A dual-engine decision dynamic obstacle avoidance method and system for an aerial vehicle

Abstract

The application relates to an aerial vehicle dual-engine decision dynamic obstacle avoidance method and system, which comprises the following steps: obtaining a radar echo, performing spectrum transformation and constant false alarm detection to generate a point cloud containing relative distance, azimuth angle and radial velocity; calculating the relative speed of the obstacle relative to the aerial vehicle; performing multi-model interaction filtering and state extrapolation based on the radial velocity to generate a future probability position distribution of the obstacle; calculating a threat index and judging whether a preset threshold is reached; if not, performing space-time grid search and collision detection based on the probability position and the threat index to generate a flyable trajectory; if yes, performing artificial potential field calculation based on the relative distance and the radial velocity to generate an emergency obstacle avoidance instruction; and outputting the instruction to a flight control system. The application realizes seamless switching of conventional path planning and emergency obstacle avoidance by fusing radar multi-model prediction and a dual-engine decision mechanism, and improves the real-time response capability and flight safety of the aerial vehicle in a complex dynamic environment.

Description

Technical Field

[0001] This application relates to the field of automatic control technology for aircraft, and in particular to a dynamic obstacle avoidance method and system for dual-engine aircraft. Background Technology

[0002] With the widespread use of aircraft in low-altitude airspace and the continuous increase in air traffic density, the collision risks faced by aircraft when performing missions in complex and dynamic environments are becoming increasingly prominent. To ensure flight safety, aircraft must possess a high degree of autonomous perception and obstacle avoidance capabilities. Existing obstacle avoidance systems typically rely on sensors such as airborne radar to acquire physical information about the surrounding environment and use pre-set algorithms to identify, track, and predict the trajectory of obstacles, thereby planning a safe flight path.

[0003] In the complex dynamic airspace of low altitudes, aircraft not only need to cope with stationary terrain obstacles, but also with other high-speed moving aircraft or sudden objects. This requires obstacle avoidance systems to complete the entire process from environmental perception to control command generation in an extremely short time. Due to the high flight speed and large inertia of aircraft, obstacle avoidance systems not only need to ensure the smoothness and feasibility of the planned path, but also need to have millisecond-level emergency response capabilities when facing extreme and sudden threats, in order to avoid catastrophic consequences caused by obstacle avoidance delays.

[0004] However, traditional target tracking and trajectory prediction algorithms mostly rely solely on target position information for state filtering and extrapolation. When facing dynamic obstacles that are moving at high speed or suddenly changing direction, they often exhibit severe prediction lag, failing to accurately predict the obstacle's future movement trend and leading to distortion of the situational information upon which obstacle avoidance planning is based. Existing obstacle avoidance decision-making systems typically employ a single path planning architecture. When faced with sudden, high-risk approach situations, complex global spatiotemporal search algorithms, limited by massive computational loads and iteration delays, often cannot meet the real-time requirements of extreme avoidance, easily missing the optimal obstacle avoidance opportunity.

[0005] Furthermore, conventional obstacle avoidance strategies lack in-depth quantification of the instantaneous approach of the target and adaptive matching with the current flight conditions of the aircraft when assessing collision risks. This makes it easy for the system to misjudge or miss when distinguishing between general disturbances and real high-risk threats, making it difficult to achieve reliable emergency response in extreme and sudden scenarios while ensuring the global optimization of the conventional flight path. Summary of the Invention

[0006] To address the aforementioned technical issues, this application provides a method and system for dynamic obstacle avoidance with dual engines in aircraft.

[0007] Firstly, this application provides a dynamic obstacle avoidance method for dual-engine aircraft, employing the following technical solution: The radar echo signal is acquired, and the radar echo signal is subjected to spectrum transformation and constant false alarm rate (CFAR) detection processing to generate point cloud data; wherein, the point cloud data includes the relative distance, azimuth angle and radial velocity between the aircraft and the obstacle; Based on the radial velocity, the point cloud data is subjected to multi-model interactive filtering and state extrapolation to generate probability position distribution data of obstacles at future times; The velocity vector of the aircraft is obtained, and the relative velocity of the obstacle with respect to the aircraft is calculated by vector difference based on the velocity vector of the obstacle in the same coordinate system output by the multi-model interactive filtering. Based on the relative distance, relative speed, and radial speed, the threat index of the obstacle is calculated and generated. Determine whether the threat index has reached a preset risk threshold; If the threat index does not reach the preset risk threshold, then the conventional planning engine performs spatiotemporal grid search and collision detection processing based on the probability location distribution data and the threat index to generate flyable trajectory data. If the threat index reaches a preset risk threshold, an emergency obstacle avoidance command is generated by performing artificial potential field calculation based on the relative distance and radial velocity through the emergency reflection engine. The flyable trajectory data or the emergency obstacle avoidance command is output to the flight control system.

[0008] By adopting the above technical solutions, an aviation obstacle avoidance architecture with deep software and hardware collaboration and dual-engine redundant decision-making was constructed. High-precision Doppler radial velocity was used as the core feature throughout the entire process of perception, prediction, assessment, and emergency response, solving the problem of lag in predicting high-speed maneuvering targets in traditional algorithms. Through adaptive quantization and route determination of the four-dimensional threat index, an organic combination of conventional spatiotemporal global planning and emergency artificial potential field reflection was achieved. This ensures both optimal path and stability for conventional flight while guaranteeing millisecond-level emergency response capabilities in extreme and sudden scenarios, comprehensively improving the survivability and autonomous obstacle avoidance capabilities of aircraft in complex and dynamic airspace.

[0009] Secondly, this application provides a dual-engine decision-making dynamic obstacle avoidance system for aircraft, which adopts the following technical solution: A radar signal processing module is used to acquire radar echo signals, perform spectrum transformation and constant false alarm rate (CFAR) detection on the radar echo signals, and generate point cloud data; wherein, the point cloud data includes the relative distance, azimuth angle, and radial velocity between the aircraft and the obstacle; The trajectory prediction module is used to perform multi-model interactive filtering and state extrapolation processing on the point cloud data based on the radial velocity to generate probability position distribution data of obstacles at future times. The relative velocity calculation module is used to obtain the velocity vector of the aircraft and, based on the obstacle velocity vector in the same coordinate system output by the multi-model interactive filtering, calculate the relative velocity of the obstacle relative to the aircraft through vector difference. The threat assessment module is used to calculate and generate a threat index for the obstacle based on the relative distance, relative speed, and radial speed. The risk assessment module is used to determine whether the threat index has reached a preset risk threshold; The conventional obstacle avoidance planning module is used to generate flyable trajectory data by performing spatiotemporal grid search and collision detection processing based on the probability location distribution data and the threat index through the conventional planning engine when the threat index does not reach the preset risk threshold. The emergency obstacle avoidance calculation module is used to generate an emergency obstacle avoidance command by performing artificial potential field calculation based on the relative distance and radial velocity through the emergency reflection engine when the threat index reaches a preset risk threshold. The command output module is used to output the flyable trajectory data or the emergency obstacle avoidance command to the flight control system.

[0010] Thirdly, this application provides a computer-readable storage medium, which adopts the following technical solution: A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as in any of the methods in the first aspect. Attached Figure Description

[0011] Figure 1 This is a first flowchart illustrating a dual-engine decision-making dynamic obstacle avoidance method for aircraft, one embodiment of this application.

[0012] Figure 2 This is a second flowchart illustrating the dynamic obstacle avoidance method for dual-engine aircraft decision-making, one embodiment of this application.

[0013] Figure 3 This is a schematic diagram of the third process of a dual-engine decision-making dynamic obstacle avoidance method for aircraft, which is one embodiment of this application.

[0014] Figure 4 This is a schematic diagram of the fourth process of a dual-engine decision-making dynamic obstacle avoidance method for aircraft, according to one embodiment of this application.

[0015] Figure 5 This is a schematic diagram of the fifth process of a dual-engine decision-making dynamic obstacle avoidance method for aircraft, according to one embodiment of this application.

[0016] Figure 6 This is a schematic diagram of the sixth process of a dual-engine decision-making dynamic obstacle avoidance method for aircraft, which is one embodiment of this application.

[0017] Figure 7 This is a schematic diagram of the seventh process of a dual-engine decision-making dynamic obstacle avoidance method for aircraft, according to one embodiment of this application.

[0018] Figure 8 This is a schematic diagram of the eighth process of a dual-engine decision-making dynamic obstacle avoidance method for aircraft, according to one embodiment of this application.

[0019] Figure 9 This is a block diagram of a single-channel PE unit circuit structure according to one embodiment of this application. Detailed Implementation

[0020] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1-9 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0021] Currently, the autonomous obstacle avoidance capability of aircraft in complex airspace environments is a core element in ensuring flight safety. Millimeter-wave radar, with its unique advantages such as long-term stable operation, strong resistance to electronic interference, and the ability to directly measure the radial velocity of targets based on the Doppler effect, has become one of the core sensors for achieving highly reliable environmental perception on aviation platforms, playing a crucial role in obstacle avoidance perception in complex scenarios.

[0022] However, existing airborne obstacle avoidance systems based on millimeter-wave radar still face the following technical bottlenecks: First, the limited airborne computing power and the contradiction between real-time performance make it difficult for traditional embedded processing platforms to simultaneously meet the high throughput requirements of radar signal processing and the iterative calculations of complex path planning algorithms. In aviation scenarios, flight speeds are high (relative speeds can reach over 200 m / s), and every 10 ms increase in system latency means a relative displacement of more than 2 meters. Existing solutions generally have latency of 50-100 ms, which severely restricts the effectiveness of obstacle avoidance. Second, there is a lack of predictive ability for high-speed maneuvering targets. Existing systems are mostly passive obstacle avoidance systems based on "detection-response," lacking the ability to predict the future motion state of targets. Aviation targets have high maneuverability (instantaneous overload can reach 5-9g), and their trajectories are difficult to describe with a single model. Traditional filtering algorithms are prone to tracking drift or even loss when targets are maneuvering. The lack of advanced predictive ability means that when facing oncoming high-speed approaching targets, response lag will directly result in the loss of the avoidance window. Third, the airspace threat assessment dimension is singular. Most systems only judge the risk level based on distance, ignoring relative speed and radial approach. Key parameters such as speed and airspace geometry indicate fundamental differences in collision risk between oncoming high-speed targets and lateral crossing targets. Single-dimensional assessment cannot support correct obstacle avoidance decisions. Furthermore, risk preferences differ significantly across flight phases such as takeoff / climb, cruise, penetration, and landing, and existing systems lack adaptive weight adjustment mechanisms. Fourth, the real-time performance and safety of path planning are at odds in dynamic airspace. Traditional algorithms such as A* are limited to two-dimensional static space and cannot handle dynamic obstacles. While random sampling algorithms such as RRT can cope with changing environments, path quality is unstable and real-time convergence is difficult to guarantee. The serial architecture that separates planning and collision detection further exacerbates latency, making it difficult to meet the real-time replanning requirements in dynamic airspace. Fifth, there is a lack of redundant safety assurance mechanisms in emergency scenarios. Any complex planning algorithm has a fixed iteration delay. In extreme scenarios such as sudden threats (e.g., missile attacks, obstacles suddenly entering the flight path), waiting for the main planning engine to complete path replanning will miss the best opportunity for hazard avoidance. Existing systems lack an emergency response channel independent of the main planning engine, leaving flight safety without a last line of defense.

[0023] Based on this, this application discloses a dual-engine decision-making dynamic obstacle avoidance method for aircraft. Through hardware acceleration and a dual-engine decision architecture, it can solve the problems of limited onboard computing power, delayed prediction of maneuvering targets, and lack of emergency avoidance, and achieve high real-time and high-reliability autonomous obstacle avoidance capability in complex airspace.

[0024] Firstly, at the hardware level, this application proposes an integrated system design based on an FPGA+ARM heterogeneous computing platform to address the real-time obstacle avoidance requirements of airborne platforms for dynamic obstacles. The core idea is "heterogeneous collaboration, prediction-first, and dual-engine redundancy." Specifically, leveraging the high parallelism of the FPGA hardware pipeline, it handles fixed logic tasks such as radar 1D / 2D-FFT, CFAR detection, and 64-channel parallel collision detection, compressing latency to the nanosecond level. The ARM module is responsible for IMM multi-model filtering, STA* node expansion, and global system scheduling. The IMM module runs three Kalman filters (CV / CA / CT) in parallel, using Doppler velocity as the direct observation to output the probability position distribution within the next 2 seconds, enabling the system to have "advanced prediction" capabilities. At the decision-making level, the system adopts a dual-engine architecture of STA and APF. In normal scenarios, the STA main engine searches for the optimal path in a three-dimensional spatiotemporal grid. A heuristic function fuses a four-dimensional threat index composed of distance, velocity, geometric collision zone coefficient, and radial velocity square term to achieve a balance between safety and efficiency. When the threat index is ≥0.8, the APF emergency reaction arc module directly takes over. Its repulsion coefficient is proportional to the radial velocity square, and it outputs an emergency obstacle avoidance command within 5ms, forming a dual insurance mechanism of "STA* for normal use and APF for emergencies". The overall solution fully meets the high real-time and high reliability requirements of airborne obstacle avoidance through deep collaboration between hardware acceleration and software decision-making.

[0025] Reference Figure 1 A dual-engine decision-making dynamic obstacle avoidance method for aircraft, which employs a dual-engine decision-making architecture including a conventional planning engine and an emergency response engine, specifically including: Step S101: Acquire radar echo signal, perform spectrum transformation and constant false alarm rate (CFAR) detection on radar echo signal to generate point cloud data; wherein, point cloud data includes the relative distance, azimuth angle and radial velocity between aircraft and obstacles. The spectral transformation of the radar echo signal is performed to convert the time-domain echo signal to the frequency domain, thereby calculating the target's distance and velocity relative to the aircraft. Constant false alarm rate (CFAR) detection, on the other hand, dynamically adjusts the detection threshold based on the surrounding noise level in complex environments with ground clutter and electromagnetic interference, filtering out clutter with a constant false alarm probability to extract the true target point.

[0026] Specifically, the millimeter-wave radar transmits continuous waves and receives reflected echoes. The system first digitally samples the simulated echoes. Then, it performs Fast Fourier Transforms (FFTs) in the range and velocity dimensions, obtaining range and velocity parameters by analyzing spectral peaks. In this process, radial velocity is calculated based on the Doppler frequency shift principle. Compared to traditional sensing methods that rely solely on position information, radial velocity can reflect target maneuvering changes in a single frame. The constant false alarm rate (CFAR) algorithm, by setting protection and reference cells and statistically analyzing local background noise power, ensures stable separation of effective target points containing signal-to-noise ratio (SNR) information in clutter environments with varying intensity.

[0027] Ultimately, in the generated point cloud data, the relative distance represents the spatial position of the target, the azimuth angle represents the target's orientation relative to the carrier aircraft, and the radial velocity accurately reflects the instantaneous motion state of the target toward or away from the aircraft. Furthermore, in some embodiments, the point cloud data may also include a signal-to-noise ratio (SNR) to characterize the intensity of the echo signal and the data's reliability, thus forming four-dimensional point cloud data and providing a quality basis for subsequent data screening.

