynamics constraints of the unmanned aerial vehicle includes: performing discrete sampling in a preset control command parameter space and combining the sampled values ​​of each dimension into multiple candidate command vectors; verifying each candidate command vector according to the flight dynamics constraints of the unmanned aerial vehicle; eliminating candidate command vectors that do not meet any constraint conditions, and integrating the remaining candidate command vectors into a candidate control command set.

[0055] Specifically, the candidate control commands cover three core control dimensions: forward speed, yaw heading, and vertical climb rate. A full range of candidate combinations are generated through discrete sampling, and then each combination is verified and screened in conjunction with the physical execution boundaries of the aircraft to ensure that the final output candidate commands are all within the executable range of the flight control system. This avoids the problem of the planning results being out of sync with the flight control capabilities from the command generation end.

[0056] First, discrete sampling is performed within the preset control command parameter space. The parameter space contains three independent dimensions: one is the desired forward velocity v. cmd In [0, v max Within the range, samples are taken at fixed intervals with a sampling step size of 0.5 to 1.0 m / s, covering the entire feasible speed range from hovering to maximum cruise speed; secondly, the expected yaw change ψ cmd , in [-ψ max ,ψ max Within the range, sampling is performed in angular steps, with step sizes ranging from 5° to 15°, covering the entire feasible heading adjustment range for left and right turns; thirdly, the expected rate of climb h cmd ,exist Within the range, sampling is performed at speed steps, with step sizes ranging from 0.2 to 0.5 m / s, covering the vertical maneuvering range of climb and descent. All sampled values ​​from the three dimensions are fully combined to obtain several sets of candidate command vectors. Each set of candidate command vectors contains three control components: desired forward velocity, desired yaw change, and desired rate of climb. Discrete sampling can cover the main feasible control space with limited computational resources, ensuring real-time obstacle avoidance decision-making under airborne computing power constraints. The sampling step size for each dimension can be flexibly adjusted according to the aircraft's maneuvering characteristics and computing power conditions.

[0057] After the candidate command vectors are generated, all candidate command vectors are verified one by one according to the flight dynamics constraints of the unmanned aerial vehicle. The verification includes four core boundaries: maximum speed constraint, yaw rate constraint, vertical rate constraint, and horizontal acceleration constraint. All constraints are based on the output command value of the previous control cycle as the differential reference to avoid the interference of noisy measured states amplified by differential operation, which would cause high-frequency jitter in the constraint judgment.

[0058] The first term is the maximum speed constraint, used to limit the commanded speed from exceeding the aircraft's power and flight control safety envelope. The constraint expression is: , Among them, v cmd v represents the desired forward velocity in the candidate instruction. max This is the maximum horizontal speed of the aircraft, typically ranging from 10 to 15 m / s. The specific value is determined by the aircraft's dynamic characteristics and flight control safety envelope.

[0059] The second term is the yaw rate constraint, which limits the rate of change of heading within a single cycle to ensure it does not exceed the yaw maneuverability of the flight control system. The constraint expression is: , Where, ψ cmd ψ is the desired yaw angle in the candidate command. prev This is the yaw angle command value output in the previous control cycle. The maximum yaw rate of the aircraft is 45° / s. Using the previous cycle's command value as a baseline for calculating changes avoids amplifying sensor measurement noise after differential calculation, preventing high-frequency jitter in constraint decisions and improving the stability of command selection.

[0060] The third term is the vertical rate constraint, used to limit the rate of altitude change within a single cycle to not exceed the aircraft's climb and descent capabilities. The constraint expression is: , Among them, h cmd For the desired climb rate instruction among the candidate instructions, h prev The climb rate command value output in the previous control cycle. The maximum climb rate of the aircraft is set to 3 m / s to ensure that vertical maneuvers are within the range that the flight control can execute.

