An amphibious platform dynamic obstacle avoidance method fusing scene cognition

Abstract

The application relates to a kind of amphibious platform dynamic obstacle avoidance methods of fusion scene cognition, belong to the technical field of obstacle avoidance path planning, solve the problem that the obstacle avoidance method in prior art relies on artificial experience, poor safety, cannot cope with complex scene, and poor real-time of obstacle avoidance judgment action. Based on the clustering result of obstacle information to determine the active cluster, based on the obstacle information in the active cluster and the obstacle avoidance scene type corresponding to the active cluster to determine the obstacle avoidance mode, based on the obstacle avoidance mode to determine the obstacle avoidance duration and lateral offset value, based on the obstacle avoidance duration, lateral offset value and trajectory generation model to obtain multiple candidate trajectories, based on the preset kinematic constraint to filter the candidate trajectories to obtain the candidate set, calculate the cost function value of each candidate trajectory in the candidate set, and the trajectory with the minimum cost function value is taken as the optimal obstacle avoidance trajectory. A kind of obstacle avoidance method with high safety and good real-time is realized.

Description

Technical Field

[0001] This invention relates to the field of obstacle avoidance path planning technology, and in particular to a dynamic obstacle avoidance method for amphibious platforms that integrates scene cognition. Background Technology

[0002] When amphibious platforms navigate on water, they face not only the risk of collision with static obstacles but also the risk of collision with dynamic obstacles. A safe and real-time obstacle avoidance planning method is particularly important for amphibious platforms to achieve safe autonomous navigation. Existing obstacle avoidance planning methods mainly use fixed obstacle avoidance timing, and the weight of the cost function for evaluating the obstacle avoidance trajectory is mainly determined by expert experience. Although existing methods can meet the basic trajectory tracking and static obstacle avoidance functions of amphibious platforms in some simple water scenarios, they still have the following shortcomings: (1) Insufficient ability to handle multi-obstacle scenarios, and the collision risk assessment index for dynamic obstacles is singular, often only considering the nearest encounter distance (DCPA) and the nearest encounter time (TCPA); (2) In terms of obstacle avoidance decision-making, fixed obstacle avoidance trigger timing is usually used. When an obstacle suddenly approaches, the fixed obstacle avoidance trigger timing causes the platform's obstacle avoidance action to be too late, which easily leads to collision.

[0003] Therefore, there is a need to provide an obstacle avoidance method that is both safe and has good real-time performance. Summary of the Invention

[0004] Based on the above analysis, the present invention aims to provide a dynamic obstacle avoidance method for amphibious platforms that integrates scene cognition, in order to solve the problems of existing obstacle avoidance methods relying on human experience, having poor safety, being unable to cope with complex scenes, and having poor real-time performance in obstacle avoidance judgment.

[0005] This invention provides a dynamic obstacle avoidance method for amphibious platforms that integrates scene cognition, the method comprising: Obtain obstacle information; Clustering operations are performed on the obstacle information to obtain obstacle avoidance scene clusters under different obstacle avoidance scene types; Calculate the collision risk index for each obstacle avoidance scenario cluster, and select the obstacle avoidance scenario cluster with the highest collision risk index as the active cluster; The obstacle avoidance mode is determined based on the obstacle information in the activated cluster and the obstacle avoidance scene type corresponding to the activated cluster. The obstacle avoidance duration and lateral offset value are determined based on the obstacle avoidance mode. Multiple candidate trajectories are obtained based on the obstacle avoidance time, the lateral offset value, and the trajectory generation model; Candidate trajectories are selected based on preset kinematic constraints to obtain a candidate set; Calculate the cost function value for each candidate trajectory in the candidate set, and select the trajectory with the smallest cost function value as the optimal obstacle avoidance trajectory.

[0006] Based on a further improvement of the above method, the acquisition of obstacle information includes: Obtain the relative motion parameters of the amphibious platform with each obstacle; The relative motion parameters include relative speed, relative distance, relative azimuth angle, relative heading angle, and speed ratio.