[0028] For example, in a scenario where a drone is flying low over a canyon, the mountain walls generate a large amount of radar clutter. Constant false alarm rate (CFAR) detection can effectively eliminate this stationary or slowly moving background clutter. Meanwhile, the point cloud data of another drone flying in the opposite direction contains not only distance information, but also radial velocity information indicating high-speed approach, azimuth angle changes, and high signal-to-noise ratio characteristics. This feature provides the most direct physical parameters for subsequent threat assessment and trajectory prediction.

[0029] Step S102: Based on radial velocity, perform multi-model interactive filtering and state extrapolation on the point cloud data to generate probability position distribution data of obstacles at future times; The core of multi-model interactive filtering is to simultaneously run multiple filters describing different motion states, adapting to abrupt changes in target motion through probability transfer and weighted fusion between models. State extrapolation, based on the optimal estimated state at the current moment, extrapolates forward according to the time step to predict the target's possible positions over a future period. The probability position distribution data not only includes the mean prediction of future positions but also a covariance matrix representing prediction uncertainty, providing a probabilistic boundary for subsequent collision detection.

[0030] In this embodiment, the system runs Kalman filters for three motion models—uniform velocity, uniform acceleration, and constant angular velocity turning—in parallel. Traditional filtering often relies solely on position observations, leading to tracking lag during target maneuvers. This scheme uses the radial velocity generated in the previous step as a direct observation input, establishing an analytical relationship between it and the Cartesian velocity, enabling the filter to sensitively capture changes in the target's motion state within a single frame. Subsequently, the system calculates the Gaussian likelihood based on the matching degree between each model and the observation data, dynamically updates the probability weights of each model, weights and fuses them to output the current optimal estimate, and extrapolates this state to multiple future moments along the time axis.

[0031] For example, when an aircraft suddenly turns, the probability weight of the constant angular velocity turning model increases rapidly, while the weight of the uniform velocity model decreases. This adaptive switching avoids trajectory drift under a single model. The extrapolated probability location distribution data essentially provides a probability region where obstacles may appear in the future. The further into the future the prediction, the larger this region becomes. This expression of uncertainty provides sufficient safety margin for subsequent path planning.

[0032] Step S103: Obtain the velocity vector of the aircraft, and calculate the relative velocity of the obstacle with respect to the aircraft by vector difference based on the obstacle velocity vector in the same coordinate system output by multi-model interactive filtering. The velocity vector of an aircraft is typically provided by an onboard inertial measurement unit (IMU) or flight control system, using a navigation coordinate system (such as a north-south coordinate system) as a reference, reflecting the real-time motion state of the aircraft. The velocity vector of an obstacle, on the other hand, is output by an IMM filter. Initially, it is a component in the radar coordinate system and needs to be transformed to the navigation coordinate system through an attitude matrix to align with the aircraft's velocity vector, thus avoiding motion distortion caused by coordinate system differences.

[0033] In this embodiment, to unify the coordinate system, the obstacle velocity in the radar coordinate system needs to be converted to the navigation coordinate system. The conversion process uses the pitch, roll, and yaw angles measured by the IMU to calculate the attitude matrix, and projects the obstacle velocity vector onto the navigation coordinate system. Subsequently, the relative velocity is calculated through vector difference, that is, the relative velocity vector is obtained by subtracting the aircraft velocity vector from the obstacle velocity vector, and then its magnitude is calculated to obtain the relative velocity scalar.

[0034] For example, during the cruise phase of an aircraft, the IMU measures the aircraft's velocity vector as [50,0,0] m / s (in the navigation coordinate system, along the positive x-axis), while the IMM filter outputs the obstacle's velocity vector as [60,10,0] m / s (in the radar coordinate system). Through attitude matrix transformation, the obstacle's velocity vector becomes [55, 5,0] m / s in the navigation coordinate system. After vector difference, the relative velocity vector is [5,5,0] m / s, with a magnitude of approximately 7.07 m / s. This value reflects the overall approach strength of the obstacle relative to the aircraft.

[0035] Step S104: Calculate the threat index of the generated obstacle based on relative distance, relative velocity, and radial velocity; In addition, a single parameter such as distance or speed cannot accurately describe the complex airspace situation. For example, the collision risk of a target approaching at high speed head-on is drastically different from that of a target moving laterally at a constant speed across the flight path, even if the distances are the same. Therefore, the threat index is used to comprehensively measure the degree of danger posed by obstacles to aircraft.

[0036] In this embodiment, a geometric collision zone coefficient and an adaptive weighting coefficient are also introduced when calculating the threat index. The geometric collision zone coefficient is determined based on the relative heading angle between the aircraft and the obstacle; a high-risk value is used for an oncoming situation, a medium value for a lateral situation, and a low value for a distant situation. Specifically, by obtaining the current flight speed vector direction information of the aircraft, the relative heading angle between the aircraft and the obstacle can be calculated based on the flight speed vector direction information and the azimuth angle.

[0037] More importantly, the adaptive weighting coefficients are dynamically adjusted according to the current flight phase of the aircraft. For example, during takeoff, distance perception is emphasized; during cruise, radial velocity weighting is strengthened; and during landing, geometric collision zone risk is prioritized. This mechanism breaks the limitations of fixed parameter assessments, allowing the threat index to adapt to different flight conditions.

[0038] For example, during the high-speed cruise phase of an aircraft, if a target is detected approaching at an extremely high radial velocity, the contribution of the squared radial velocity term to the threat index increases dramatically. Combined with the higher weight given to this term during the cruise phase, the threat index will spike instantly. This quantification method directly reflects the extreme danger of a high-speed oncoming collision, providing reliable data for triggering emergency obstacle avoidance.

[0039] Step S105: Determine whether the threat index has reached the preset risk threshold; if not, proceed to step S106; if yes, proceed to step S107. The risk threshold serves as a critical criterion for the system to switch decision-making modes, distinguishing between normal safe situations and emergency high-risk situations. Due to the significant inertia and high speed of aircraft, waiting for the system to complete complex global path planning may delay hazard avoidance opportunities in certain extreme scenarios. Therefore, this threshold is set as a routing switch for the dual-engine architecture. When the threat index is below this risk threshold, the system considers there sufficient time for global planning; once the threshold is exceeded, the system determines that an imminent collision risk is present.

[0040] Specifically, in this embodiment, the preset risk threshold is set as an empirical constant. The system compares the real-time calculated threat index with this threshold during each frame of data processing. This judgment logic is extremely lightweight, does not involve complex iterative calculations, and ensures extremely low latency in decision routing. Based on the comparison results, the system outputs different control signals to guide the data flow into the main planning engine or the emergency response arc engine.

[0041] For example, when an aircraft is flying normally in a canyon, the threat index of static obstacles such as mountain walls is usually kept at a low level, far below the risk threshold, and the system continues to operate in normal planning mode. However, if a bird or drone suddenly cuts into the flight path from the side at high speed, the surge in its radial and relative speeds will cause the threat index to exceed the threshold in a very short time. At this point, the judgment logic is immediately triggered, and the system enters an emergency response state.

[0042] Step S106: Using a conventional planning engine, spatiotemporal grid search and collision detection are performed based on probability location distribution data and threat index to generate flyable trajectory data. Among these features, the spatiotemporal grid search breaks through the limitations of traditional two-dimensional spatial planning by incorporating the time dimension into the search space. This ensures that each grid node not only contains spatial coordinates but also is bound to a specific timestamp. This strong spatiotemporal binding mechanism enables the planning algorithm to predict the position of dynamic obstacles in the future, thereby achieving proactive avoidance in the time domain. Collision detection processing utilizes previously generated probability position distribution data, treating the uncertain boundaries of obstacles as constraints rather than simple rigid points, thus improving detection accuracy. The threat index is incorporated as a cost penalty into the search heuristic function, guiding the path to proactively avoid high-threat areas.

[0043] In this embodiment, the system constructs a three-dimensional grid with forward, lateral, and temporal dimensions. During node expansion, for each candidate node, the system extracts the obstacle prediction distribution parameters corresponding to that moment and calculates the Mahalanobis distance between the candidate node and the obstacle probability distribution, using this as a collision criterion. If deemed safe, the node is added to the open list. Its cost in the heuristic function includes not only distance but also a threat index penalty, causing the algorithm to tend to select lower-risk areas while searching for the shortest path. After the search is complete, spline interpolation is used to smooth the polyline path, generating a flyable trajectory that satisfies the kinematic constraints of the aircraft.

[0044] For example, when an aircraft encounters multiple moving obstacles ahead, the spatiotemporal grid search considers whether the obstacles have also moved to a given point when the aircraft arrives at that point. If the probability distribution ellipse of the obstacles covers the candidate node, collision detection will eliminate it. Furthermore, even if a path has the shortest distance, if it passes through an area with a high threat index, its total cost will increase, and the algorithm will ultimately select a slightly longer but safer smooth trajectory.

[0045] It should be noted that when the point cloud data is empty or the threat index of all obstacles is below the minimum threshold, the pre-configured preset cruise trajectory can be output directly.

[0046] Step S107: Using the emergency reflection engine, an artificial potential field is calculated based on the relative distance and radial velocity to generate an emergency obstacle avoidance command. The artificial potential field calculation and processing is a fast obstacle avoidance algorithm that simulates the interaction of gravity and repulsion, serving as the system's emergency reaction arc engine. In emergency scenarios, complex global search algorithms are limited by computational iteration latency and cannot meet millisecond-level obstacle avoidance requirements. This solution improves upon the traditional artificial potential field by positively correlating the repulsive force intensity with the square of the obstacle's radial velocity, resulting in a significant repulsive force from a high-speed approaching obstacle, thus achieving an adaptive nonlinear amplification of the obstacle avoidance force.

[0047] Specifically, when the threat index exceeds the limit, the system immediately interrupts the regular path planning link and directly reads the latest radar point cloud data. For each detected obstacle, the system calculates the reciprocal of its relative distance as the basic repulsive force, and adds the product of the dynamic repulsive force coefficient and the square of the radial velocity. If the relative distance is less than the near-field danger threshold, the system will directly output the maximum emergency repulsive force vector to ensure that no collision occurs. The summation of the repulsive force vectors of all obstacles is converted into an acceleration command, which is then limited to prevent aircraft instability.

[0048] For example, when faced with a missile or high-speed drone that suddenly enters the flight path, its radial velocity is extremely high. The repulsive force calculated by the artificial potential field will increase exponentially. The generated emergency obstacle avoidance command can drive the aircraft to make the maximum permissible evasive acceleration maneuver within milliseconds.

[0049] Step S108: Output flyable trajectory data or emergency obstacle avoidance command to the flight control system.

[0050] The flight control system is the underlying hardware mechanism that executes changes in the attitude and trajectory of an aircraft. Whether it's a smoothed global flyable trajectory in normal mode or an acceleration obstacle avoidance command generated instantaneously in emergency mode, both ultimately need to be translated into control signals for execution. Simultaneously, during the command output process, the system also needs to handle the smooth transition during dual-engine switching to prevent sudden changes in control commands from causing the aircraft to lose control.

[0051] Specifically, the system sends trajectory point sequences or acceleration commands to the flight controller via a highly reliable bus interface. When recovering from emergency mode to normal mode, the system does not immediately cut off emergency commands but continuously monitors the threat index. Only when the threat index falls below the risk threshold, and the regular spatiotemporal grid search has completed a full replanning and generated a new flyable trajectory, does the system smoothly switch control and output regular trajectory data, ensuring a smooth transition from extreme evasive maneuvering to normal cruise.

[0052] For example, after an aircraft successfully avoids an oncoming target using an emergency obstacle avoidance command, the threat index decreases as the relative distance increases. During this period, the system restarts a spatiotemporal grid search in parallel. Once a new path is generated and confirmed to be safe, the control weight of the artificial potential field command is gradually reduced, and flight control authority is transferred to the new flyable trajectory data, thus preventing the aircraft from experiencing trajectory jitter or heading loss after the obstacle avoidance maneuver.

[0053] In this embodiment, the above steps constitute a dual-engine decision-making dynamic obstacle avoidance architecture for aircraft. "Dual-engine" specifically refers to two sets of decision-making logic branches working in parallel and collaboratively: the logic branch that performs spatiotemporal grid search and collision detection based on probabilistic position distribution data and threat index constitutes the system's "conventional planning engine." This engine focuses on global optimal path search, using spatiotemporal grids and probability distributions to calculate long-term optimal flyable trajectory data, ensuring flight stability and economy. The logic branch that performs artificial potential field calculation based on relative distance and radial velocity constitutes the system's "emergency reflection engine." This engine, based on the principle of artificial potential field, directly generates repulsive force using radar-measured radial velocity and relative distance, thereby outputting emergency obstacle avoidance commands within milliseconds. The system acts as a "clutch" by real-time monitoring of the relationship between the threat index and preset risk thresholds. Under normal circumstances, the conventional planning engine dominates; under extremely dangerous circumstances, it immediately switches to the emergency reflection engine, achieving a perfect fusion of global optimal planning and local real-time risk avoidance.

[0054] The above implementation constructs an aviation obstacle avoidance architecture with deep hardware and software collaboration and dual-engine redundant decision-making. It utilizes high-precision Doppler radial velocity as a core feature spanning the entire chain of perception, prediction, assessment, and emergency response, solving the problem of lag in predicting high-speed maneuvering targets in traditional algorithms. Through adaptive quantization and route determination of the four-dimensional threat index, it achieves an organic combination of conventional spatiotemporal global planning and emergency artificial potential field reflection. This ensures both optimal path and stability for conventional flight while guaranteeing millisecond-level emergency response capabilities in extreme and sudden scenarios, comprehensively enhancing the survivability and autonomous obstacle avoidance capabilities of aircraft in complex and dynamic airspace.

[0055] Reference Figure 2 As one implementation of step S101, the steps of acquiring radar echo signals, performing spectrum transformation and constant false alarm rate (CFAR) detection on the radar echo signals, and generating point cloud data include: Step S201: Perform analog-to-digital conversion on the radar echo signal to generate a digital echo signal; Analog-to-digital conversion (ADC) refers to using an analog-to-digital converter to discretize and quantize the amplitude of the analog echo signal according to a preset sampling rate, and then converting it into a digital echo signal sequence that can be recognized by a computer.

[0056] Specifically, the system is deployed on the programmable logic side of a heterogeneous computing platform, which is equipped with a multi-channel analog-to-digital conversion interface. The system synchronously acquires analog echo signals from multiple radar intermediate frequency channels at a fixed sampling rate, completing the synchronous conversion from analog to digital signals. By employing a hardware design that combines differential transmission and high-speed sampling, the system can completely preserve the amplitude and phase information of the original echo time domain in airborne environments with strong electromagnetic interference, providing lossless digital source data for subsequent high-precision spectrum analysis.

[0057] Step S202: Perform range-dimensional fast Fourier transform and velocity-dimensional fast Fourier transform on the digital echo signal in sequence to extract and generate target range parameters and Doppler frequency parameters; The range-dimensional Fast Fourier Transform (FFT) maps the one-dimensional time-domain digital echo signal to the range domain, extracting the target range parameters by analyzing the peak positions in the spectrum. The velocity-dimensional FFT, building upon the range-dimensional processing, performs a secondary frequency domain transformation on echo signals with multiple slow time periods within the same range cell, used to analyze the Doppler frequency parameters caused by target motion.