[0061] The fourth constraint is the horizontal acceleration constraint, which limits the velocity change within a single cycle to ensure it does not exceed the aircraft's maximum acceleration / deceleration capability. The constraint expression is as follows: , Among them, a cmd v is the single-cycle horizontal acceleration corresponding to the current candidate instruction. prev a is the forward speed command value output in the previous control cycle. max The maximum horizontal acceleration of the aircraft is taken as 4 m / s². 2 This avoids situations where the flight control cannot respond to rapid acceleration or deceleration commands within a single cycle.

[0062] After all constraint checks are completed, candidate command vectors that do not meet any constraint conditions are eliminated. The candidate command vectors that pass all constraint checks are integrated into a candidate control command set, which serves as the input basis for the subsequent command optimization stage. This entire generation and screening mechanism places flight dynamics constraints at the candidate command generation stage, rather than performing amplitude limiting after planning is completed. This ensures that all candidate commands are physically executable, avoiding the problem of flight control not being able to respond promptly to commands such as sharp turns, sudden stops, and sharp climbs output by traditional planning methods. It ensures that the obstacle avoidance commands output by the subsequent optimization can be accurately executed by the flight control system, improving the actual control effect of low-altitude obstacle avoidance.

[0063] For step S6, the candidate control command set is used to perform trajectory extrapolation. Based on the risk value of the extrapolated trajectory in each sector and the deviation from the mission trajectory, the comprehensive cost is calculated, and the one with the best comprehensive cost is selected as the execution command output.

[0064] In this step, the trajectory extrapolation of the candidate control command set, and the calculation of the comprehensive cost based on the risk value of the extrapolated trajectory in each sector and its deviation from the mission trajectory, includes: using each candidate command in the candidate control command set as input, recursively generating the corresponding predicted trajectory within a preset prediction time domain; accumulating the risk value of the sectors traversed along each predicted trajectory to obtain the risk cost corresponding to each candidate command; calculating the deviation between the end position of each predicted trajectory and the corresponding target point on the current mission trajectory to obtain the trajectory deviation cost corresponding to each candidate command; determining the energy consumption cost based on the degree of difference between each candidate command and the current flight state; determining the command smoothness cost based on the degree of difference between each candidate command and the command executed in the previous cycle; determining the unknown area crossing cost based on the proportion of steps falling into the unknown area in the predicted trajectory; and weighted summing the risk cost, the trajectory deviation cost, the energy consumption cost, the command smoothness cost, and the unknown area crossing cost to obtain the comprehensive cost corresponding to each candidate command.

[0065] Specifically, each candidate command in the candidate control command set is first used as a constant control input to perform short-time trajectory extrapolation, generating a predicted trajectory under the corresponding candidate command. The extrapolation process employs a uniform velocity approximation model, progressively extrapolating the aircraft position along the forward velocity, yaw rate of change, and climb rate of the candidate command. The prediction time domain is consistent with the obstacle motion prediction, and the extrapolation step size matches the control cycle. The aircraft position at each moment in the prediction time domain is output point by point, forming a complete predicted trajectory. Rolling trajectory extrapolation based on candidate commands can complete the safety pre-assessment of each path under limited onboard computing power, providing spatial location basis for subsequent multi-dimensional cost calculations.

[0066] After the predicted trajectory is generated, the cost components corresponding to each candidate command are calculated from five dimensions: mission tracking, safety risk, maneuver energy consumption, command smoothness, and unknown area risk, forming a multi-dimensional comprehensive evaluation system.

[0067] The first term is the track deviation cost, used to quantify the degree of deviation between the candidate track and the preset mission track, ensuring that the obstacle avoidance process follows the mission path as closely as possible and avoids excessive detours that could affect mission execution efficiency. Track Deviation Cost E track The formula for calculating (u) is: , Where: p pred (u,Tp) represents the candidate instruction u applied to the prediction time domain T. p At the terminal phase, the corresponding three-dimensional position of the spacecraft. target The position of the forward target point on the current mission track is provided in real time by the mission planning layer along the preset track, and is usually the nearest forward target point within visual range along the mission track. normThe normalization scale for track deviation is set to 10m. It is used to map the position deviation to the normalization interval of [0,1]. The larger the value, the farther the candidate track deviates from the mission path.