[0007] Based on a further improvement to the above method, the step of performing a clustering operation on the obstacle information to obtain obstacle avoidance scene clusters under different obstacle avoidance scene types includes: The K-means algorithm is used to cluster the relative motion parameters corresponding to each obstacle to obtain obstacle avoidance scenario clusters under different obstacle avoidance scenario types; wherein, the obstacle avoidance scenario types include encounter situation type, left diagonal crossing type, right diagonal crossing type, and overtaking situation type.

[0008] Based on a further improvement to the above method, the calculation of the collision risk index corresponding to each obstacle avoidance scenario cluster includes: Calculate the collision risk score for each obstacle in the obstacle avoidance scenario cluster, and sum the collision risk scores of all obstacles in the obstacle avoidance scenario cluster to obtain the collision risk index; wherein, the collision risk score is: , , , , , These represent the nearest encounter time risk membership degree, nearest encounter distance risk membership degree, relative distance risk membership degree, relative azimuth angle risk membership degree, and speed ratio risk membership degree of the j-th obstacle in the i-th obstacle avoidance scenario cluster, respectively. , , , , These are the nearest encounter time risk membership weight, nearest encounter distance risk membership weight, relative distance risk membership weight, relative azimuth angle risk membership weight, and speed ratio risk membership weight, respectively, for the j-th obstacle in the i-th obstacle avoidance scenario cluster. i=1,2,3,4, j=1,2,3,…,N, where N is the number of obstacles in the i-th obstacle avoidance scenario cluster.

[0009] A further improvement to the above method, the step of determining the obstacle avoidance mode based on the obstacle information in the activated cluster and the obstacle avoidance scene type corresponding to the activated cluster, includes: The obstacle with the highest collision risk score in the activated cluster is selected as the target obstacle; Calculate the target collision risk change rate and the target's closest encounter time corresponding to the target obstacle; Obtain the preset risk threshold and preset encounter time threshold corresponding to this obstacle avoidance scenario type; The obstacle avoidance mode is determined based on the target collision risk change rate, the preset risk threshold, the target's most recent encounter time, and the preset encounter time threshold.

[0010] Based on a further improvement of the above method, the trajectory generation model is a fifth-order polynomial trajectory sampling model based on the Frenet coordinate system.

[0011] Based on a further improvement of the above method, the preset kinematic constraints include: , , Where s is the cumulative longitudinal arc length of the amphibious platform along the global reference path. Let be the instantaneous curvature of the amphibious platform at arc length s. , These are the first and second derivatives of the lateral displacement of the amphibious platform with respect to the arc length s, respectively. The maximum curvature for the amphibious platform's turning. Let be the lateral acceleration of the amphibious platform at time t. The maximum acceleration of the amphibious platform is given. , , , These are the coefficients of the lateral trajectory model in the fifth-order polynomial trajectory sampling model.

[0012] Based on a further improvement of the above method, the cost function is: , in, , , These are the safety cost weight, the tracking cost weight, and the smoothing cost weight, respectively. Let k be the k-th candidate trajectory in the candidate set, where k = 1, 2, 3, ..., N, and N is the number of candidate trajectories in the candidate set. , , These are the safety cost, the tracking cost, and the smoothing cost for the k-th candidate trajectory, respectively.

[0013] Based on a further improvement of the above method, the safety cost weight, tracking cost weight, and smoothing cost weight are obtained in the following way: The basic weights of safety cost, tracking cost, and smoothing cost are determined using an offline hierarchical method. The basic weights of the safety cost, the basic weights of the tracking cost, and the basic weights of the smoothing cost are adjusted based on the collision risk score of the target obstacle.

[0014] A further improvement to the above method involves revising the basic weights of the safety cost, the basic weights of the tracking cost, and the basic weights of the smoothing cost based on the collision risk score of the target obstacle, including: , , , , , , in, , , These are the basic weights for the safety cost, the basic weights for the tracking cost, and the basic weights for the smoothing cost, respectively. , , These are the safety weight surge factor, the tracking weight decay factor, and the smoothing weight fine-tuning factor, respectively. The collision risk score is given to the target obstacle.