[0058] In this embodiment, the system first performs a first-level Fast Fourier Transform (FFT) on the digital echo signal, achieving extremely high range resolution in the range domain by leveraging the wideband characteristics of the radar signal, thus accurately distinguishing nearby adjacent targets. After completing the range dimension processing, the system performs a second-level FFT on the continuous echo data of the same range cell. Since the relative motion of the target causes changes in the echo phase, the velocity dimension transform can accurately extract the Doppler frequency parameters corresponding to this phase change, thereby providing frequency domain data support for subsequent analysis of the instantaneous velocity of the target.

[0059] For example, when multiple obstacles are close to each other but moving at different speeds in front of an aircraft, they are difficult to distinguish in the time or range domains alone. By using a range-dimensional Fast Fourier Transform (FFT), the system can spatially separate these targets; and by using a velocity-dimensional FFT, the system can further distinguish the unique motion characteristics of each target on the Doppler frequency axis, making the stationary mountain clutter and the moving aircraft exhibit distinctly different peak distributions in the two-dimensional spectrum, laying the foundation for subsequent target detection and parameter calculation.

[0060] Step S203: Generate radial velocity based on Doppler frequency parameters; There is a strict physical proportionality between the Doppler frequency and the radial velocity of the target. Radial velocity refers to the velocity component of the target relative to the aircraft along the line connecting the two, directly representing the instantaneous motion of the target approaching or moving away from the aircraft.

[0061] In this embodiment, after acquiring the Doppler frequency parameters, the system directly calculates the radial velocity based on the correspondence between the radar transmission wavelength and the Doppler frequency shift. Compared to traditional systems that estimate velocity through position differences across consecutive frames, this method of directly resolving radial velocity from the Doppler frequency eliminates inter-frame delay, resulting in extremely high real-time performance and measurement accuracy. This high-precision radial velocity can directly meet the observation input requirements of the backend multi-model filter without requiring additional data smoothing processing, simplifying the algorithm architecture while ensuring the continuity of velocity tracking.

[0062] For example, when an enemy drone approaches head-on at extremely high speed, its radial velocity is considerable. By analyzing the Doppler frequency parameters, the system can directly obtain this approaching radial velocity value within a very short time of processing a single frame of radar data. This physical quantity directly reflects the urgency of the collision risk. If the velocity is estimated using traditional position difference estimation, not only is there a lag, but it will also produce a large calculation error when the target is maneuvering at high speed, thus delaying the obstacle avoidance opportunity.

[0063] Step S204: The target distance parameters are processed by the constant false alarm rate detection algorithm to remove clutter interference and generate point cloud data including the relative distance, azimuth angle, radial velocity and signal-to-noise ratio between the aircraft and the obstacle.

[0064] Specifically, in airborne radar detection scenarios, the intensity of ground clutter, weather interference, and internal system noise are dynamically changing. If a fixed threshold is used for detection, a large number of false alarms will occur when the background noise increases, while missed alarms will occur when the background noise decreases. The constant false alarm rate detection algorithm sets reference and protection units around the target unit, statistically analyzes the power level of local background noise in real time, and dynamically calculates the detection threshold accordingly, ensuring that the true target can be extracted with a constant false alarm probability even when the clutter environment changes.

[0065] Specifically, the system employs a cell-average constant false alarm rate (CFAR) detection algorithm on a two-dimensional spectrum. The system incorporates protection units to prevent the target signal's energy from leaking into reference cells, while simultaneously utilizing the average local noise from surrounding reference cells. If the signal strength of the target cell exceeds a threshold dynamically calculated based on the noise average, it is determined to be a valid target; otherwise, it is discarded as clutter.

[0066] After this screening process, the system integrates discrete target points into structured point cloud data, which includes relative distance representing spatial location, azimuth angle representing the target's orientation relative to the carrier aircraft, radial velocity representing instantaneous approach situation, and signal-to-noise ratio representing signal quality.

[0067] For example, when flying at low altitudes in mountainous canyons, mountain walls and trees generate large areas of high-intensity clutter. A constant false alarm rate (CFAR) detection algorithm can automatically raise the detection threshold based on the local intensity of this clutter, eliminating large areas of clutter peaks. The actual dynamic aircraft that need to be avoided have signal strengths far exceeding the local background noise, allowing them to pass through the dynamic threshold filtering process. The data is ultimately output as point cloud data containing multi-dimensional physical features, providing clean and high-value environmental perception input for subsequent trajectory prediction and threat assessment systems.

[0068] In the above implementation, the system starts from the underlying hardware signal acquisition and, through the deep collaboration of analog-to-digital conversion, multidimensional fast Fourier transform, and constant false alarm rate (CFAR) detection algorithms, not only achieves efficient extraction of real targets in complex and cluttered environments, but more importantly, it directly calculates high-precision radial velocity information at the source of signal processing and integrates it into the point cloud data structure. This preprocessing method, which directly analyzes Doppler frequency shift into physical velocity parameters, provides a high-quality data source for subsequent multi-model trajectory prediction and dynamic threat assessment based on Doppler observations. It breaks through the perception lag bottleneck caused by traditional obstacle avoidance systems relying solely on position information, and improves the aircraft's perception accuracy and response capability in high-speed dynamic airspace environments.

[0069] Reference Figure 3 As one implementation of step S102, the step of generating probability location distribution data of obstacles at future times by performing multi-model interactive filtering and state extrapolation processing on point cloud data based on radial velocity includes: Step S301: Perform data association processing on the point cloud data using the nearest neighbor algorithm; The core purpose of point cloud data association is to accurately match the point cloud observation data detected by the radar in the current frame with the historical tracks stored in the system. In complex dynamic airspace, a single radar scan often outputs point clouds of multiple targets. Without a reliable association mechanism, the system will be unable to distinguish which point cloud belongs to which obstacle, leading to tracking drift or target loss when tracks intersect. The basic logic of the nearest neighbor algorithm is to calculate the spatial distance between the new point cloud and the predicted positions of all existing historical tracks, and bind the observation point with the smallest distance to the corresponding track, thereby ensuring the continuity and data accuracy of multi-target tracking.

[0070] In this embodiment, after receiving point cloud data containing relative distance, relative velocity, and radial velocity generated by the previous module, the system extracts the spatial coordinate information of each point cloud and compares it with the predicted center position of existing obstacle trajectories maintained by the system. The system sets a reasonable association threshold. When a new point cloud falls within the association threshold of a certain historical trajectory, and the spatial distance between the two is the smallest among all candidate matches, the system updates the state of the point cloud to the corresponding trajectory. For isolated point clouds that cannot be associated with any existing trajectory, the system marks them as newly discovered targets and establishes an initial trajectory for them, thereby achieving accurate data continuity in multi-target environments.

[0071] For example, when an aircraft encounters two drones flying in overlapping airspace, a single radar scan will output two point clouds that are close to each other. The system uses a nearest neighbor algorithm to calculate the deviation between these two point clouds and the predicted positions of the two drones in the previous frame, and assigns the point cloud with the smallest deviation to the corresponding trajectory. In this way, even if the flight paths of the two drones briefly intersect, the system can correctly distinguish their respective observation data based on the optimal matching principle of spatial distance, avoiding tracking confusion or target identification errors.

[0072] Step S302: Start three Kalman filter models in parallel: uniform speed model, uniform acceleration model, and constant angular velocity steering model. Understandably, because aircraft have extremely high degrees of freedom of maneuver, their motion state may frequently switch between level flight, acceleration / deceleration dive, or rotation. A single motion model cannot fully describe such complex motion characteristics and is prone to tracking lag when the target is maneuvering.

[0073] In this embodiment, the Kalman filter is an algorithm that uses the state equation of a linear system to optimally estimate the system state through system input and output observation data. The core framework of the interactive multi-model algorithm is formed by launching multiple motion models in parallel: a uniform velocity model is used to match steady-state cruise targets, a uniform acceleration model is used to capture variable-speed maneuvering targets, and a constant angular velocity steering model is specifically designed for changing-direction turning targets. These three models operate in parallel to cover all possible motion modes of the target.

[0074] Specifically, for each dynamic obstacle trajectory confirmed through data association, the system synchronously runs three independent Kalman filters in the background. Each filter performs independent state prediction and covariance update based on its own inherent state transition matrix. These three models do not exist in isolation, but rather constitute a model set, providing multivariate underlying state estimates for subsequent probability assessment and fusion. This parallel architecture ensures that regardless of the obstacle's current flight condition, the system always has one or more models whose outputs closely match its actual trajectory. For example, when a commercial airliner suddenly encounters turbulence during cruise and begins to accelerate and turn, a single uniform velocity model will produce a large prediction error due to the actual acceleration and angular velocity of the target. However, the parallel uniform acceleration model and constant angular velocity turning model in this system can capture the rate of change of velocity and the rate of change of heading, respectively, and output state predictions that closely match the actual maneuver, thus ensuring that the system maintains a complete lock on the target's state even when facing instantaneous high-G maneuvers.

[0075] Step S303: Using the radial velocity in the point cloud data as a direct observation input, construct the observation equations for radial velocity and Cartesian velocity to update the state of each Kalman filter model; In this context, radial velocity is the velocity component of the target along the electromagnetic line of sight, directly measured by radar through the Doppler effect, while Cartesian velocity is the velocity vector of the target in a Cartesian coordinate system. Traditional filtering algorithms typically update the state using only the position coordinates measured by radar as the observation input. This approach requires multiple frames of data to converge velocity estimation when the target is maneuvering at high speeds, resulting in significant response lag. The purpose of constructing the observation equation is to establish an analytical mapping relationship between the radial velocity in polar coordinates and the Cartesian velocity component in Cartesian coordinates, enabling the filter to directly utilize high-precision Doppler observations to correct internal state variables.

[0076] In this embodiment, the system uses the high-precision radial velocity resolved by the front-end signal processing module as a direct observation input, constructing an observation equation that includes the analytical relationship between radial velocity and Cartesian velocity. During the Kalman filter update phase, this equation is used to calculate the information between the actual and predicted observations. By introducing this radial velocity information, the filter can quickly detect minute changes in the target's motion state within a single frame data period and instantly correct the velocity state estimates and error covariance of each model.

[0077] For example, when an enemy aircraft approaches head-on at extremely high speed, although its lateral position changes very little in the radar field of view and its position observation appears stable, its radial velocity is extremely large and changes drastically. If only position observation is used, the filter may misjudge the target as stationary or moving at low speed; however, by introducing radial velocity as a direct observation, the system can capture this dangerous situation of high-speed approach within a single frame, immediately reflecting the target's true high-speed motion characteristics in the state update, and completely eliminating the response lag of traditional algorithms.

[0078] Step S304: Update the posterior probability weights of the three Kalman filter models based on the Gaussian likelihood function and fuse them in a weighted manner to generate fused state data; The Gaussian likelihood function quantifies the degree to which each motion model matches the current actual observation data. The closer the model's predicted observations are to the actual radar observations, the higher its likelihood probability density. The posterior probability weight represents the probability that a particular model can correctly describe the target's true motion state under the current observation conditions.

[0079] In this embodiment, since the target's motion mode is dynamically changing, fixed weights cannot adapt to this change. The system adaptively increases the weights of models that fit the current maneuver and decreases the weights of failed models by calculating the likelihood value and updating the weights in real time. Weighted fusion then sums the state estimates of the three models according to their respective update weights, outputting an optimal state estimate that integrates the advantages of all models.

[0080] Specifically, the system uses the innovation output and covariance matrix of each Kalman filter to calculate the probability density of each model by substituting it into the Gaussian likelihood function. Then, combining this with a pre-set model transition probability matrix, the posterior probability weights of each model are updated using the Bayesian law of total probability. Finally, the system weights and fuses the state vectors and covariance matrices of the three models—uniform velocity, uniform acceleration, and constant angular velocity turning—accelerates them according to the updated weights to generate the global fused state data for the current moment. This dynamic optimization mechanism allows the system to primarily rely on the uniform velocity model when the target is flying smoothly, while automatically and smoothly transitioning to the acceleration or turning model when the target is maneuvering.

[0081] For example, when the target begins to enter a turning maneuver, the prediction results of the constant angular velocity steering model are highly consistent with the actual radar observations, and its likelihood function value increases sharply, leading to a significant increase in the posterior probability weight of this model. Meanwhile, the prediction of the uniform velocity model is biased because it does not consider the turn, and its weight decreases accordingly. The system, through weighted fusion, outputs a comprehensive state that retains the stability characteristics of the uniform velocity model while fully incorporating the steering model's accurate capture of maneuver trends, ensuring the continuity and smoothness of the target's trajectory during attitude transitions.

[0082] Step S305: Extrapolate the fused state data to a future preset time period to generate probability location distribution data containing the location mean vector and the location covariance matrix.

[0083] State extrapolation refers to extrapolating the target's trajectory over a future period based on the optimal fused state estimate at the current moment, using kinematic equations. Since the prediction of future moments inherently involves uncertainty, and this uncertainty gradually amplifies as the extrapolation time increases, the prediction result is no longer a definite coordinate point, but a probability distribution. The position mean vector represents the most likely spatial location of the target at the future moment, while the position covariance matrix describes the uncertainty range of this predicted location in various directions, i.e., the shape and size of the probability envelope, providing a safety margin for subsequent obstacle avoidance planning.

[0084] In this embodiment, after generating the fused state data for the current moment, the system performs progressive extrapolation calculations based on motion parameters such as position and velocity contained in the state, according to a preset future time period. While extrapolating the position mean vector, the system simultaneously calculates the evolution of the prediction covariance matrix. Due to the accumulation of process noise, the further the extrapolation time, the larger the range of the probability ellipse corresponding to the generated covariance matrix. The system ultimately outputs this series of probabilistic position distribution data containing the mean and covariance, providing a predictive envelope with spatiotemporal uncertainty for subsequent threat assessment and path planning.

[0085] For example, the system predicts that a moving obstacle will reach a certain coordinate point in two seconds. However, due to the obstacle's potential slight maneuvers, its actual position after two seconds may diverge. The system-generated probabilistic position distribution data not only provides the most likely center point but also a probabilistic ellipse boundary that expands over time. When performing collision detection, the path planning module no longer treats the obstacle as a rigid point but instead uses this probabilistic ellipse envelope as a danger zone for avoidance calculations. This provides sufficient safety margin for aircraft and effectively addresses potential collision risks arising from prediction uncertainties.

[0086] In the above implementation, the system constructs an interactive multi-model prediction architecture based on direct observation of Doppler radial velocity. The nearest neighbor algorithm ensures accurate data correlation in multi-target scenarios, and the parallel interaction and adaptive weighted fusion of three models—uniform velocity, uniform acceleration, and constant angular velocity steering—solves the problem of single-model drift when tracking highly maneuverable targets. In particular, using high-precision radial velocity as a direct observation input enables the filter to respond to sudden changes in target maneuvering in a single frame, achieving a qualitative leap from passive detection to proactive prediction. The final output, containing probability position distribution data including the mean vector and covariance matrix, provides a spatiotemporal prediction foundation that balances foresight and safety margin for subsequent path planning, enhancing the situational awareness and hazard avoidance capabilities of aircraft in complex dynamic airspace.