[0068] The second item is the risk cost, used to assess the cumulative safety risk of the predicted trajectory crossing the risk grid, and is the core safety constraint for candidate instruction selection. Risk Cost E risk The formula for calculating (u) is: , Among them, Risk m (k+t) represents the risk value of the m-th sector at the t-th prediction time, overlap(u, sector) m The value is a coverage flag for the predicted trajectory of candidate command u and the m-th sector, taking the value 0 or 1. It is set to 1 if the predicted trajectory falls into the sector at any time, and 0 otherwise. This formula takes the maximum risk value of the sectors traversed by the trajectory at each prediction time, and then calculates the average over the entire prediction time domain. This reflects the average highest risk level of the candidate trajectory within the entire prediction window; a higher value indicates a greater safety risk for the trajectory. Using the maximum risk at each time point allows for a more conservative assessment of trajectory safety, avoiding safety hazards caused by local high-risk points being masked by the mean.

[0069] The third item is energy consumption cost, used to quantify the maneuver intensity corresponding to candidate commands, suppress unnecessary large accelerations, decelerations, and turns, and reduce flight energy consumption. Energy consumption cost E energy The formula for calculating (u) is: , Among them, v current This represents the aircraft's current flight speed. The maneuver intensity is quantified by normalizing the square of the speed difference; a larger value indicates a greater maneuver amplitude and higher energy consumption, and the guidance system prioritizes obstacle avoidance paths with gentler maneuvers.

[0070] The fourth item is the command smoothness cost, used to quantify the degree of difference between the candidate command and the command executed in the previous cycle, avoiding frequent abrupt changes in control commands, reducing execution jitter in the flight control system, and improving the smoothness of the flight path. Command smoothness cost E smooth The formula for calculating (u) is: , Among them, u prev This is the execution command output from the previous control cycle. The difference is calculated based on the command from the previous cycle, which constrains the change in commands between adjacent control cycles, preventing abrupt changes in obstacle avoidance commands and ensuring the continuity and stability of flight control execution.

[0071] The fifth item is the cost of traversing unknown areas, used to quantify the proportion of predicted flight paths crossing unknown sectors. This guides aircraft to prioritize directions with sufficient perception coverage, reducing potential safety risks from unknown areas. The cost of traversing unknown areas is E. unknown The formula for calculating (u) is: , Where m is the sector index, u m The m-th sector is marked as an unknown region. `count{*}` is a statistical function used to count the total number of sectors that satisfy the condition "the sector is an unknown region and the predicted trajectory passes through this sector". The calculated result ranges from [0,1]. A larger value indicates that the candidate trajectory traverses more unknown regions, resulting in lower path perception reliability. This cost avoids prioritizing sensor blind zone directions in obstacle avoidance decisions, thus strengthening safety redundancy in unknown environments from the path planning level.

[0072] After calculating the five cost components, the risk cost, track deviation cost, energy consumption cost, command smoothness cost, and unknown area crossing cost are weighted and summed to obtain the comprehensive cost corresponding to each candidate command. The calculation formula is as follows: ; Wherein, b1 is the trajectory deviation cost weight coefficient, with a value range of [0.3, 0.5]; b2 is the risk cost weight coefficient, with a value range of [0.3, 0.5]; b3 is the energy consumption cost weight coefficient, with a value range of [0.05, 0.15]; b4 is the smoothness cost weight coefficient, with a value range of [0.05, 0.15]; and b5 is the unknown area crossing cost weight coefficient, with a value range of [0.05, 0.15]. The sum of the five weights is 1. The weight ratio can be flexibly adjusted according to task requirements. In scenarios with high safety priority, the weight of the risk item can be increased; in scenarios with high task priority, the weight of the trajectory deviation item can be increased. The weight parameters can be optimized and calibrated in the simulation environment with the goal of minimizing the weighted sum of obstacle avoidance failure rate and false alarm rate.