[0015] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. This invention provides a dynamic obstacle avoidance method for amphibious platforms that integrates scene cognition. It divides obstacles into multiple obstacle avoidance scenarios using a clustering algorithm, determines activation clusters based on these scenarios, identifies risk patterns based on these clusters, determines the weights of the cost function based on the risk patterns, and finally selects the optimal obstacle avoidance trajectory based on the cost function value. Compared to existing technologies, this method can be applied to high-density obstacle environments, with low computational cost and fast response speed. By introducing an adaptive obstacle avoidance triggering mechanism, it switches between emergency, gradual, and fine-tuning obstacle avoidance modes in real time based on the CRI change rate and TCPA, overcoming the limitation of fixed triggering timing in responding to sudden approach or velocity changes of obstacles, and significantly reducing collision risk.

[0016] 2. This invention provides a dynamic obstacle avoidance method for amphibious platforms that integrates scene cognition. It uses the analytic hierarchy process to determine the three benchmark weights of safety, tracking, and smoothness in the trajectory evaluation cost function. Then, it dynamically adjusts the weights based on the real-time CRI through a weight distortion function. This design prioritizes safety under high collision risk and prioritizes maintaining line-hugging driving and smooth movement under low risk, taking into account the adaptability to both extreme and normal working conditions.

[0017] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0018] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is an example diagram of a dynamic obstacle avoidance method for amphibious platforms that integrates scene cognition, as described in an embodiment of the present invention. Detailed Implementation

[0019] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0020] An amphibious platform refers to equipment / systems capable of maneuvering and operating in both land and water environments. Its core features include cross-domain mobility, mode switching, and environmental adaptation, encompassing both civilian and unmanned applications. This platform belongs to a hierarchical autonomous control system, typically composed of a perception layer, a planning layer, a control layer, and an execution layer. The method proposed in this invention focuses on the planning layer; specifically, it is a dynamic obstacle avoidance method for amphibious platforms that integrates scene cognition. This method primarily aims to establish obstacle avoidance routes within the current scenario, where scene data refers to obstacle information provided by the perception layer.

[0021] A specific embodiment of the present invention discloses a dynamic obstacle avoidance method for amphibious platforms that integrates scene cognition, such as... Figure 1 As shown, the method includes: S1: Obtain obstacle information.

[0022] The amphibious platform utilizes sensors such as lidar and visual cameras to acquire obstacle information under varying weather conditions and distances. Specifically, the data acquired by each sensor is preprocessed. Clustering algorithms are used to identify obstacles in the preprocessed point cloud data, and a target detection model is employed to identify obstacles in the preprocessed visual images. After obtaining the obstacle data, it undergoes time and spatial synchronization processes before multi-sensor data fusion is performed. This yields basic obstacle information, including the obstacle's speed, size, position, and direction of movement.

[0023] Obstacle information is calculated based on the positional relationship between each obstacle and the current amphibious platform. This obstacle information includes the relative motion parameters between the amphibious platform and each obstacle. These relative motion parameters include relative velocity, relative distance, relative azimuth, relative heading angle, and velocity ratio.

[0024] For example, obstacle information can be calculated based on the location of the amphibious platform. ,speed Heading angle and the location of obstacles ,speed heading angle The relative motion parameters between the amphibious platform and the obstacles are calculated. The formulas for calculating each relative motion parameter are as follows: , in, RD is the relative speed, RB is the relative azimuth, RH is the relative heading, and K is the speed ratio.

[0025] S2: Perform clustering operations on the obstacle information to obtain obstacle avoidance scene clusters under different obstacle avoidance scene types.

[0026] This step uses the relative motion parameters of each obstacle as reference information to achieve clustering. The clustering algorithm can be any existing clustering algorithm such as K-means, hierarchical clustering, DBSCAN, Mean Shift, etc. This invention does not limit it, as long as it can classify multiple obstacles based on relative motion parameters. K-means clustering is an unsupervised learning method that divides the data into K clusters based on a similarity criterion. Its goal is to minimize the sum of squared errors within each cluster, i.e.: , in, It is a sample and its cluster center The square of the Euclidean distance between them.