[0087] Reference Figure 4As one implementation of step S104, the step of calculating the threat index of the generated obstacle based on relative distance, relative velocity, and radial velocity includes: Step S401: Based on the relative heading angle between the aircraft and the obstacle, the geometric collision zone coefficient is generated by segmented fixed values ​​and linear interpolation. The relative heading angle refers to the angle between the direction of the aircraft's flight velocity vector and the direction of the aircraft's line of sight to the obstacle. This angle directly determines the geometric encounter situation between the two in the airspace. In high-speed flight scenarios, even if an oncoming target and a laterally crossing target have the same distance and speed, the probability of collision and the ease of avoidance are fundamentally different. The geometric collision zone coefficient is used to quantify the risk weight of this geometric situation; it is a dimensionless value within a specific range. Piecewise fixed values ​​refer to directly assigning fixed empirical risk values ​​to different heading angle ranges, while linear interpolation is used to perform smooth transition calculations between two adjacent ranges to avoid abrupt changes in risk assessment at the angle critical point.

[0088] In this embodiment, the system first acquires the flight speed vector direction information of the aircraft and calculates the relative heading angle. Then, the system uses a preset segmented logic for judgment: when the relative heading angle is within the oncoming situation range, the geometric collision region coefficient is set to its highest value; when it is within the lateral situation range, it is set to a medium value; and when it is within the far-from-the-situation range, it is set to its lowest value. For transition angles outside these preset ranges, the system uses a linear interpolation algorithm to calculate the geometric collision region coefficient proportionally between two adjacent fixed interval values ​​based on the current angle, thereby generating a parameter that can continuously and smoothly reflect the geometric probability of airspace collisions.

[0089] For example, when an aircraft detects a target ahead during cruise, if the target is within the heading angle range of a near-head-on approach, the system directly assigns it the highest geometrical collision zone coefficient to represent an extremely high collision probability. If the target is within the range of a large lateral angle approach, the system assigns a medium geometrical collision zone coefficient. When the target's heading angle is in the critical transition region between head-on and lateral approaches, the system uses a linear interpolation algorithm to smoothly transition the geometrical collision zone coefficient from the highest value to a medium value as the angle gradually increases. This prevents abrupt changes in risk assessment levels due to small changes in angle, ensuring the continuity of situation assessment.

[0090] In some embodiments, |θ rel |<30° represents a head-on stance, A collision Take 1.0; 60° < |θ rel |<120° represents a horizontal orientation, A collision Take 0.5; |θ rel |>150° represents a distanced position, A collisionWe set the value to 0.2; for the remaining angle ranges, we used linear interpolation to ensure a smooth transition of the situation.

[0091] Step S402: Read the adaptive weighting coefficients corresponding to the current flight phase of the aircraft; Flight phases refer to the different motion conditions an aircraft experiences throughout its journey from takeoff to landing, typically including takeoff / climb, cruise, landing, and emergency maneuvers. During different flight phases, the aircraft's flight envelope, maneuverability, and risk tolerance vary significantly. Adaptive weighting coefficients are a set of dynamic parameters used to adjust the proportion of different risk dimensions, such as distance, speed, and geometric collision zone, in the overall threat assessment. By reading weighting coefficients that match the current flight phase, the system can overcome the limitations of traditional fixed-parameter assessments, deeply adapting the threat assessment logic to the aircraft's actual flight characteristics.

[0092] In this embodiment, the system has a pre-built weight configuration mapping table for each typical flight phase. The system acquires the aircraft's flight control status data in real time to identify the current flight phase and retrieves a corresponding set of adaptive weight coefficients from the mapping table. For example, during takeoff / climb, the system reads a combination of weight coefficients that emphasizes near-field perception; during cruise, it reads a combination of weight coefficients that enhances radial velocity perception. This process enables the system to have full-condition adaptive capability, dynamically adjusting the risk assessment strategy without manual intervention.

[0093] For example, during the descent phase, an aircraft is extremely sensitive to lateral geometric disturbances due to the need for precise attitude control and accurate descent trajectory maintenance. At this time, the system will significantly increase the weighting of the geometric collision zone coefficient, giving high priority to even distant lateral targets. However, once the aircraft enters the cruise phase, with high speed and stable altitude, the system automatically switches to a combination that enhances radial velocity perception, allowing the system to focus on preventing deadly threats from high-speed approaching objects.

[0094] Step S403: Calculate the threat index according to the following formula: ; Where TI is the threat index, R is the relative distance, and R max v represents the radar's maximum effective detection range. rel V is the relative velocity. max A is the maximum relative velocity threshold preset for the system. collision v is the geometric collision region coefficient. radial α represents the radial velocity, and α, β, γ, and δ are adaptive weighting coefficients.

[0095] Specifically, the threat index is a dimensionless quantitative indicator that comprehensively reflects the degree of danger posed by obstacles to aircraft, and is used as a direct criterion for triggering subsequent obstacle avoidance actions. The adaptive weighting coefficient satisfies weighting engineering constraints (e.g., α+β+γ+δ=1) and can be dynamically configured according to the flight phase (takeoff / climb, cruise, landing, emergency maneuver). For example, during takeoff, the focus is on distance risk (increase α); during cruise, the radial approach risk is strengthened (increase δ); and during landing, the geometric area coefficient is emphasized (increase γ), thus achieving a dynamic fusion assessment of multi-dimensional risks.

[0096] For example, this application differentiates the weights for four typical flight states: takeoff / climb, cruise, landing, and emergency maneuvering: during takeoff / climb, the distance dimension is emphasized (α=0.6); during cruise, the radial velocity weight is strengthened (δ=0.3); during landing, the risk of airspace geometric collision zones is emphasized (γ=0.3); and during emergency maneuvering, the radial velocity weight is maximized (δ=0.5). This mechanism deeply matches the threat assessment logic with the airborne mission scenario and flight risk characteristics, achieving adaptive and accurate situational quantification across all operating conditions. After the module completes the threat index (TI) calculation, it outputs the threat results using a three-level classification mechanism: TI < 0.6 indicates low risk and safe airspace situation; 0.6 ≤ TI < 0.8 indicates medium risk and requires continuous monitoring; TI ≥ 0.8 indicates high risk and requires immediate obstacle avoidance response.

[0097] In this embodiment, the system extracts the relative distance R and relative velocity v generated in the previous stage. rel Radial velocity v radial And the calculated geometric collision region coefficient A collision Simultaneously, the currently read adaptive weight coefficients α, β, γ, and δ are substituted into the weighted summation formula for comprehensive calculation. In this process, the synergistic effect of the radial velocity squared term and the adaptive weights allows the system to instantly increase the calculated threat index to a high level when facing an oncoming high-speed target, effectively amplifying the urgency of high-risk collision scenarios. The final generated threat index TI, as a scalar value, is directly output to the decision module for threshold comparison.

[0098] For example, when an aircraft encounters a high-speed oncoming target during the cruise phase, even though the target may be far away, its radial velocity is extremely high, causing the squared term of this velocity to contribute dramatically to the calculation. Combined with the higher adaptive weighting of radial velocity during the cruise phase, the threat index will quickly exceed the high-risk threshold. Conversely, if the target is flying slowly in the same direction, even at a closer distance, its radial velocity is extremely low, resulting in a weak contribution from the squared term, and the calculated threat index will remain at a low level. This calculation logic effectively avoids the system overreacting to non-threatening targets, ensuring that genuine collision risks can be quantified and identified immediately.

[0099] In the above implementation, the system constructs a four-dimensional threat assessment model that integrates distance, velocity, geometric situation, and Doppler radial velocity. Combined with an adaptive weighting mechanism during flight phases, it solves the problem that traditional single-dimensional assessments cannot adapt to complex dynamic airspace. In particular, the introduction of the radial velocity square term and the synergistic effect of dynamic weights enable the system to nonlinearly amplify the threat level of extremely dangerous head-on high-speed approach attacks, achieving accurate quantification of airspace situation under all operating conditions. This provides a reliable and robust decision-making basis for mode switching in the dual-engine obstacle avoidance architecture.

[0100] Reference Figure 5 As one implementation of step S106, the step of generating flyable trajectory data by performing spatiotemporal grid search and collision detection processing based on probability location distribution data and threat index includes: Step S501: Construct a three-dimensional spatiotemporal grid that includes forward space, lateral space and time dimension; The forward and lateral spaces constitute the two-dimensional physical plane of the aircraft's motion, covering its forward detection range and lateral maneuver envelope. The introduction of the time dimension strongly binds spatial location with timestamps, so that each node in the grid represents not only a physical coordinate but also the specific time when the aircraft arrived at that coordinate.

[0101] In this embodiment, the system optimizes parameters based on airborne flight conditions, setting the range and resolution of the forward space, the range and resolution of the lateral space, and the step size and total prediction period of the time dimension. The system combines these three-dimensional variables to construct a three-dimensional search grid. Since the preceding interactive multi-model prediction module has already output the probability distribution of obstacle positions at future moments, the construction of the three-dimensional spatiotemporal grid enables the system to accurately look up the table to obtain the spatial state of obstacles at a specific moment when expanding path nodes, thereby providing a spatiotemporal alignment framework for subsequent dynamic collision detection.

[0102] For example, when an aircraft flies forward at a certain speed, the system divides the state of the next few seconds into fixed steps in the time dimension. If there is a moving obstacle ahead, the system considers not only whether the path is spatially unobstructed when planning the path, but also whether the obstacle will have moved to a certain point when the aircraft reaches that point. The construction of a three-dimensional spatiotemporal grid enables the system to actively avoid intersections with dynamic obstacles in the time domain, solving the problem that traditional static planning cannot avoid moving targets.

[0103] Step S502: Expand the nodes in the three-dimensional spatiotemporal grid to obtain the coordinates of the candidate nodes and the predicted distribution parameters at the corresponding time. Node expansion is the process by which the pathfinding algorithm moves from the current node to adjacent reachable nodes. During expansion, the system generates a series of candidate nodes to be evaluated. These candidate nodes not only contain forward and lateral spatial coordinates but also are bound to corresponding timestamps. The predicted distribution parameters are the probability distribution data of the future positions of obstacles output by the preceding interactive multi-model filtering module, including the position mean vector and covariance matrix.

[0104] In this embodiment, the system expands outwards into the neighborhood from the starting point within a three-dimensional spatiotemporal grid according to preset kinematic rules, generating new candidate node coordinates. For each generated candidate node, the system reads its timestamp and extracts the corresponding obstacle prediction distribution parameters from the prediction data sequence based on the timestamp. This process ensures strict synchronization between spatial location and temporal state, avoiding false alarms or missed alarms due to time misalignment, and enabling each candidate node to undergo security verification on an accurate time slice.

[0105] For example, the system can extend to a candidate node representing the aircraft's arrival 50 meters ahead in two seconds. The system will automatically retrieve the mean and covariance matrix of the obstacle's probability position two seconds later from the output of the previous prediction module. In this way, the candidate node no longer faces a static obstacle, but rather the actual possible distribution of that obstacle two seconds later, making subsequent collision determination entirely based on dynamic prediction.

[0106] Step S503: Calculate the squared Mahalanobis distance of the candidate nodes based on the predicted distribution parameters, and compare it with a preset confidence threshold to generate a collision flag. In situations where obstacle location prediction is uncertain, traditional Euclidean distance can only perform rigid point-to-point collision detection, while squared Mahalanobis distance can compare the probability envelope of candidate nodes with that of obstacles, achieving probabilistic collision detection for soft boundaries. Preset confidence thresholds are typically based on a chi-square distribution, corresponding to a specific confidence probability range. The collision flag is a binary determination result used to indicate whether the candidate node falls within the obstacle's danger probability envelope.

[0107] This application innovatively represents obstacles as a probability distribution N(μ,Σ), and uses Mahalanobis distance to measure the statistical deviation between candidate nodes and obstacles, which is defined as: ; In the above formula, p=(x,y) represents the candidate node position, μ=(μ_x,μ_y) represents the mean obstacle position predicted by the IMM multi-model filter, and Σ is a 2×2 covariance matrix. To facilitate FPGA hardware implementation, complex matrix operations are transformed into deterministic multiply-add sequences by predefined inverse matrix elements (inv_xx=σ_yy / det, inv_xy=-σ_xy / det, inv_yy=σ_xx / det, det=σ_xx·σ_yy-σ_xy^2). For a multivariate Gaussian distribution with 2 degrees of freedom, the Mahalanobis distance follows a chi-square distribution. When d_M<3.0, the node is located within the 3σ confidence ellipse (confidence probability 98.9%). Therefore, this application uses d_M^2<9.0 as the collision risk criterion, and a collision risk is determined when the candidate node is located within the 99% probability ellipse.

[0108] In this embodiment, the system uses a hardware parallel processing unit to read the position mean vector and covariance matrix from the predicted distribution parameters of the candidate node coordinates. The system calculates the squared Mahalanobis distance between the candidate node and the center of the obstacle probability distribution using an algebraic expansion, eliminating complex matrix inversion and floating-point division. The calculated squared Mahalanobis distance is then fed into a comparator and compared with a preset confidence threshold representing a high confidence probability. If the calculated value is less than the threshold, it indicates that the candidate node is highly likely to collide with the obstacle, and the system immediately generates a collision flag indicating a collision risk; otherwise, a safety flag is generated.

[0109] For example, when the prediction module has low uncertainty about the future position of an obstacle, its covariance matrix corresponds to a small probability ellipse. In this case, even if a candidate node is physically close to the mean center of the obstacle, its squared Mahalanobis distance may have exceeded a preset confidence threshold. The system determines this to be safe and allows the aircraft to pass, avoiding unnecessary detours. Conversely, if the prediction uncertainty is high and the candidate node is close to the center, the squared Mahalanobis distance will be less than the threshold, and the system immediately outputs a collision flag to remove the node.

[0110] Step S504: Add candidate nodes that are safe according to the collision flag to the search list, and incorporate the threat index as a cost penalty into the heuristic function to search for the optimal path and generate the original planned path. The search list is a set of nodes to be expanded, maintained by the pathfinding algorithm, used to store safe nodes that have passed collision detection. The heuristic function is a mathematical expression that guides the search direction and evaluates the overall cost of nodes. Incorporating a threat index as a cost penalty term into the heuristic function means that the cost of a node depends not only on the path length but also on the risk level of the airspace in which the path is located. Optimal path search involves iteratively updating the search list to find the sequence of nodes with the minimum overall cost from the starting point to the target point, thereby generating the original planned path.

[0111] In this embodiment, the system filters out candidate nodes with safe collision flags and adds them to the open search list. When calculating the comprehensive cost of these nodes, the system adds a threat index calculated by the previous module as an additional penalty on top of the base distance cost. This means that even if a node in a high-threat airspace is closer, its total cost will increase significantly due to the risk penalty. During the iteration process, the system prioritizes expanding on nodes with the lowest total cost, thereby enabling the final generated original planned path to proactively bypass high-risk areas, achieving the dual goals of optimal distance and lowest risk.

[0112] For example, during the search, the system might discover two paths to the target point: one shorter but passing through an area with a high threat index, and the other slightly longer but with an extremely low threat index. Because the heuristic function incorporates a threat index penalty, the shorter path actually has a higher total cost. The system will ultimately choose the slightly longer but safer path, one that is further away from high-risk airspace. This ensures that the originally planned path is not only optimized in terms of geometric distance but also optimal in terms of obstacle avoidance safety.