[0073] Finally, the comprehensive cost of all candidate commands is considered, and the candidate command with the lowest comprehensive cost and that satisfies the minimum safe distance constraint is selected as the output command for this control cycle. This multi-objective weighted comprehensive evaluation mechanism unifies flight safety, mission tracking, energy consumption control, command smoothness, and unknown risk avoidance under the same evaluation framework. It can balance mission execution efficiency and flight quality while ensuring low-altitude flight safety, outputting obstacle avoidance control commands that best meet actual flight requirements.

[0074] For step S7, the execution instruction is executed, and the process returns to S1 in the next control cycle.

[0075] When executing the execution command in this step, a graded obstacle avoidance strategy is implemented according to the risk level: when the risk value reaches a first threshold, the target heading is corrected by a preset angle; when the risk value reaches a second threshold greater than the first threshold, deceleration is performed and the lateral search weight is increased; when the risk value reaches a third threshold greater than the second threshold, if a low-risk passage exists laterally or above, lateral detour or climb is performed; if the risk values ​​of all sectors traversed by the candidate control command set are higher than the third threshold, a safety protection mode is executed. This graded obstacle avoidance strategy is configured with a hysteresis handling mechanism: action entry thresholds and action exit thresholds are set for each level of obstacle avoidance action; when executing the execution command, the obstacle avoidance action is triggered only when the number of control cycles in which the risk value continuously exceeds the corresponding risk level threshold reaches the action entry threshold; after the obstacle avoidance action is triggered, the current obstacle avoidance action is allowed to exit only when the number of control cycles in which the risk value continuously falls below a preset release threshold reaches the action exit threshold; wherein, the action exit threshold is greater than the action entry threshold.

[0076] Specifically, the selected execution command in step S6 is sent to the airborne flight control system. The flight control system then uses power output and control surface deflection to execute the corresponding maneuvers for forward speed, yaw angle, and climb rate. The system operates continuously in a fixed control cycle. After completing each round of obstacle avoidance calculation and command output, it returns to step S1 in the next control cycle to perform a new round of perception data acquisition, preprocessing, and full-link obstacle avoidance calculation, forming a real-time rolling closed-loop control link that dynamically responds to changes in obstacles in the low-altitude surrounding environment and adjusts the flight status.

[0077] During obstacle avoidance, the system matches the corresponding intensity of obstacle avoidance action according to the risk level of the current sector, and divides the response into four levels from low to high, so as to maintain the continuity of the mission track as much as possible while ensuring flight safety.

[0078] When the sector risk value reaches the first threshold, a first-level correction is triggered, which only makes a small adjustment to the target heading by a preset angle. The first threshold corresponds to a risk value ≥0.30 or a collision time ≤5.0s, which is a light adjustment at the warning level. Throughout the process, the flight speed and altitude remain basically unchanged, and low-risk obstacles are avoided in advance by making small heading deviations, so as to follow the original mission trajectory as closely as possible and reduce unnecessary maneuvers that interfere with mission execution.

[0079] When the sector risk value reaches a second threshold greater than the first threshold, a second-level deceleration is triggered, performing deceleration and increasing the lateral search weight. The second threshold corresponds to a risk value ≥ 0.50 or a collision time ≤ 2.0s. At this point, the forward collision risk is significantly increased. By reducing the forward flight speed, the available obstacle avoidance response time is extended. At the same time, the lateral search weight is increased in the candidate command selection stage to detect passable lateral paths in advance, reserving sufficient buffer space for subsequent obstacle avoidance actions.

[0080] When the sector risk value reaches a third threshold that exceeds the second threshold, if a low-risk passage exists to the side or above, a level three response is triggered, executing a lateral detour or climb maneuver. The third threshold corresponds to a risk value ≥ 0.70 or a collision time ≤ 1.0s, which is considered a high-risk level. The forward path no longer meets safe passage conditions, and the system actively selects a low-risk passage to the side or vertical to execute a detour or climb maneuver, proactively avoiding high-risk obstacles.