[0027] For example, the K-means algorithm is used to cluster the relative motion parameters corresponding to each obstacle to obtain obstacle avoidance scenario clusters under different obstacle avoidance scenario types. These obstacle avoidance scenario types include encounter situations, left-side diagonal crossings, right-side diagonal crossings, and overtaking situations. Specifically, based on the relative motion parameters of the obstacles and referring to international maritime collision avoidance regulations, the K-means clustering method is used to divide the obstacles into four clusters, each corresponding to one of the four different obstacle avoidance scenarios (encounter situation, left-side diagonal crossing, right-side diagonal crossing, and overtaking situation). Taking the direction pointed to by the platform's stern as the reference, with clockwise as the positive direction, based on the relative heading angle, 354°~6° represents an encounter situation, 112.5°~247.5° represents an overtaking situation, 6°~112.5° represents a right-side diagonal crossing, and 247.5°~354° represents a left-side diagonal crossing.

[0028] After processing by S21-S22, the clustering results of obstacle information can be obtained, that is, obstacle avoidance scene clusters under different obstacle avoidance scene types.

[0029] S3: Calculate the collision risk index corresponding to each obstacle avoidance scenario cluster, and take the obstacle avoidance scenario cluster with the largest collision risk index as the active cluster.

[0030] S31: Calculate the collision risk index corresponding to each obstacle avoidance scenario cluster, including: S311: Calculate the collision risk score for each obstacle in the obstacle avoidance scenario cluster.

[0031] The collision risk score is calculated by weighting the collision risk assessment index parameters (i.e., closest encounter time, closest encounter distance, relative distance, relative azimuth, and speed ratio) with their corresponding weights. Specifically, the collision risk score is: , in, , , , , These are the nearest encounter time risk membership degree, nearest encounter distance risk membership degree, relative distance risk membership degree, relative azimuth angle risk membership degree, and speed ratio risk membership degree of the j-th obstacle in the i-th obstacle avoidance scenario cluster, respectively. , , , , These are the nearest encounter time risk membership weight, nearest encounter distance risk membership weight, relative distance risk membership weight, relative azimuth angle risk membership weight, and speed ratio risk membership weight, respectively, for the j-th obstacle in the i-th obstacle avoidance scenario cluster. i=1,2,3,4, j=1,2,3,…,N, where N is the number of obstacles in the i-th obstacle avoidance scenario cluster.

[0032] The membership degree of the most recent time-related risk is: , , in, Minimum meeting time, The TCPA (Closest Encounter Time) is the time of closest encounter between the amphibious platform and the obstacle. TCPA > 0 indicates the obstacle is approaching the amphibious platform, TCPA < 0 indicates the obstacle is moving away from the amphibious platform, and TCPA = 0 indicates the obstacle is closest to the amphibious platform at this moment. The membership degree of the nearest distance risk is: , , in, Minimum meeting distance For safe encounter distance. DCPA is the closest encounter distance between the amphibious platform and the obstacle.

[0033] The relative distance risk membership degree is: , in, To allow for emergency avoidance distance, For a safe distance, .

[0034] The relative azimuth risk membership degree is: .

[0035] Speed-to-risk membership degree is: , in, This is the reference speed ratio, which is generally taken as 1.

[0036] The collision risk assessment index parameters (i.e., the time of the nearest encounter, the distance of the nearest encounter, the relative distance, the relative azimuth angle, and the speed ratio) are mapped to the interval [0,1] as the risk membership degree of each item. The closer the membership degree value is to 1, the greater the collision risk, and the closer it is to 0, the safer it is.

[0037] Among them, the membership weights of the nearest encounter time risk, the nearest encounter distance risk, the relative distance risk, the relative azimuth risk, and the speed ratio risk are all greater than 0, and their sum is 1, that is: , .

[0038] S312: The collision risk index is obtained by summing the collision risk scores of all obstacles in the obstacle avoidance scenario cluster.

[0039] Based on step S311, the collision risk score corresponding to each obstacle can be obtained. By accumulating the collision risk scores of all obstacles in each obstacle avoidance scenario cluster, the collision risk index corresponding to that obstacle avoidance scenario cluster can be obtained.

[0040] S32: Select the obstacle avoidance scenario cluster with the highest collision risk index as the activation cluster.

[0041] The present invention also provides another method for determining active clusters, as an alternative to the method for determining active clusters given in S3.