[0113] Step S505: The original planned path is smoothed by B-spline interpolation algorithm and curvature and acceleration constraints are applied to generate flyable trajectory data.

[0114] The original planned path is a polyline composed of a series of discrete spatiotemporal grid nodes, containing numerous hard inflection points, which does not conform to the dynamic characteristics of an aircraft. B-spline interpolation is a mathematical method that generates smooth, continuous curves using control points, eliminating sharp angles in the path. Curvature constraints limit the severity of trajectory curvature, while acceleration constraints limit abrupt changes in velocity; both are physical boundaries ensuring flight safety and passenger comfort. The flyable trajectory data is the final output to the flight control system, consisting of continuous, smooth trajectory commands that satisfy physical constraints.

[0115] In this embodiment, the system uniformly samples several control points from the original planned path and uses a B-spline interpolation algorithm to fit and generate a continuous, smooth curve. During the fitting and generation process, the system monitors the geometric characteristics of the curve in real time, strictly limiting its maximum curvature to not exceed the aircraft's maneuverability limits, while ensuring that the acceleration along the trajectory does not exceed the load tolerance range. After this smoothing and constraint processing, the originally discrete and rigid polygonal path is transformed into flyable trajectory data that conforms to the aircraft's kinematic characteristics, which can be directly used for closed-loop tracking in the flight control system.

[0116] For example, the original planned path might contain a near-right-angle turn at a certain obstacle avoidance point. If the aircraft follows this path, it will stall and disintegrate due to excessive lateral overload. The system uses a B-spline interpolation algorithm to transform this right-angle turn region into a smooth arc with continuously varying curvature, ensuring that the radius of curvature of this arc is greater than the aircraft's minimum turning radius. Ultimately, the generated flyable trajectory data eliminates all maneuvers exceeding physical limits, guaranteeing the aircraft's attitude stability and flight safety during autonomous obstacle avoidance.

[0117] In the above implementation, the system constructs a spatiotemporally three-dimensional linked path planning architecture, and uses Mahalanobis distance to achieve probabilistic collision detection that considers prediction uncertainties, avoiding false alarms and missed alarms of traditional rigid detection. Simultaneously, the system deeply integrates the threat index into the heuristic function, enabling path search to have predictive capabilities for proactively avoiding high-risk airspace. Finally, through joint processing of spline interpolation and physical constraints, the discrete algorithm path is transformed into a flyable trajectory that conforms to aerodynamic characteristics, achieving a high degree of unity between global optimality, safety, and engineering feasibility in path planning in dynamic airspace.

[0118] Reference Figure 6 As one implementation of step S503, the step of calculating the squared Mahalanobis distance of candidate nodes based on the predicted distribution parameters and comparing it with a preset confidence threshold to generate a collision flag includes: Step S601: Read the location mean vector and location covariance matrix from the predicted distribution parameters; The position mean vector represents the spatial coordinate center point where the obstacle is most likely to appear at a certain future moment, while the position covariance matrix describes the degree of uncertainty divergence of the predicted position in various dimensions such as the lateral and forward directions.

[0119] In this embodiment of the application, after the system expands the candidate nodes to be evaluated in the three-dimensional spatiotemporal grid, the processing system will accurately read the position mean vector and position covariance matrix at the corresponding time from the output sequence of the interactive multi-model prediction module according to the timestamp carried by the candidate nodes.

[0120] Step S602: Invert the position covariance matrix and extract the element parameters of the inverse matrix; In calculating Mahalanobis distance, the inverse matrix must be used to weight the coordinate deviations. The parameters of the inverse matrix are scalar values ​​extracted from the inverse of the position covariance matrix, representing the shape of the obstacle probability ellipse in different directions and the uncertainty weights.

[0121] Specifically, the system pre-inverts the position covariance matrix at the processing end and extracts the result as inverse matrix element parameters according to the requirements of algebraic expansion. These parameters are then broadcast to the configuration register of the hardware parallel processing unit as fixed weighting coefficients for subsequent multiply-accumulate operations, enabling the hardware pipeline to directly call these parameters for extremely fast calculations in a stateless manner.

[0122] For example, when the uncertainty of an obstacle is large in the lateral direction but small in the forward direction, the element parameters corresponding to the inverse of its position covariance matrix in the lateral direction will have a large penalty weight. The system extracts these element parameters in advance, so that subsequent calculations can directly use these scalar values ​​to weight the coordinate deviation, eliminating the need to repeatedly perform the time-consuming matrix inversion process at each node detection, thus greatly improving the detection efficiency of a single node.

[0123] Step S603: Calculate the squared Mahalanobis distance d according to the following formula. 2 M: ; Where Δx and Δy are the differences between the candidate node coordinates and the mean vector of their positions, inv xx 、inv xy 、inv yy The parameters are the elements of the inverse matrix; Specifically, the squared Mahalanobis distance is a measure of the statistical deviation between the coordinates of a candidate node and the center of the obstacle probability distribution. In this formula, the difference between the candidate node coordinates and the mean vector of the position represents the geometric distance deviation between the predicted position of the aircraft and the predicted center of the obstacle on each spatial axis, while the parameters of the inverse matrix serve as weighting coefficients for these deviations.

[0124] In this embodiment, to address the real-time collision detection requirements of massive nodes in spatiotemporal path planning, the system designs a dedicated hardware parallel processing unit on the FPGA PL side to implement the probabilistic collision detection algorithm based on Mahalanobis distance in hardware. Traditional Mahalanobis distance calculation involves complex matrix inversion and floating-point division, which is difficult to implement efficiently in hardware. This application utilizes a hardware parallel processing unit to calculate the square of the Mahalanobis distance. The hardware parallel processing unit is based on a multi-path PE parallel array architecture and implements the algebraic expansion calculation of the square of the Mahalanobis distance through pure combinational logic circuits. All PE operation units perform synchronous parallel operations.

[0125] Specifically, the array contains 64 parallel PE channels, with each PE arithmetic unit processing one spatiotemporal candidate node. To achieve extremely low latency, the internal circuitry of a single PE channel is implemented using pure combinational logic, abandoning the traditional clock-synchronous register pipeline design. The logic flow of the PE arithmetic unit circuit includes: a first-level subtractor for calculating coordinate differences, a second-level multiplier array for performing the product operation of squared terms and cross terms, a third-level multiplier for loading inverse matrix coefficients for weighting, and a fourth-level addition tree structure for performing the final accumulation and summation.

[0126] In some embodiments, the four-level combinational logic pipeline architecture in the PE computing unit specifically includes: The first stage is the coordinate deviation calculation circuit, which consists of two parallel 16-bit fixed-point subtractors. These subtractors receive the candidate node coordinates (x, y) and the pre-configured predicted ellipse center (μ_x, μ_y) from the ARM, respectively, and synchronously calculate the deviations Δx = x - μ_x and Δy = y - μ_y. Fixed-point representation is used to eliminate floating-point operation overhead. The subtractors are purely combinational logic, with an output delay of less than 2ns.

[0127] The second stage: the squaring and multiplication circuit, utilizes three dedicated DSP48E1 multiplier cores operating in parallel to calculate Δx², Δy², and Δx·Δy without blocking. The DSP hard core resources complete an 18×18-bit multiplication in a single cycle. This stage performs basic algebraic operations, transforming geometric relationships into primitives for subsequent weighted calculations.

[0128] The third stage: constant weighting circuit, which is the essence of the algorithm's hardware implementation. First, three parallel scalar multipliers multiply the intermediate terms Δx², Δy², and Δx·Δy by the elliptic parameters inv_xx, inv_yy, and 2·inv_xy read from the configuration register, respectively. These parameters are obtained by inverting the prediction covariance matrix Σ by the ARM, followed by algebraic expansion and pre-scaling, representing the shape, orientation, and uncertainty of the obstacle probability ellipse. Subsequently, a three-input adder tree accumulates the weighted results, outputting the final Mahalanobis distance squared value D². This process transforms complex matrix operations into a fixed multiply-add sequence.

[0129] Level 4: Accumulation and Comparison Circuit. The three-input adder tree accumulates the weighted three terms to obtain the squared Mahalanobis distance D². D² is then fed into a high-speed comparator, which compares it with the dynamic threshold T (3σ corresponds to 9.0) read from a register, outputting a 1-bit collision flag. A true flag indicates that the node falls within the obstacle probability ellipse.

[0130] This purely combinational logic parallel architecture enables the system to complete collision detection of 64 nodes in parallel within one clock cycle. Compared to the hundreds of microseconds required for serial execution of software algorithms on the ARM side, this hardware acceleration module compresses the latency of a single batch detection to within microseconds, fundamentally breaking through the computing power bottleneck of airborne platforms in dynamic and complex environments, and supporting the high-frequency and rapid iteration of the STA* algorithm.

[0131] Step S604: Compare the squared Mahalanobis distance with a preset confidence threshold. If the squared Mahalanobis distance is less than the preset confidence threshold, generate a collision flag indicating a collision risk.

[0132] The pre-set confidence threshold is a boundary criterion based on the statistical properties of a multivariate Gaussian distribution. In two-dimensional space, the squared Mahalanobis distance follows a chi-square distribution; setting a specific threshold corresponds to a confidence probability interval. Comparing the squared Mahalanobis distance with this threshold determines whether a candidate node falls within a high-probability obstacle region. The collision flag is the output of this decision logic, used to characterize whether the candidate node poses a collision risk, thus guiding subsequent path search algorithms to retain safe nodes or eliminate dangerous nodes.

[0133] Specifically, the system feeds the calculated squared Mahalanobis distance into a high-speed comparator and compares it with a preset confidence threshold read from a register. To cover the vast majority of potential collision scenarios, the system typically sets the threshold to a value corresponding to an extremely high confidence probability. If the comparison result shows that the squared Mahalanobis distance is less than the threshold, it indicates that the candidate node is within the probability envelope of the obstacle, and the system immediately generates a collision flag indicating a collision risk and removes it; otherwise, the node is deemed safe and allowed to be added to the search list.

[0134] For example, when the system sets the preset confidence threshold to a value corresponding to an extremely high confidence probability, if the squared Mahalanobis distance of a candidate node is less than this threshold, it means that there is a very high probability that the aircraft will overlap with the obstacle when it reaches that node. In this case, the system generates a collision flag indicating a collision risk, promptly preventing the node from expanding further. This probability-based soft-boundary judgment method can more scientifically accommodate prediction errors, neither overlooking any potential threats nor excessively compressing the safe flight space of the aircraft.

[0135] In the above implementation, the system reconstructs the complex probabilistic collision detection algorithm from matrix operations at the software level to a combinational logic multiplication-addition sequence at the hardware level, eliminating the computational bottleneck by extracting inverse matrix element parameters and algebraic expansion. By fusing candidate node deviations and probability distribution parameters into the squared Mahalanobis distance, and combining it with a pre-set confidence threshold for soft boundary determination, the system ensures extremely high detection accuracy while compressing the single-node collision detection latency to the nanosecond level. This breaks through the computational power limitations of real-time verification of massive nodes in dynamic spatial domains, providing solid underlying computational power support for the rapid iteration and convergence of spatiotemporal path planning.

[0136] Reference Figure 7 As one implementation of step S107, the step of generating an emergency obstacle avoidance command by performing artificial potential field calculation based on relative distance and radial velocity includes: Step S701: Calculate the dynamic repulsion coefficient based on the radial velocity; wherein the dynamic repulsion coefficient is positively correlated with the square of the radial velocity; Specifically, in emergency mode, the system directly reads the radial velocity from the point cloud data generated by the front-end signal processing module, squares it, and multiplies it by a preset scaling factor to calculate the dynamic repulsion coefficient in real time. This coefficient is no longer a fixed constant but a variable that updates in real time with the target's maneuvering speed. Through this mechanism, the system deeply integrates Doppler physical characteristics with the mechanical model at the algorithm level, ensuring that the intensity of the repulsion field can accurately match the actual collision urgency when facing targets with different approach speeds.

[0137] For example, when an aircraft encounters a target flying slowly in the same direction ahead, its radial velocity is extremely small, and the calculated dynamic repulsion coefficient is close to zero, so the system will not produce an excessive repulsion response. However, when an enemy aircraft flies towards it at an extremely high speed, the extremely large radial velocity causes the dynamic repulsion coefficient to increase exponentially, directly raising the overall repulsion barrier of the target, forcing the aircraft to produce a strong tendency to avoid it, effectively making up for the shortcomings of traditional range potential fields in responding slowly to high-speed threats.

[0138] Step S702: Calculate the repulsive force vector of a single obstacle according to the following formula. : ; Where, k rep The basic repulsion coefficient, k dop R is the dynamic repulsion coefficient, and R is the relative distance. v is a unit distance vector. radial Radial velocity, The radial velocity direction is a unit vector; Specifically, the single obstacle repulsion vector is the repulsive force exerted by a single obstacle on an aircraft in an artificial potential field, including both magnitude and direction. The basic repulsion coefficient ensures that the basic avoidance distance can be maintained even at low speeds or when stationary, while the superposition of dynamic repulsion coefficients allows the repulsive force to be dynamically amplified according to the Doppler characteristics.

[0139] In this embodiment, the system extracts the relative distance of a single obstacle in the current point cloud, and calculates the comprehensive repulsive force amplitude of the obstacle by combining the preset basic repulsive force coefficient and the dynamic repulsive force coefficient calculated in the previous step. Subsequently, the system calculates the distance unit vector based on the spatial orientation of the obstacle relative to the aircraft, multiplies the repulsive force amplitude by the unit vector, and finally generates a single obstacle repulsive force vector with a clear spatial orientation.

[0140] For example, when facing an object cutting in from the side, although the distance is relatively short, its radial velocity is not large, and its single obstacle repulsion vector is mainly determined by the basic repulsion coefficient and the relative distance, so it will not cause the aircraft to make violent maneuvers. However, when facing a high-speed target approaching from the front, the basic repulsion and the extremely large dynamic repulsion are superimposed, and the resulting single obstacle repulsion vector will be extremely large, pointing directly to the rear or side of the aircraft, forcibly driving the aircraft to quickly deviate from the current flight path.

[0141] Step S703: Determine whether the relative distance is less than the near-field danger threshold; if yes, proceed to step S704; otherwise, proceed to step S705. The near-field hazard threshold is a physical distance threshold set by the system to determine whether an aircraft has entered an extreme hazard zone that cannot be reversed by conventional force field calculations. Within this distance, the aircraft's aerodynamic inertia and control system delays mean that conventional repulsive force vectors may not be able to pull it away from the hazard zone in time.

[0142] In this embodiment, while generating a single obstacle repulsion vector, the system simultaneously extracts the relative distance to the obstacle and compares it with a preset near-field danger threshold. This comparison process has the highest priority in the computational chain. Once the condition is met, the system will no longer execute subsequent complex vector summation and global coordination logic, but will directly switch to the highest-level single-target emergency bypass to ensure that the most decisive action is taken within the millisecond-level time window of life and death.

[0143] For example, when an aircraft suddenly stops abruptly upon encountering a large bird, and the distance between them instantly shrinks to within the near-field danger threshold, conventional force field calculations may result in the repulsive force vectors canceling each other out due to the presence of other surrounding obstacles. In this situation, the system uses this judgment step to quickly identify this irreversible physical critical state, immediately interrupting conventional logic to prepare for directly outputting the maximum reaction force.