[0081] If the risk value of all sectors traversed in all directions covered by the candidate control command set is higher than the third threshold, and there is no reliable low-risk channel in all directions, then the level four safety protection mode is triggered, and hovering, deceleration, exiting along the original route, returning to home, or entering the preset safety control mode are executed to prioritize the safety of the aircraft itself and avoid entering uncontrollable high-risk areas.

[0082] To prevent frequent fluctuations in risk values ​​near critical thresholds due to perceived noise and transient disturbances, which could lead to repeated switching of obstacle avoidance actions, the tiered obstacle avoidance strategy incorporates a hysteresis handling mechanism. This mechanism suppresses command jitter through a dual design of asymmetric duration requirements and risk threshold bands. For each level of obstacle avoidance action, corresponding action entry and exit thresholds are set, both measured in units of continuous control cycles, to determine the duration conditions for action triggering and exiting. Simultaneously, corresponding risk level trigger and release thresholds are set, with the risk release threshold being lower than the risk trigger threshold, forming an asymmetric risk threshold band with a buffer zone.

[0083] Furthermore, an independent counting register is configured for each risk level. Only when the risk value of a sector in a certain direction continuously exceeds the risk trigger threshold for the corresponding level, and the number of consecutive control cycles reaches the action entry threshold for that level, is the obstacle avoidance action for that level formally triggered. The action entry threshold is typically set to 3 to 5 control cycles, which can effectively filter out false triggers caused by single-frame sensing noise or instantaneous disturbances, ensuring the reliability of obstacle avoidance action initiation.

[0084] After an obstacle avoidance action is triggered, it will not immediately degrade and exit due to a brief drop in risk value. Only when the risk value in the corresponding direction remains consistently below the risk clearance threshold for that level, and the number of consecutive control cycles reaches the action exit threshold for that level, will the current obstacle avoidance action be exited and the strategy downgraded to a lower level. The action exit threshold is typically set to 10 to 20 control cycles, and its value is greater than the action entry threshold, creating an asymmetric hysteresis characteristic of fast entry response and high exit delay. In a typical configuration, the risk clearance threshold is approximately 0.7 times the risk trigger threshold; for example, when the high-risk trigger threshold is 0.65, the corresponding clearance threshold is 0.45.

[0085] This asymmetric hysteresis design prioritizes safety, enabling it to respond quickly to obstacle avoidance needs when risks escalate and maintain stability when risks decrease. It avoids frequent switching of obstacle avoidance actions when risk values ​​oscillate within the threshold range, effectively suppressing command jitter and track oscillation, and significantly improving flight smoothness and control stability during low-altitude obstacle avoidance.

[0086] In summary, this solution employs a three-dimensional sector risk grid architecture in a body coordinate system. It quantifies sector risk using multi-dimensional parameters including nearest obstacle distance, collision time, measurement confidence, terrain height margin, unknown area marking, and obstacle type, significantly improving the comprehensiveness and accuracy of risk assessment in complex low-altitude scenarios. Through obstacle time-series tracking and motion trend prediction, it performs targeted modulation of sector risks corresponding to dynamic obstacles, enabling early response to moving targets and effectively improving the problem of delayed obstacle avoidance response in dynamic scenarios. Perception blind spots and over-range areas are marked as... Unknown areas are incorporated into risk calculations and trajectory cost constraints, preventing unperceived areas from being misjudged as safe passages and enhancing flight safety redundancy in unknown environments. Candidate control commands are generated under flight dynamics constraints, ensuring that all output commands are within the flight control's executable range and avoiding the problem of difficult-to-implement commands during rapid acceleration / deceleration and super-maneuvering. Finally, through trajectory extrapolation and multi-objective comprehensive cost optimization, the system balances mission trajectory tracking accuracy, command smoothness, and energy consumption control while ensuring flight safety, thus improving the overall reliability and engineering adaptability of autonomous obstacle avoidance for low-altitude unmanned aerial vehicles.

[0087] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the present invention in any way, and all technical solutions obtained by equivalent substitution or equivalent transformation fall within the protection scope of the present invention.