[0042] A31: Calculate the collision risk score for each obstacle in the obstacle avoidance scenario cluster.

[0043] This step is implemented in the same way as step S311.

[0044] A32: For each obstacle avoidance scenario cluster, sort the obstacles in the cluster according to the Collision Risk Rating (CRI), and select the three obstacles with the highest CRI as the representatives of the cluster's collision risk. If there are fewer than three obstacles in the cluster, use placeholders with CRI=0 to fill in the gaps. Calculate the collision risk index for the cluster. , in, Let k be the collision risk score representing the k-th collision risk. Let be the importance attenuation factor representing the k-th collision risk. The higher the obstacle collision risk, the larger the corresponding attenuation factor value. For example, the importance attenuation factor can be set empirically, but it needs to ensure... .For example, , , .

[0045] A33: Set the collision risk index threshold for active clusters to... When the cluster's collision risk index is greater than When the cluster is active, it is marked as a candidate active cluster; otherwise, it is an inactive cluster.

[0046] For example, the collision risk index threshold can be set to 0.7.

[0047] A34: Select the obstacle avoidance scenario cluster with the highest collision risk index from the candidate activation clusters as the activation cluster.

[0048] To reduce the computational cost of the algorithm, if multiple active clusters exist at the same time, only the cluster with the highest collision risk index is selected as the active cluster, that is, at most only one active cluster exists. This selection method can further improve the real-time performance of the system.

[0049] This invention proposes two methods for calculating activation clusters. Method 1 has low computational complexity and good real-time performance, accurately identifying high-risk areas formed by clusters of multiple low- and medium-risk obstacles. The computational logic remains consistent regardless of the number of obstacles within the cluster, and the computation time remains relatively stable, preventing delays caused by sudden increases in the number of obstacles. The system's real-time performance is guaranteed, but it is insensitive to individual extremely high-risk obstacles. Method 2 focuses on the top three highest-risk obstacles within a cluster, aligning with the decision-making logic of human drivers prioritizing the most dangerous targets. It exhibits strong anti-interference capabilities, good robustness, and adaptability to extreme scenarios, but its computational complexity is high and may underestimate the risk of cluster obstacles. Therefore, when the detected obstacle density is low, Method 1 can be used to ensure system real-time performance; when the detected obstacle density is high, Method 2 should be used to ensure the accuracy of risk assessment.

[0050] Furthermore, if the current obstacle density is low, Method 1 is used to determine the active cluster. However, if a high-risk obstacle with CRI>0.8 is found in step S311, Method 2 can be directly used to determine the active cluster.

[0051] S4: Determine the obstacle avoidance mode based on the obstacle information in the activated cluster and the obstacle avoidance scene type corresponding to the activated cluster.

[0052] In aquatic environments, surface conditions are complex and constantly changing. The amphibious platform's speed and direction relative to obstacles change in real time, leading to a corresponding increase in collision risk. When a dynamic obstacle suddenly appears ahead or its speed suddenly increases, obstacle avoidance decision-making triggering mechanisms with fixed parameters struggle to respond promptly. Therefore, it is necessary to design an adaptive obstacle avoidance triggering mechanism that adjusts parameters accordingly. This process includes: S41: Select the obstacle with the highest collision risk score in the activated cluster as the target obstacle.

[0053] S42: Calculate the target collision risk change rate and the target nearest encounter time corresponding to the target obstacle.

[0054] The collision risk change rate is as follows: , in, The collision risk score for the target obstacle at time t. The collision risk score for the target obstacle at time t-1.

[0055] The next likely time to meet the target is: .

[0056] S43: Obtain the preset risk threshold and preset encounter time threshold corresponding to this obstacle avoidance scenario type.

[0057] The system pre-sets risk thresholds (i.e., collision risk change rate thresholds) and encounter time thresholds (i.e., TCPA thresholds) for four obstacle avoidance scenarios: encounter, left-side diagonal crossing, right-side diagonal crossing, and overtaking. This allows for dynamic adjustment of the obstacle avoidance triggering mechanism for different scenarios. The risk and encounter time thresholds can be set based on factors such as the amphibious platform type, size, weight, mission type, and current aquatic environment. For frequently used aquatic environments, simulation methods can also be used to obtain the thresholds.