[0144] Step S704: Output the maximum emergency repulsion vector; Among them, the direction of the maximum emergency repulsion vector is the composite direction of the repulsion vectors of all near-field obstacles; when multiple targets enter the near-field at the same time, the maximum emergency repulsion vectors of each near-field target are summed and then output with amplitude limiting.

[0145] Step S705: Sum the repulsive force vectors of all obstacles to generate the total repulsive force; The maximum emergency repulsion vector is a system-preset limit repulsion force directed away from the current near-field obstacle, its function being to force maximum overload maneuvering at the extreme distance. The total repulsion force is the vector sum of the single obstacle repulsion vectors generated by all detected obstacles in the airspace under normal emergency conditions.

[0146] Understandably, through branching, the system achieves comprehensive multi-target avoidance in general emergency situations by using total repulsion force to find the safest resultant force direction; while in extreme situations, it achieves single-target avoidance by using the maximum emergency repulsion force vector, ensuring the robustness of emergency response under extreme conditions.

[0147] Specifically, if the previous step determines that the relative distance is less than the near-field danger threshold, the system immediately retrieves the preset maximum emergency repulsion vector from memory as the current output; if it is not less than the threshold, the system iterates through the single obstacle repulsion vectors of all obstacles in the current frame, superimposing them pairwise according to the vector addition rule, and finally generating a total repulsion force that reflects the overall airspace repulsion trend. Furthermore, when multiple obstacles simultaneously satisfy the condition that their relative distance is less than the near-field danger threshold, the obstacle with the highest threat index is selected as the direction reference for the maximum emergency repulsion vector.

[0148] For example, when an aircraft is surrounded by multiple drones but still has room to maneuver, the system calculates the total repulsive force and finds the best escape direction by combining the repulsive forces in all directions; and once a drone suddenly approaches to the very near field, the system directly outputs the maximum emergency repulsive force vector, driving the aircraft away from the deadly threat with maximum acceleration.

[0149] Step S706: Convert the maximum emergency repulsion vector or total repulsion into an acceleration command, and perform amplitude limiting processing within a preset amplitude range to generate an emergency obstacle avoidance command.

[0150] Acceleration commands are motion control parameters that the flight control system can directly recognize and execute. The preset amplitude range is the maximum acceleration boundary set based on the aircraft's aerodynamic layout, structural strength, and engine thrust limitations. Amplitude limiting is used to prevent the artificial potential field from calculating overload commands exceeding the aircraft's physical capabilities in extremely dense or high-speed approach scenarios, thus avoiding structural disintegration or stall and crash due to violent maneuvers. Emergency obstacle avoidance commands are the final control signals issued after physical constraint verification.

[0151] In this embodiment, the system maps the acquired maximum emergency repulsion vector or total repulsion force proportionally to a three-axis acceleration command based on the aircraft's current mass and flight attitude. During the mapping process, the system monitors each axis component of the acceleration command in real time. If any axis component is found to exceed a preset amplitude range, it is truncated to a boundary value using a limiting algorithm. The command after limiting is finally encapsulated into an emergency obstacle avoidance command and sent directly to the flight control system for execution via the bus.

[0152] For example, when multiple oncoming targets result in an extremely large total repulsive force, the system directly converts it into a corresponding acceleration command. However, this command may require the aircraft to generate extremely large lateral overloads. The limiting processing with a preset amplitude range restricts this to within the aircraft's maximum permissible lateral overload. The generated emergency obstacle avoidance command ensures that the aircraft tries its best to avoid obstacles without losing attitude control due to excessive overload, achieving an engineering balance between survivability and flight safety.

[0153] In the above implementation, the system constructs an artificial potential field emergency reflection architecture based on dynamic enhancement of Doppler radial velocity. By utilizing a dynamic repulsion coefficient positively correlated with the square of the radial velocity, it solves the problem of delayed response of traditional potential fields to high-speed approaching threats. Simultaneously, by setting an extreme near-field danger threshold and a maximum emergency repulsion bypass, it avoids the fatal risks caused by multi-target repulsion cancellation. Furthermore, with strict physical limiting processing, it ensures that the system generates emergency obstacle avoidance commands within milliseconds, enabling extreme avoidance without causing loss of control of the aircraft, thus constructing a solid safety defense for the aircraft.

[0154] In practical applications, the specific execution process of the above method can be illustrated by the following six stages: Phase 1: After system startup, the MIMO radar transmits FMCW continuous waves. Target echoes enter the FPGA PL terminal via four LVDS interfaces. The ADC synchronously completes the digitization acquisition of the four-channel intermediate frequency signals at a 10MHz sampling rate, preserving complete time-domain amplitude and phase information. The hardware pipeline then begins: First, a 256-point 1D-FFT is performed to transform the time-domain signal to the range domain, achieving a range resolution of 3.75cm using a 4GHz bandwidth. Next, a 128-point 2D-FFT is performed on the continuous echoes of the same range cell to extract Doppler frequencies and calculate radial velocity with an accuracy of 0.1m / s. Finally, a CA-CFAR detector (4 protection cells + 16 reference cells) filters valid targets in the range-Doppler spectrum, eliminating ground clutter and electromagnetic interference. The entire processing delay is less than 5ms, and the output, including range R, azimuth θ, radial velocity v_r, and signal-to-noise ratio SNR, is sent to shared memory for use by the ARM terminal.

[0155] Phase Two: After the ARM module reads the point cloud from shared memory, it first performs data association using the nearest neighbor algorithm, matching the new point cloud with existing historical trajectories. Point clouds that cannot be associated are used as the starting point of the new trajectory. For each dynamic obstacle, the IMM filter updates three Kalman filters in parallel: CV (uniform velocity), CA (uniform acceleration), and CT (constant angular velocity turning). Each filter predicts the state based on its own motion model, then fuses the new observations (position + Doppler velocity) to update the state estimate. Next, it calculates the Gaussian likelihood function based on the observation residuals to update the model probability. Finally, it performs weighted fusion to output the fused state and a 2×2 covariance matrix. The prediction module extrapolates the fused state to future times of 0.5s, 1.0s, 1.5s, and 2.0s, outputting the probability position distribution N(μ,Σ), completing the upgrade from "real-time detection" to "advanced prediction."

[0156] Phase 3: The ARM side calculates the four-dimensional threat index TI for each obstacle. First, it solves for the relative distance R and the relative velocity scalar v_rel. Then, it calculates the geometric collision zone coefficient A based on the relative heading angle θ_rel. collision (|θ) rel |<30° take 1.0, |θ rel |Take a value of 0.5 between 60° and 120°,|θ rel (>150° is taken as 0.2, the rest are linearly interpolated), and finally weighted and fused according to the formula including the radial velocity square term. The weighting coefficients are dynamically adjusted according to the flight phase: takeoff / climb emphasis distance, cruise enhanced radial velocity, landing emphasis area coefficient, and emergency maneuver maximizes radial velocity perception. The radial velocity square term makes the threat of oncoming high-speed targets increase quadratically. TI value enters mode determination: if TI≥0.8 triggers an emergency state, the system immediately skips the STA planning link and activates the APF emergency mode; if TI<0.8, it enters the normal mode and starts STA path planning.

[0157] Phase Four: In normal mode, the ARM side initiates STA* spatiotemporal search, constructing a three-dimensional spatiotemporal grid of X (0~200m, resolution 2m), Y (-50~50m, resolution 2m), and T (0~5s, resolution 0.1s). After initializing the starting point, the algorithm enters the main loop, retrieving the node with the minimum cost from the OPEN list and expanding its spatiotemporal neighborhood to generate candidate child nodes. For each candidate node, the ARM writes the node coordinates and timestamp along with the corresponding obstacle prediction distribution into the FPGA BRAM. The FPGA's 64-channel parallel collision detection unit completes the Mahalanobis distance calculation within 0.8μs and returns a collision flag. Safe nodes enter the OPEN list, and their cost function integrates the basic distance cost, cumulative threat cost, and heuristic term. The threat index TI serves as an additional cost penalty, enabling the path to actively avoid high-risk airspace. After the search is completed, cubic B-spline interpolation is used to smooth the original path. Five control points are uniformly sampled from the path to generate a continuous smooth trajectory, while strictly limiting the maximum curvature to ≤0.1m⁻¹ and the maximum acceleration to ≤5m / s².

[0158] Phase 5: When TI ≥ 0.8 triggers emergency mode, the system immediately interrupts STA* planning, and the APF module responds instantaneously based on the latest radar point cloud. The core algorithm is a Doppler-enhanced repulsive field, where the repulsive coefficient is proportional to the square of the radial velocity, causing the repulsive force against oncoming high-speed targets to increase non-linearly on a quadratic basis, while maintaining a basic level against targets in the same direction or far away. After summing the repulsive forces of all obstacles, the amplitude is limited to [-50, 50] m / s² to generate an emergency acceleration command. For extremely close-field high-risk scenarios, the algorithm directly outputs the maximum emergency command. The entire response chain, from TI exceeding the threshold to the output of flight control commands on the CAN bus, has a delay of less than 5ms. The system continues to run in APF mode; if TI < 0.8 and STA* has completed a full replanning, it automatically switches back to normal mode.

[0159] Phase Six: Both the flyable trajectory planned by STA* and the emergency acceleration commands generated by APF are transmitted to the flight control system via the CAN bus interface. The CAN bus uses differential signal transmission, which has high resistance to electromagnetic interference, ensuring that obstacle avoidance commands reach the flight control actuators stably and reliably in the complex electromagnetic environment of the airborne system, thus completing closed-loop control.

[0160] Reference Figure 8 As a further implementation of the dual-engine decision-making dynamic obstacle avoidance method for aircraft, after the step of outputting flyable trajectory data or emergency obstacle avoidance commands to the flight control system, the method further includes: Step S801: When outputting the emergency obstacle avoidance command, continuously obtain the updated threat index; Understandably, emergency obstacle avoidance commands are real-time control quantities generated by the emergency reaction arc engine when the system faces extremely high collision risks. Their core purpose is to quickly escape the current dangerous situation through extreme maneuvers. However, the state of obstacles and the relative position of the aircraft in dynamic airspace are constantly changing, and a single emergency command output cannot guarantee the complete elimination of the threat. Continuously acquiring updated threat indices during the output of emergency obstacle avoidance commands is to establish a closed-loop monitoring mechanism for the effectiveness of obstacle avoidance. This allows the system to monitor in real time whether the current maneuver has effectively increased the distance to the obstacle or reduced the relative approach speed, thus providing the most up-to-date state criteria for subsequent state transitions.

[0161] In this embodiment, when the system decision-making route determines that it has entered emergency mode and begins issuing emergency obstacle avoidance commands to the flight control system, the underlying radar signal processing and threat assessment links are not interrupted. The system continuously receives the latest frame of radar point cloud data, recalculates the relative distance and radial velocity, and, in conjunction with the adaptive weighting coefficients of the current flight phase of the aircraft, calculates and generates an updated threat index in real time. This continuous updating process runs through the entire emergency obstacle avoidance execution cycle, ensuring that the system's perception of the airspace situation does not have any time blind spots.

[0162] For example, when an aircraft executes an emergency obstacle avoidance command and makes a sharp turn to evade a high-speed approaching target, the relative distance and radial velocity between the two will change drastically as the aircraft maneuvers. By continuously acquiring updated threat indices, the system can keenly sense that the relative distance is gradually increasing and the radial velocity is decreasing, thus confirming the effectiveness of the current avoidance maneuver. Conversely, if the target changes course, causing the threat index to remain high, the system will maintain emergency mode and continue searching for an escape opportunity.

[0163] Step S802: When the updated threat index is detected to be lower than the preset risk threshold, and the spatiotemporal grid search and collision detection processing has completed one path replanning, switch to outputting flyable trajectory data.

[0164] The preset risk threshold is the critical criterion for distinguishing between normal safety situations and emergency high-risk situations. Spatiotemporal grid search and collision detection processing are used to find the optimal path that meets dynamic constraints in a three-dimensional spatiotemporal grid. Path replanning refers to the system's need to reconstruct the search space and generate a new path to the target point after an aircraft deviates from its original flight path due to an emergency maneuver.

[0165] In this embodiment, while maintaining the output of emergency obstacle avoidance commands, the background conventional planning engine is not completely shut down. Instead, it utilizes remaining computing power to initiate spatiotemporal grid search and collision detection processing based on the aircraft's real-time yaw position. When the system detects that the updated threat index has fallen below a preset risk threshold, indicating that the collision hazard has been temporarily eliminated, the system checks whether the conventional planning engine has completed at least one complete path replanning iteration and generated new flyable trajectory data. Only when both conditions are logically met simultaneously does the system trigger a smooth switching mechanism, switching the output link from emergency obstacle avoidance commands to the newly planned flyable trajectory data, ensuring a smooth transition of the aircraft from extreme obstacle avoidance state to normal cruise state.

[0166] For example, an aircraft successfully avoids an oncoming target through an emergency sideslip, and the updated threat index drops below the risk threshold. However, due to the emergency maneuver, the aircraft has deviated from its original flight path by tens of meters. If the emergency command is abruptly cut off at this point, the aircraft will be in an uncontrolled drifting state. Upon detecting that the threat has been cleared, the system waits for the background to complete a spatiotemporal grid search based on the current position and generate a smooth regression trajectory. Once this flyable trajectory data is ready, the system immediately switches its output, guiding the aircraft to smoothly correct its course along the new flyable trajectory data, effectively avoiding the sudden trajectory changes and attitude instability that may occur during mode switching.

[0167] In the above implementation, the system constructs a complete obstacle avoidance decision-making closed loop and a smooth state recovery mechanism. By continuously monitoring the threat index and executing path replanning in parallel during emergency obstacle avoidance, the system not only ensures the continuity and safety of the avoidance process, but also achieves a seamless handover between the emergency reaction arc and the normal global planning the moment the threat is cleared. This avoids abrupt changes in control commands during mode switching, ensuring that the aircraft can autonomously and smoothly recover to normal flight status after extreme avoidance, thus improving the robustness and engineering reliability of the dynamic obstacle avoidance system.

[0168] To verify the effectiveness of the proposed dual-engine dynamic obstacle avoidance method for aircraft, this application employs a hexacopter UAV equipped with a heterogeneous computing platform and millimeter-wave radar. Testing was conducted in mountainous canyons with typical terrain features such as sharp bends and steep climbs, with both static and dynamic obstacles set up. The system was compared with traditional serial processing schemes and schemes without emergency modules. Experimental results show that the system, utilizing FPGA acceleration, achieves a single-batch collision detection time of only 0.8 μs, with end-to-end response latency as low as 28 ms and 4.7 ms in normal and emergency scenarios, respectively. The obstacle avoidance success rate reaches 96%, and the average path curvature and threat index are significantly better than the comparison group. Furthermore, the resource utilization rates of FPGA logic units, DSP slices, and BRAM are all below 6%, meeting the airborne resource requirements.

[0169] Furthermore, in extreme stress tests against sudden rockfalls, the system triggered the APF emergency mode through threat assessment within an extremely short intersection window, responding rapidly and achieving zero collisions in 99.4% of the flight mileage. This verifies the significant advantages of the heterogeneous collaborative and dual-engine redundant architecture in terms of real-time performance and reliability of dynamic obstacle avoidance.