Claims

1. A method for autonomous obstacle avoidance control of a low-altitude unmanned aerial vehicle sector risk grid, characterized in that, Includes the following steps: S1. Acquire the motion state of the unmanned aerial vehicle and the environmental perception data of the surrounding environment, preprocess the environmental perception data and mark the areas without effective perception data as unknown areas; S2. Using the body coordinate system of the unmanned aerial vehicle as a reference, the surrounding space is divided into multiple sectors according to azimuth, pitch and range, and the distance to the nearest obstacle and the marking of unknown areas are determined according to the preprocessed environmental perception data. S3. Track obstacles with continuous temporal sequence and predict the occupancy trend of each obstacle in each sector in the future time period; S4. Calculate the risk value of each sector based on the distance to the nearest obstacle in each sector, the unknown area marker, the relative approach speed, the measurement confidence level, the terrain height margin, and the obstacle type. The risk value of the sector corresponding to the dynamic obstacle is determined based on the occupancy trend of the obstacle. S5. Generate a candidate control command set under the flight dynamics constraints of the unmanned aerial vehicle; S6. Perform trajectory deduction on the candidate control command set, calculate the comprehensive cost based on the risk value of the deduced trajectory in each sector and the deviation from the mission trajectory, and select the one with the best comprehensive cost as the execution command output. S7. Execute the execution instruction and return to S1 in the next control cycle.

2. The autonomous obstacle avoidance control method for low-altitude unmanned aerial vehicles (UAVs) based on a sector risk grid according to claim 1, characterized in that, The preprocessing of the environmental sensing data and the marking of areas with no valid sensing data as unknown areas includes: The environmental perception data is synchronized in time and transformed in spatial coordinates to be unified into the body coordinate system. Invalid points are removed and field of view validity is marked on the environmental perception data after unifying the coordinate system; Areas where no valid perception data falls after invalid point removal and field of view validity marking are marked as unknown areas.

3. The autonomous obstacle avoidance control method for low-altitude unmanned aerial vehicles (UAVs) based on a sector risk grid according to claim 1, characterized in that, The tracking of obstacles with continuous temporal sequence and the prediction of the occupancy trend of each obstacle in each sector in future time periods include: Obstacle clustering and inter-frame association are performed on the preprocessed environmental perception data to determine the temporal correspondence of the same obstacle; The process noise covariance of the Kalman filter is configured according to the motion properties of the obstacle; static obstacles are configured with a low noise level, and dynamic obstacles are configured with a high noise level. Using the time-series correspondence as the observation input, and the three-dimensional position and velocity of the obstacle in the body coordinate system as the state variables, recursive state estimation is performed through the configured Kalman filter. Based on the state estimation results at the current moment, the obstacle movement trajectory within the preset time domain is extrapolated forward, and the occupancy trend of each sector is determined according to the obstacle movement trajectory.

4. The autonomous obstacle avoidance control method for low-altitude unmanned aerial vehicles (UAVs) based on a sector risk grid according to claim 1, characterized in that, The calculation of the risk value for each sector based on the distance to the nearest obstacle within each sector, the unknown area marker, the relative approach speed, the measurement confidence level, the terrain height margin, and the obstacle type includes: The collision time is determined based on the distance to the nearest obstacle and the relative approach speed. The measurement confidence level is determined based on the sensor echo signal-to-noise ratio and the effective field of view coverage. Calculate the terrain height margin based on the current flight altitude and the terrain height in the sector direction; The risk value of each sector is obtained by weighting and fusing the nearest obstacle distance, collision time, measurement confidence level, terrain height margin, unknown area marker, and obstacle type.

5. The autonomous obstacle avoidance control method for low-altitude unmanned aerial vehicles (UAVs) based on a sector risk grid according to claim 4, characterized in that, The sector risk value corresponding to the dynamic obstacle is determined based on the obstacle's occupancy trend, specifically including: For the dynamic obstacle, the risk value of each sector obtained by weighted fusion is used as the base risk value; Based on the occupancy trend, the base risk value of the sector predicted to be occupied by the dynamic obstacle in the future time period is modulated upwards; The base risk value of the sector currently occupied by the dynamic obstacle but no longer occupied in the future is down-modulated.