[0058] Among them, the preset risk thresholds include the high risk change rate threshold range, the potential risk change rate threshold range, and the safe change rate threshold range, and the preset encounter time thresholds include the high risk encounter time threshold range, the potential risk encounter time threshold range, and the safe encounter time threshold range.

[0059] For example, the preset risk threshold can be set to a high-risk change rate threshold range. The potential risk change rate threshold range is The safety change rate threshold range is The preset meeting time threshold can be set to a high-risk meeting time threshold range. The potential risk will be encountered within a time threshold range. The safe meeting time threshold range is .

[0060] S44: Determine the obstacle avoidance mode based on the target collision risk change rate, the preset risk threshold, the target's most recent encounter time, and the preset encounter time threshold.

[0061] The obstacle avoidance modes include emergency mode, progressive mode, and fine-tuning mode.

[0062] This step can be represented as a preset function MODE: , When the target collision risk change rate is within the high-risk change rate threshold range or the target TCPA is within the high-risk encounter time threshold range, an emergency mode is triggered, indicating that an obstacle has suddenly appeared ahead or the obstacle's speed has suddenly increased, requiring emergency avoidance. When the target collision risk change rate is within the potential risk change rate threshold range and the target TCPA is within the potential risk encounter time threshold range, a progressive mode is triggered, indicating that the current encounter situation is stable and a balanced obstacle avoidance strategy is sufficient. When the target collision risk change rate is within the safe change rate threshold range and the target TCPA is within the safe encounter time threshold range, a fine-tuning mode is triggered. In this case, the amphibious platform's operating environment is safe, and a strategy that prioritizes tracking the global reference path should be adopted. Thus, the obstacle avoidance mode of the amphibious platform is determined based on the target collision risk change rate and the target's most recent encounter time.

[0063] S5: Determine the obstacle avoidance duration and lateral offset value based on the obstacle avoidance mode.

[0064] The obstacle avoidance time and lateral offset value can be set according to factors such as the type, size, and weight of the amphibious platform, the type of mission being performed, and the current aquatic environment.

[0065] For example, in emergency mode, the obstacle avoidance duration T can be set to 3-5 seconds; in progressive mode, the obstacle avoidance duration T can be set to 8-10 seconds; and in fine-tuning mode, the obstacle avoidance duration T can be set to 10-15 seconds. In an encounter situation, the lateral offset can be set to 3 times the obstacle width; for a left diagonal crossing, the lateral offset can be set to 2.5 times the obstacle width; for a right diagonal crossing, the lateral offset can be set to 2.5 times the obstacle width; and for an overtaking situation, the lateral offset can be set to 2 times the obstacle width.

[0066] S6: Based on the obstacle avoidance time, the lateral offset value, and the trajectory generation model, multiple candidate trajectories are obtained.

[0067] The trajectory generation model is a fifth-order polynomial trajectory sampling model based on the Frenet coordinate system. To solve for the longitudinal and lateral motion of the amphibious platform, the global Cartesian coordinate system is converted to the path-following Frenet coordinate system. A pre-set global reference trajectory is used as the baseline of the Frenet coordinate system, and any trajectory point in the global coordinate system is converted to Frenet coordinates. Specifically, lateral and longitudinal motions use polynomials of different orders to ensure the continuity and smoothness of the trajectory. For the lateral trajectory, complete kinematic constraints (position, velocity, acceleration) are applied to the trajectory start and end points to ensure a smooth trajectory without abrupt changes. The calculation formula is as follows: , The kinematic constraints are set as follows: For the starting point, i.e., t=0: , The initial lateral position, velocity, and acceleration are determined by the current state of the platform. For the endpoint, i.e., when t=T: , In this case, the final velocity and acceleration are usually set to 0.

[0068] Then construct a system of linear equations and solve for the coefficients of the polynomials: .

[0069] For the longitudinal trajectory, a fourth-order polynomial is used to constrain the velocity of the trajectory: , The constraints are: , Then construct a system of linear equations and solve for the coefficients of the polynomials: .