[0170] In summary, the technical solution of this application has the following beneficial effects: First, it possesses the real-time advantage of deep collaborative processing between FPGA and ARM heterogeneous systems. This application adopts an FPGA+ARM heterogeneous platform, utilizing the high-concurrency pipeline characteristics of the FPGA PL end to handle fixed logic tasks such as radar FFT transformation, CFAR detection, and 64-channel parallel collision detection, compressing the computation latency to the nanosecond level. At the same time, the ARM PS end is responsible for IMM multi-model filtering, STA* node expansion, and global scheduling. The two collaborate through the on-chip AXI-4 bus, eliminating data interaction bottlenecks and effectively supporting the low-latency operation of the entire system in high-speed dynamic scenarios.

[0171] Second, it possesses Doppler-enhanced predictive sensing capabilities. Unlike traditional tracking methods that rely solely on position observation, this application uses the high-precision radial velocity output by the radar as the direct observation input of the IMM filter to construct an analytical observation model. This enables the three Kalman filters (CV, CA, and CT) to iterate in parallel and respond to target motion changes in a single frame. This design achieves a qualitative leap from "real-time detection" to "advanced prediction," and can output the probabilistic position distribution at future moments, improving the ability to plan ahead and avoid high-speed maneuvering targets.

[0172] Third, it possesses full-condition adaptive threat assessment capabilities. This application constructs a four-dimensional threat index model that includes the reciprocal of distance, relative velocity scalar, geometric collision zone coefficient, and radial velocity squared term. The radial velocity squared term can accurately amplify the urgency of high-risk collision scenarios. At the same time, the system introduces a flight phase adaptive weighting mechanism, which can dynamically adjust the weight coefficients of each threat dimension according to different phases such as takeoff / climb, cruise, landing, and emergency maneuvers, achieving accurate situational quantification across all flight conditions without manual intervention.

[0173] Fourth, it possesses decision-making capabilities for spatiotemporal path planning and threat fusion optimization. This application constructs a three-dimensional spatiotemporal grid search space of X, Y, and T, strongly binding spatial location with the time dimension, endowing path planning with the ability to dynamically avoid obstacles in a temporal sequence; the heuristic function innovatively incorporates a four-dimensional threat index as an additional cost penalty into the calculation, enabling the planning engine to actively avoid high-threat airspace, achieving dual-objective optimization of "optimal distance" and "lowest risk", and improving the safety and rationality of the path.

[0174] V. It possesses advantages in full-stack acceleration and high resource utilization based on hardware circuit mapping. This application fully implements the probabilistic collision detection algorithm based on Mahalanobis distance in hardware. Through algebraic expansion, complex matrix operations are transformed into a pure combinational logic multiply-accumulate sequence, and a 64-way PE parallel array architecture is adopted. This design compresses the time for a single batch detection to 0.8 microseconds, improving performance by three orders of magnitude compared to pure software implementation, while the logic resource utilization rate is less than 6%. While breaking through the bottleneck of onboard computing power, it also has excellent energy efficiency and deployability.

[0175] VI. High reliability advantage of dual-engine redundant security architecture. This application designs a STA* and APF dual-engine layered redundancy architecture. In normal scenarios, the STA* main engine plans the globally optimal path. When the threat index reaches the threshold, it instantly switches to the APF emergency reaction arc module. The APF algorithm uses the Doppler enhanced repulsive field to directly output obstacle avoidance commands with a response latency of less than 5 milliseconds. Moreover, the dual engines have a smooth recovery mechanism, forming a double insurance of "normal planning + emergency backup", which greatly improves the robustness and survivability of the system in extreme scenarios.

[0176] This application also discloses a dual-engine decision-making dynamic obstacle avoidance system for aircraft.

[0177] A dual-engine decision-based dynamic obstacle avoidance system for aircraft, specifically comprising: The radar signal processing module is used to acquire radar echo signals, perform spectrum transformation and constant false alarm rate (CFAR) detection on the radar echo signals, and generate point cloud data. The point cloud data includes the relative distance, azimuth angle, and radial velocity between the aircraft and the obstacle. The trajectory prediction module is used to perform multi-model interactive filtering and state extrapolation on point cloud data based on radial velocity to generate probability location distribution data of obstacles at future times. The relative velocity calculation module is used to obtain the velocity vector of the aircraft and, based on the obstacle velocity vector in the same coordinate system output by multi-model interactive filtering, calculate the relative velocity of the obstacle relative to the aircraft through vector difference. The threat assessment module is used to calculate the threat index of generated obstacles based on relative distance, relative speed, and radial speed. The risk assessment module is used to determine whether the threat index has reached a preset risk threshold. The conventional obstacle avoidance planning module is used to generate flyable trajectory data by performing spatiotemporal grid search and collision detection based on probability location distribution data and threat index when the threat index does not reach the preset risk threshold. The emergency obstacle avoidance calculation module is used to generate emergency obstacle avoidance commands when the threat index reaches a preset risk threshold, by performing artificial potential field calculations based on relative distance and radial velocity through the emergency reflection engine. The command output module is used to output flyable trajectory data or emergency obstacle avoidance commands to the flight control system.

[0178] The dual-engine decision-making dynamic obstacle avoidance system for aircraft according to the embodiments of this application can implement any of the above methods, and the specific working process of each module in the system can refer to the corresponding process in the above method embodiments.

[0179] In actual deployment, this system consists of a heterogeneous computing platform, six functional modules, and three data interfaces. The heterogeneous computing platform serves as the core hardware carrier of the system, integrating FPGA programmable logic (PL) and ARM processing system (PS), achieving hardware-software collaboration through on-chip high-speed interconnect. The FPGA PL handles fixed logic, high parallelism, and low latency tasks, including radar 1D / 2D-FFT spectrum transformation, CA-CFAR target detection, and a 64-channel parallel collision detection pipeline, achieving nanosecond-level latency through a hardware pipeline architecture. The ARM PS handles variable logic, complex scheduling, and serial computing tasks, including IMM multi-model filter updates, STA* node expansion control, path smoothing, and system-wide task scheduling, leveraging the flexibility of the instruction set.

[0180] The six specific functional modules are: radar signal processing module (responsible for raw echo signal processing and point cloud generation), IMM interactive multi-model trajectory prediction module (responsible for obstacle trajectory tracking and future state extrapolation), four-dimensional threat assessment module (responsible for airspace threat index calculation and classification), STA* path planning module (responsible for generating the globally optimal path in normal scenarios), FPGA parallel collision detection acceleration module (responsible for probabilistic collision detection in the STA* algorithm), and APF emergency reflection arc module (responsible for generating emergency obstacle avoidance commands in emergency scenarios). Three data interfaces are provided: radar input interface (4-channel LVDS), flight control output interface (CAN bus), and IMU attitude compensation interface (SPI), ensuring highly reliable interconnection between the system and external devices. The on-chip AXI-4 bus provides a 1.2GB / s transmission bandwidth, eliminating data interaction bottlenecks between the PS and PL ends. Task allocation is based on hardware and software characteristics: the PL end is adapted for repetitive, high-concurrency computation, while the PS end is adapted for dynamic algorithm iteration and multi-task scheduling, achieving optimal computational efficiency.

[0181] It should be understood that the above six functional modules are a physical division based on the actual hardware architecture and algorithm processing flow. From the perspective of logical function implementation, the system also includes a relative velocity calculation module, a risk judgment module, and a command output module. The relative velocity calculation module is configured to perform differential calculations based on the velocity vectors of the aircraft and obstacles; this function can be integrated into the radar signal processing module or the threat assessment module. The risk judgment module is configured to determine whether the threat index reaches a preset risk threshold; this logical control function is executed by the trajectory prediction module in collaboration with the threat assessment module or by the main control unit. The command output module is configured to send the trajectory or command to the flight control system as the system's output interface. Furthermore, during the conventional obstacle avoidance planning process, the FPGA parallel collision detection acceleration module operates as a hardware acceleration sub-module of the conventional obstacle avoidance planning module (i.e., the STA* path planning module), specifically undertaking high-frequency collision detection tasks and being called by the conventional obstacle avoidance planning module to support its spatiotemporal grid search process. During the emergency obstacle avoidance calculation process, the APF emergency reflection arc module, as the core execution unit of the emergency obstacle avoidance calculation module, directly generates emergency obstacle avoidance commands, bypassing the conventional planning link to achieve millisecond-level emergency response.

[0182] As one implementation of the radar signal processing module, the radar signal processing module is the system's sensing front end, responsible for receiving, analyzing, and extracting target information from the raw radar echo signal. It converts the time-domain echo into four-dimensional point cloud data containing range, azimuth, radial velocity, and signal-to-noise ratio. Module parameters include: ADC data acquisition (10MHz sampling rate, 4 signal channels), 1D-FFT range dimension transformation (256 points), 2D-FFT velocity dimension transformation (128 points), and CA-CFAR target detection (4 protection cells + 16 reference cells). The workflow is as follows: After system startup, the ADC synchronously digitizes the analog signals of the 4 radar intermediate frequency channels at a fixed sampling rate of 10MHz, preserving complete time-domain amplitude and time information; then, a 256-point 1D-FFT is performed to map the time-domain signal to the range domain, achieving an ultra-high range resolution of 3.75cm based on the 4GHz signal bandwidth; next, a 128-point 2D-FFT is performed on the same range cells to extract Doppler frequencies and calculate radial velocity; finally, the CA-CFAR algorithm is used to filter valid targets, remove clutter and interference, and output a four-dimensional point cloud. The module design adopts FPGA PL-end hardware pipeline parallel processing to replace the traditional DSP serial architecture, and the radial velocity detection accuracy reaches 0.1m / s, which meets the observation input requirements of the back-end IMM filter.

[0183] As one implementation of the IMM (Interactive Multi-Model) trajectory prediction module, this module is the core prediction unit of the radar perception system. It receives four-dimensional point cloud data of obstacles from the radar signal processing module and, based on historical and real-time observation information, continuously tracks the target trajectory and extrapolates its future state. This provides reliable time-series prediction data for obstacle avoidance, path planning, and probabilistic collision detection, achieving an upgrade from "real-time detection" to "advanced prediction." The core of this module adopts an interactive multi-model (IMM) filtering framework, specifically designed for airborne perception scenarios, solving the problem that traditional single Kalman filters cannot adapt to the maneuvering, speed-changing, and directional-changing flight of aerial targets. The algorithm input is radar point cloud data, and the output is the probabilistic position distribution at future times.

[0184] To adapt to aviation scenarios, three types of aviation-specific motion models are built: CV (uniform speed level flight), CA (uniform acceleration maneuver), and CT (constant angular velocity turn). The CV model is suitable for uniform speed level flight scenarios of UAVs and civil aircraft; the CA model is suitable for rapid climbs, dives, and other variable speed conditions; and the CT model is suitable for turning, circling, and other directional maneuvers. The three models are computed independently and iterated in parallel to accurately match the motion characteristics of various aircraft. The algorithm is deeply adapted to the detection advantages of airborne radar, using the high-precision Doppler radial velocity output from the front-end as the core observation input. Unlike traditional airborne algorithms that rely solely on position information filtering, this algorithm can quickly identify sudden changes in target flight state through single-frame velocity residuals. The second step of the algorithm performs model interaction and state pre-estimation. Relying on the aviation scenario-optimized transition probability matrix (such as the fixed model transition probability matrix [[0.8,0.1,0.1],[0.1,0.8,0.1],[0.1,0.2,0.7]]), it achieves smooth switching between multiple flight models, avoiding trajectory jumps during sudden changes in maneuver state. Each model combines the noise characteristics of the airborne system and the noise covariance of radar observations to iteratively update the posterior state estimate and error covariance matrix, outputting the optimal flight state solution for a single model. Finally, based on the observation residuals and Gaussian likelihood functions of each model, the posterior probability weights of the CV, CA, and CT models are adjusted in real time. The states and covariances of the three types of models are fused using dynamic probabilities as weighting coefficients to output the optimal flight state and uncertainty matrix of the airspace target. This accurately adapts to complex operating conditions such as steady-state flight, instantaneous maneuvers, and attitude switching of aircraft, improving the robustness and prediction accuracy of airborne multi-target tracking.

[0185] As one implementation of the four-dimensional threat assessment module, this module serves as the situational awareness hub of the airborne radar perception system. It receives obstacle position, velocity, and trajectory status information from the preceding trajectory prediction module, and performs a quantitative assessment and classification of airspace obstacle risk. This addresses the challenge of assessing the varying risk emphases at different stages of airborne flight, providing quantitative and reliable risk data for obstacle avoidance, maneuvering decisions, and situational analysis. The module's workflow is as follows: First, it receives real-time obstacle trajectory information and calculates basic risk parameters (relative distance and relative velocity scalars). Then, it combines the relative geometric attitudes of the aircraft and obstacles to calculate the geometric collision zone coefficient. Subsequently, it integrates the four-dimensional parameters (distance, relative velocity, airspace collision zone, and radial velocity) through the core threat formula, dynamically adjusts the weighting coefficients according to the current flight stage, and finally outputs low, medium, and high risk levels according to threshold rules.

[0186] As one implementation of the STA* path planning module, the STA* path planning module is the main decision engine of the airborne perception and planning system. It takes the obstacle threat index and temporal trajectory prediction results output by the preceding four-dimensional threat assessment module and autonomously generates the globally optimal obstacle avoidance path under normal flight scenarios. Compared with traditional two-dimensional spatial path planning algorithms, this module adopts the spatiotemporal A* (STA*) algorithm to construct a three-dimensional search system of space and time. It can adapt to the dynamic motion characteristics of airspace obstacles and solve the problems that traditional planning algorithms cannot handle dynamic obstacles, strong path rigidity, and poor threat situation fit. Finally, it outputs a smooth, safe, and flyable trajectory that conforms to airborne motion constraints, providing core decision support for the aircraft's autonomous obstacle avoidance and cruise maneuvering.

[0187] The module operation process adopts a full-link pipeline architecture: First, a three-dimensional spatiotemporal grid is constructed, consisting of X (0~200m, which is larger than the instantaneous physical detection range of radar, aiming to provide sufficient prediction and maneuver space for the aircraft), Y (-50~50m), and T (0~5s), strongly binding spatial location with the time dimension. Each node contains three-dimensional spatiotemporal information. Then, starting from the starting point, neighboring nodes are expanded. For each candidate node, the FPGA parallel collision detection unit is called to complete the probabilistic collision determination based on Mahalanobis distance within 0.8μs. Safe nodes that pass the detection are added to the OPEN list. The cost calculation integrates the basic distance cost and the four-dimensional threat index TI, and additional penalties are imposed on high-threat airspace. The algorithm iteratively updates the OPEN / CLOSE list until the optimal path is found. Finally, after three B-spline smoothing processes and satisfying curvature and acceleration constraints, the flyable trajectory is output to the flight control system.