6. The autonomous obstacle avoidance control method for low-altitude unmanned aerial vehicles (UAVs) based on a sector risk grid according to claim 1, characterized in that, The process of generating a candidate control command set under the flight dynamics constraints of the unmanned aerial vehicle includes: Discrete sampling is performed within the preset control command parameter space, and the sampled values ​​of each dimension are combined into multiple candidate command vectors; The candidate command vectors are verified one by one according to the flight dynamics constraints of the unmanned aerial vehicle; Candidate instruction vectors that do not meet any of the constraints are removed, and the remaining candidate instruction vectors are integrated into a candidate control instruction set.

7. The autonomous obstacle avoidance control method for low-altitude unmanned aerial vehicles (UAVs) based on a sector risk grid according to claim 1, characterized in that, The process of performing trajectory extrapolation on the candidate control command set, and calculating the comprehensive cost based on the risk value of the extrapolated trajectory in each sector and its deviation from the mission trajectory, includes: Using each candidate instruction in the candidate control instruction set as input, the corresponding predicted trajectory is recursively generated within a preset prediction time domain. By accumulating the risk value of the sectors traversed along each of the predicted paths, the risk cost corresponding to each candidate instruction is obtained. Calculate the deviation between the predicted end position of each trajectory and the corresponding target point on the current mission trajectory to obtain the trajectory deviation cost corresponding to each candidate command; The energy consumption cost is determined based on the degree of difference between each candidate command and the current flight state. The instruction smoothing cost is determined based on the degree of difference between each candidate instruction and the instruction executed in the previous cycle. The cost of traversing the unknown region is determined based on the proportion of steps that fall into the unknown region in the predicted flight path. The risk cost, the track deviation cost, the energy consumption cost, the command smoothness cost, and the unknown area crossing cost are weighted and summed to obtain the comprehensive cost corresponding to each candidate command.

8. The autonomous obstacle avoidance control method for low-altitude unmanned aerial vehicles (UAVs) based on a sector risk grid according to claim 7, characterized in that, The formula for calculating the comprehensive cost is as follows: ； Where b1 is the track deviation cost weighting coefficient, E track (u) represents the track deviation cost, b2 is the risk cost weighting coefficient, and E risk (u) represents the risk cost, b3 is the energy consumption cost weighting coefficient, and E energy (u) represents the energy consumption cost, b4 is the smoothness cost weighting coefficient, and E smooth (u) represents the smoothness cost of the instruction, b5 is the weighting coefficient for the cost of traversing unknown regions, and E unknown (u) represents the cost of traversing the unknown region.

9. The autonomous obstacle avoidance control method for low-altitude unmanned aerial vehicles (UAVs) based on a sector risk grid according to claim 1, characterized in that, When executing the execution instruction, a graded obstacle avoidance strategy is implemented according to the risk level: When the risk value reaches the first threshold, the target course is corrected by a preset angle. When the risk value reaches a second threshold that is greater than the first threshold, deceleration is performed and the lateral search weight is increased; When the risk value reaches a third threshold that is greater than the second threshold, if there is a low-risk passage to the side or above, a lateral detour or ascent shall be performed. If the risk values ​​of all sectors traversed by the candidate control instruction set are higher than the third threshold, then a security protection mode is executed.

10. The autonomous obstacle avoidance control method for low-altitude unmanned aerial vehicles (UAVs) based on a sector risk grid according to claim 9, characterized in that, The hierarchical obstacle avoidance strategy includes a hysteresis handling mechanism: There are separate action entry and action exit thresholds for each level of obstacle avoidance action; When executing the execution instruction, the obstacle avoidance action is triggered only when the number of control cycles in which the risk value continuously exceeds the corresponding risk level threshold reaches the action entry threshold. After an obstacle avoidance action is triggered, the current obstacle avoidance action is allowed to exit only when the number of control cycles in which the risk value is continuously lower than the preset release threshold reaches the action exit threshold. Wherein, the action exit threshold is greater than the action entry threshold.

Images