[0070] The current lateral position, current lateral velocity, current lateral acceleration, current longitudinal position, current longitudinal velocity, and current longitudinal acceleration can be obtained by the amphibious platform's perception layer, while the target's longitudinal velocity can be determined by factors such as mission requirements, environmental constraints, and vehicle dynamics constraints.

[0071] Discrete sampling is performed on the obstacle avoidance time and lateral offset value to generate M candidate trajectories. For example, the sampling time interval can be set to 0.05s, and M is typically 50~100.

[0072] For example, a two-dimensional sampling space is set up: Among them, obstacle avoidance time The sampling time interval is ; lateral offset The sampling offset interval is , , For the width of the amphibious platform, The width of the obstacle. For safety offset, The values ​​of each parameter change dynamically. For example, in emergency mode, the obstacle avoidance duration parameter values ​​are... , , The current obstacle avoidance scenario is an encounter situation, with a safe offset. Sampling offset interval Among them, lateral offset This corresponds to the lateral offset of the target in the lateral trajectory.

[0073] After determining the sampling interval, sampling time interval, and sampling offset interval, obstacle avoidance duration sampling point sequence and lateral offset sampling point sequence are generated, respectively. Then, a Cartesian product operation is performed to obtain all possible parameter combinations. For the i-th parameter combination [obstacle avoidance duration]... Horizontal offset Substitute these equations into the linear equations corresponding to the horizontal and vertical trajectories, respectively, and solve for the coefficients of their respective polynomials. After determining the coefficients, solve for each time step. trajectory points Each trajectory point forms a candidate trajectory.

[0074] S7: Select candidate trajectories based on preset kinematic constraints to obtain a candidate set.

[0075] The preset kinematic constraints are as follows: , , Where s is the cumulative longitudinal arc length of the amphibious platform along the global reference path. Let be the instantaneous curvature of the amphibious platform at arc length s. , These are the first and second derivatives of the lateral displacement of the amphibious platform with respect to the arc length s, respectively. The maximum curvature for the amphibious platform's turning. Let be the lateral acceleration of the amphibious platform at time t. The maximum acceleration of the amphibious platform is given. , , , These are the coefficients of the lateral trajectory model in the fifth-order polynomial trajectory sampling model.

[0076] By applying the above kinematic constraints, invalid trajectories that do not meet the constraints are deleted, and trajectories that meet the constraints are retained as one of the candidate trajectories.

[0077] S8: Calculate the cost function value of each candidate trajectory in the candidate set, and take the trajectory with the smallest cost function value as the optimal obstacle avoidance trajectory.

[0078] The cost function is: , in, , , These are the safety cost weight, the tracking cost weight, and the smoothing cost weight, respectively. Let k be the k-th candidate trajectory in the candidate set, where k = 1, 2, 3, ..., N, and N is the number of candidate trajectories in the candidate set. , , These are the safety cost, the tracking cost, and the smoothing cost for the k-th candidate trajectory, respectively.

[0079] The security cost is as follows: , in, This represents the distance between each point on the trajectory and the target obstacle.

[0080] The cost of following the path is: .

[0081] The smoothing cost is: .

[0082] The safety cost weight, tracking cost weight, and smoothing cost weight are obtained as follows: S81: The offline hierarchical method (AHP) is used to determine the basic weights of safety cost, line-following cost, and smoothing cost, including: S811: Construct a hierarchical structure. A three-layer structure is constructed, including the target layer, the criterion layer, and the solution layer. The target layer outputs the optimal obstacle avoidance trajectory; the criterion layer sets evaluation metrics including safety, tracking performance, and smoothness; the solution layer generates a set of candidate trajectories.

[0083] S812: Construct the judgment matrix. Use a scaling method to assign importance to the three criteria, with importance increasing from 1 to 9. Define the judgment matrix A: , in, , .

[0084] Based on driver experience, the three criteria are ranked in order of importance: safety > line following > smoothness. Different weights are assigned to each criterion, which are then used as the values ​​of the off-diagonal elements in the judgment matrix A. For example, safety is slightly more important than line following, with an importance scale of 3; safety is significantly more important than smoothness, with an importance scale of 5; and line following is slightly more important than smoothness, with an importance scale of 3. The values ​​of the off-diagonal elements can then be obtained. .