[0188] Furthermore, the module improves upon the traditional A* algorithm by incorporating a scenario-specific heuristic function that integrates the threat index. This breaks away from the limitations of traditional algorithms that rely solely on geometric distance, achieving a dual-objective planning approach of "optimal distance + lowest risk." The heuristic function comprises three parts: basic distance cost, cumulative threat cost, and distance heuristic term. While ensuring algorithm adoptability, it integrates the TI threat index output from preceding modules into the cost calculation, imposing additional cost penalties on high-threat areas to proactively guide paths away from high-risk airspace. The algorithm maintains OPEN and CLOSE lists to iteratively update nodes and select the globally optimal path. To meet the high real-time requirements of airborne systems, the collision detection stage utilizes an FPGA acceleration unit, leveraging hardware pipelined parallel computation to significantly reduce the time consumed by security verification of massive numbers of nodes.

[0189] To address the issues of numerous inflection points, poor maneuverability, and unsuitability for direct airborne flight control in the original STA* path, the module employs a cubic B-spline interpolation algorithm for path smoothing. The algorithm uniformly samples five control points from the original planned path and generates a continuous, smooth trajectory based on interpolation fitting. Simultaneously, it strictly matches the aircraft's physical motion constraints, limiting the maximum trajectory curvature to ≤0.1m⁻¹ and the maximum acceleration to ≤5m / s², avoiding hard inflection points, sudden accelerations, and other trajectory problems that do not conform to flight characteristics, ensuring that the output trajectory can be directly adapted to airborne attitude control and maneuver execution.

[0190] As one implementation of the FPGA parallel collision detection acceleration module, this module is the core computing engine for STA* path planning. It implements a probabilistic collision detection algorithm based on Mahalanobis distance in hardware, breaking through the limitation of traditional deterministic collision detection that treats obstacles as rigid points. By introducing the probability distribution N(μ,Σ) output by the IMM prediction module, it uses Mahalanobis distance instead of Euclidean distance to measure the statistical deviation between candidate nodes and obstacles, and uses a 3σ confidence ellipse (d_M²<9.0) as the collision determination criterion. The module adopts a 64-channel PE parallel array architecture, with each PE unit realizing single-cycle computation through hardware expansion. The overall design uses pure combinational logic, and the time for a single batch detection is less than 0.8μs, providing underlying support for high-speed iteration of STA*.

[0191] Reference Figure 9The FPGA parallel collision detection acceleration module adopts a hardware acceleration architecture with a 64-channel PE parallel array and multi-stage reduction. External input simultaneously feeds 64 sets of node coordinates (x, y) to be detected along with the corresponding obstacle prediction distribution parameters (μ_x, μ_y, inv_xx, inv_xy, inv_yy) into 64 parallel computing units PE0 to PE63. After each PE independently completes the probabilistic collision detection calculation for a single node, all results are output to a 64-to-1 OR reducer, and finally integrated to output the global collision_flag flag. The core advantage of this parallel architecture is multi-node synchronous computation without waiting, enabling batch collision detection of 64 spatiotemporal nodes in a single operation. This overcomes the computational bottleneck of serial computation at the architectural level, perfectly adapting to the business requirements of massive node iterative detection using the STA* algorithm.

[0192] The single-channel PE (Probabilistic Collision Detection) unit is designed to fully map the probabilistic collision detection algorithm based on Mahalanobis distance into FPGA hardware circuitry. It is the smallest execution unit for hardware implementation of the Mahalanobis distance algorithm. Each PE unit implements the expanded calculation of the squared Mahalanobis distance through pure combinational logic circuitry. All PE units operate synchronously and in parallel, with a single batch detection time of less than 1μs. The PE unit adopts a four-stage combinational logic pipeline architecture, fully mapping the core formula of probabilistic collision detection and the calculation of the squared Mahalanobis distance into a four-stage pure combinational fixed-point arithmetic circuit consisting of a subtractor, a square multiplier, a constant multiplier, an accumulator, and a comparator. This forms a deterministic data path from coordinate input to collision flag output, requiring only approximately 8ns of combinational delay from input stabilization to output.

[0193] As one implementation of the APF (Artificial Potential Field) emergency response arc module, this module serves as a redundant safety guarantee for conventional STA* path planning, specifically designed to handle extreme emergency scenarios such as sudden high-risk airspace conflicts and instantaneous obstacle maneuvers. While conventional STA* planning algorithms can output the optimal flyable path, their iterative process has a fixed latency, making them unsuitable for millisecond-level emergency avoidance requirements. This module, based on the Artificial Potential Field (APF) architecture, establishes a dedicated airborne emergency response mechanism. Through real-time threat monitoring, Doppler-enhanced repulsive field calculation, and a simplified link directly connecting to flight control command output, it achieves microsecond-level emergency response, mitigating collision risks caused by planning delays in extreme scenarios. It acts as the last line of defense for overall aircraft flight safety.

[0194] Furthermore, to adapt to the full-scenario flight safety control requirements of airborne systems, the system is designed with a three-level operating state and an automatic recovery mechanism. Different states are precisely matched with corresponding operating engines and response latency, balancing the optimal trajectory for normal flight and the real-time response for emergency scenarios. The specific state switching rules are as follows: Normal cruise state (TI<0.6) is dominated by STA* (FPGA acceleration), with a response latency of approximately 28ms, ensuring flight stability; Attention state (0.6≤TI<0.8) is dominated by STA* (full-speed mode), with a response latency of approximately 20ms, improving the planning iteration speed to avoid risks in advance; Emergency state (TI≥0.8) is dominated by the APF emergency module, with a response latency of less than 5ms, bypassing the main planning link to directly output emergency obstacle avoidance commands. In addition, a recovery state is set (TI<0.8 and STA* replanning is completed), at which time it automatically switches back to STA* mode, smoothly returning to the normal planning mode after the risk is eliminated, ensuring a smooth attitude transition without abrupt changes.

[0195] In this embodiment, although the system employs a three-level threat classification (low, medium, high) or a four-level system state (normal, alert, emergency, recovery) to achieve refined flight control and adaptive parameter adjustment, from a decision-making logic perspective, the system's core engine switching mechanism strictly adheres to a binary judgment criterion. Specifically, the system presets a "preset risk threshold" (e.g., 0.8) as an emergency trigger benchmark: when the threat index reaches this threshold, the system determines that it has entered an emergency state and activates the APF emergency reaction arc module; when the threat index does not reach this threshold, regardless of whether it is in a low-risk or medium-risk range, the system determines that it is in a normal state and runs the STA* path planning module. The alert state or normal cruise state mentioned above are merely different operating modes under the normal state, the difference being the adjustment of internal planning parameters (such as iteration speed), and do not involve a substantial switching of the obstacle avoidance engine. Therefore, the above multi-level / multi-state description is a refinement and expansion of the "reached / not reached" judgment result in step S105 at the specific execution level, and the two are logically unified.

[0196] In the several embodiments provided in this application, it should be understood that the provided methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the division of a certain module is merely a logical functional division, and in actual implementation there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0197] This application also discloses a computer-readable storage medium.

[0198] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above in any of the dual-engine decision-making dynamic obstacle avoidance methods for aircraft.

[0199] The computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device; the program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0200] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A dynamic obstacle avoidance method for dual-engine aircraft, characterized in that, The method includes: The radar echo signal is acquired, and the radar echo signal is subjected to spectrum transformation and constant false alarm rate (CFAR) detection processing to generate point cloud data; wherein, the point cloud data includes the relative distance, azimuth angle and radial velocity between the aircraft and the obstacle; Based on the radial velocity, the point cloud data is subjected to multi-model interactive filtering and state extrapolation to generate probability position distribution data of obstacles at future times. The velocity vector of the aircraft is obtained, and the relative velocity of the obstacle with respect to the aircraft is calculated by vector difference based on the velocity vector of the obstacle in the same coordinate system output by the multi-model interactive filtering. Based on the relative distance, relative speed, and radial speed, the threat index of the obstacle is calculated and generated. Determine whether the threat index has reached a preset risk threshold; If the threat index does not reach the preset risk threshold, then the conventional planning engine performs spatiotemporal grid search and collision detection processing based on the probability location distribution data and the threat index to generate flyable trajectory data. If the threat index reaches a preset risk threshold, an emergency obstacle avoidance command is generated by performing artificial potential field calculation based on the relative distance and radial velocity through the emergency reflection engine. The flyable trajectory data or the emergency obstacle avoidance command is output to the flight control system.

2. The dual-engine decision-making dynamic obstacle avoidance method for aircraft according to claim 1, characterized in that, The steps of acquiring radar echo signals, performing spectrum transformation and constant false alarm rate (CFAR) detection on the radar echo signals, and generating point cloud data include: The radar echo signal is subjected to analog-to-digital conversion to generate a digital echo signal; The digital echo signal is subjected to a range-dimensional fast Fourier transform and a velocity-dimensional fast Fourier transform in sequence to extract and generate target range parameters and Doppler frequency parameters. The radial velocity is generated analytically based on the Doppler frequency parameters; The target distance parameters are processed by a constant false alarm rate (CFAR) detection algorithm to remove clutter interference and generate point cloud data including the relative distance, azimuth angle, radial velocity, and signal-to-noise ratio between the aircraft and the obstacle.

3. The dual-engine decision-making dynamic obstacle avoidance method for aircraft according to claim 2, characterized in that, Based on the radial velocity, the steps of performing multi-model interactive filtering and state extrapolation on the point cloud data to generate probability location distribution data of obstacles at future times include: The point cloud data is processed by the nearest neighbor algorithm to perform data association. Three Kalman filter models were launched in parallel: a uniform speed model, a uniform acceleration model, and a constant angular velocity steering model. The radial velocity in the point cloud data is used as a direct observation input to construct observation equations for radial velocity and Cartesian velocity to update the state of each Kalman filter model. The posterior probability weights of the three Kalman filter models are updated based on the Gaussian likelihood function and then weighted and fused to generate fused state data. The fused state data is extrapolated to a future preset time period to generate the probability position distribution data, which includes the position mean vector and the position covariance matrix.

4. The dual-engine decision-making dynamic obstacle avoidance method for aircraft according to claim 1, characterized in that, The steps for calculating the threat index of the obstacle based on the relative distance, relative velocity, and radial velocity include: Based on the relative heading angle between the aircraft and the obstacle, the geometric collision zone coefficient is generated by segmented fixed values ​​and linear interpolation. Read the adaptive weighting coefficients corresponding to the current flight phase of the aircraft; The threat index is calculated using the following formula: ； Where TI is the threat index, R is the relative distance, and R max v represents the radar's maximum effective detection range. rel V is the relative velocity. max A is the maximum relative velocity threshold preset for the system. collision v is the geometric collision region coefficient. radial denoted as radial velocity, and α, β, γ, and δ as adaptive weighting coefficients.

5. The dual-engine decision-making dynamic obstacle avoidance method for aircraft according to claim 4, characterized in that, The steps for generating flyable trajectory data by performing spatiotemporal grid search and collision detection based on the probability location distribution data and the threat index include: Construct a three-dimensional spatiotemporal grid that includes forward space, lateral space, and time dimensions; In the three-dimensional spatiotemporal grid, node expansion is performed to obtain the coordinates of candidate nodes and the predicted distribution parameters at the corresponding time. The squared Mahalanobis distance of the candidate node is calculated based on the predicted distribution parameters and compared with a preset confidence threshold to generate a collision flag. Candidate nodes that are safe by collision flags are added to the search list, and the threat index is incorporated as a cost penalty into the heuristic function to search for the optimal path and generate the original planned path. The original planned path is smoothed using a B-spline interpolation algorithm, and curvature and acceleration constraints are applied to generate the flyable trajectory data.

6. The dual-engine decision-making dynamic obstacle avoidance method for aircraft according to claim 5, characterized in that, The steps of calculating the squared Mahalanobis distance of the candidate nodes based on the predicted distribution parameters and comparing it with a preset confidence threshold to generate a collision flag include: Read the location mean vector and location covariance matrix from the predicted distribution parameters; Invert the position covariance matrix and extract the element parameters of the inverse matrix; The squared Mahalanobis distance d is calculated using the following formula. 2 M: ； Where Δx and Δy are the differences between the candidate node coordinates and the mean vector of their positions, inv xx 、inv xy 、inv yy The parameters of the inverse matrix; The squared Mahalanobis distance is compared with a preset confidence threshold. If the squared Mahalanobis distance is less than the preset confidence threshold, a collision flag indicating a collision risk is generated.

7. The dual-engine decision-making dynamic obstacle avoidance method for aircraft according to claim 1, characterized in that, The steps for generating emergency obstacle avoidance commands by performing artificial potential field calculations based on the relative distance and radial velocity include: The dynamic repulsion coefficient is calculated based on the radial velocity; wherein the dynamic repulsion coefficient is positively correlated with the square of the radial velocity. Calculate the repulsion vector of a single obstacle using the following formula. : ； Where, k rep The basic repulsion coefficient, k dop R is the dynamic repulsion coefficient, and R is the relative distance. v is a unit distance vector. radial Radial velocity, The radial velocity direction is a unit vector; Determine whether the relative distance is less than the near-field danger threshold; If yes, output the maximum emergency repulsive force vector; otherwise, sum the repulsive force vectors of all obstacles to generate the total repulsive force. The maximum emergency repulsive force vector or the total repulsive force is converted into an acceleration command, and the amplitude is limited within a preset range to generate the emergency obstacle avoidance command.

8. A dynamic obstacle avoidance method for a dual-engine aircraft according to any one of claims 1 to 7, characterized in that, After the step of outputting the flyable trajectory data or the emergency obstacle avoidance command to the flight control system, the method further includes: While outputting the emergency obstacle avoidance command, the updated threat index is continuously acquired; When the updated threat index is detected to be lower than the preset risk threshold, and the spatiotemporal grid search and collision detection process completes one path replanning, the system switches to outputting the flyable trajectory data.

9. A dual-engine decision-making dynamic obstacle avoidance system for aircraft, characterized in that, The system is used to execute the dual-engine decision-making dynamic obstacle avoidance method for an aircraft as described in any one of claims 1 to 8, the system comprising: The radar signal processing module is used to acquire radar echo signals, perform spectrum transformation and constant false alarm rate (CFAR) detection on the radar echo signals, and generate point cloud data; wherein, the point cloud data includes the relative distance, azimuth angle and radial velocity between the aircraft and the obstacle; The trajectory prediction module is used to perform multi-model interactive filtering and state extrapolation processing on the point cloud data based on the radial velocity to generate probability position distribution data of obstacles at future times. The relative velocity calculation module is used to obtain the velocity vector of the aircraft and, based on the obstacle velocity vector in the same coordinate system output by the multi-model interactive filtering, calculate the relative velocity of the obstacle relative to the aircraft through vector difference. The threat assessment module is used to calculate and generate a threat index for the obstacle based on the relative distance, relative speed, and radial speed. The risk assessment module is used to determine whether the threat index has reached a preset risk threshold; The conventional obstacle avoidance planning module is used to generate flyable trajectory data by performing spatiotemporal grid search and collision detection processing based on the probability location distribution data and the threat index through the conventional planning engine when the threat index does not reach the preset risk threshold. The emergency obstacle avoidance calculation module is used to generate an emergency obstacle avoidance command by performing artificial potential field calculation based on the relative distance and radial velocity through the emergency reflection engine when the threat index reaches a preset risk threshold. The command output module is used to output the flyable trajectory data or the emergency obstacle avoidance command to the flight control system.

10. A computer-readable storage medium, characterized in that: The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 8.

Images