[0085] S813: Calculate the weight vector. Construct the homogeneous linear equation: , in, Let A be the largest eigenvalue of matrix A. These are the corresponding feature vectors. The feature vectors are then normalized to obtain the baseline weights for the three criteria.

[0086] S814: Consistency Check. Design consistency indicators. Let n be the order of the judgment matrix A. Calculate the consistency ratio. RI is a random consistency index, obtained by looking up a table. When CR meets the set consistency threshold, it means that the calculated benchmark weights have passed the consistency test and there is no logical contradiction.

[0087] S82: Adjust the basic weights of the safety cost, the basic weights of the tracking cost, and the basic weights of the smoothing cost based on the collision risk score of the target obstacle.

[0088] After obtaining the baseline weights, the weights are adjusted using the highest CRI of the obstacles in the current active cluster, calculated as follows: , , , , , , in, , , These are the basic weights for the safety cost, the basic weights for the tracking cost, and the basic weights for the smoothing cost, respectively. , , These are the safety weight surge factor, the tracking weight decay factor, and the smoothing weight fine-tuning factor, respectively. The collision risk score is given to the target obstacle.

[0089] For example, it can be set , , .

[0090] The final cost function weights are obtained by dynamically adjusting the basic weights of safety cost, tracking cost, and smoothing cost based on the collision risk score of the target obstacle.

[0091] Compared with existing technologies, this embodiment provides a dynamic obstacle avoidance method for amphibious platforms that integrates scene cognition. It divides obstacles into multiple obstacle avoidance scenarios using a clustering algorithm, determines activation clusters based on these scenarios, identifies risk patterns based on these clusters, determines the weights of the cost function based on the risk patterns, and finally selects the optimal obstacle avoidance trajectory based on the cost function value. Compared with existing technologies, this method can be applied to high-density obstacle environments with low computational cost and fast response speed. By introducing an adaptive obstacle avoidance triggering mechanism, it switches between emergency, gradual, and fine-tuning obstacle avoidance modes in real time based on the CRI change rate and TCPA, overcoming the shortcomings of fixed triggering timing in dealing with sudden approach or speed changes of obstacles, and significantly reducing collision risk. The analytic hierarchy process (AHP) is used to determine the safety, tracking, and smoothness benchmark weights of the trajectory evaluation cost function, and then the weights are dynamically adjusted based on the real-time CRI using a weight distortion function. This design prioritizes safety under high collision risk and prioritizes maintaining line-hugging and smooth movement under low risk, taking into account adaptability to both extreme and normal operating conditions.

[0092] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0093] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A dynamic obstacle avoidance method for amphibious platforms integrating scene cognition, characterized in that, The method includes: Obtain obstacle information; Clustering operations are performed on the obstacle information to obtain obstacle avoidance scene clusters under different obstacle avoidance scene types; Calculate the collision risk index for each obstacle avoidance scenario cluster, and select the obstacle avoidance scenario cluster with the highest collision risk index as the active cluster; The obstacle avoidance mode is determined based on the obstacle information in the activated cluster and the obstacle avoidance scene type corresponding to the activated cluster. The obstacle avoidance duration and lateral offset value are determined based on the obstacle avoidance mode. Multiple candidate trajectories are obtained based on the obstacle avoidance time, the lateral offset value, and the trajectory generation model; Candidate trajectories are selected based on preset kinematic constraints to obtain a candidate set; Calculate the cost function value for each candidate trajectory in the candidate set, and select the trajectory with the smallest cost function value as the optimal obstacle avoidance trajectory.

2. The dynamic obstacle avoidance method for an amphibious platform integrating scene cognition according to claim 1, characterized in that, The acquisition of obstacle information includes: Obtain the relative motion parameters of the amphibious platform with each obstacle; The relative motion parameters include relative speed, relative distance, relative azimuth angle, relative heading angle, and speed ratio.

3. The dynamic obstacle avoidance method for an amphibious platform integrating scene cognition according to claim 2, characterized in that, The step of performing clustering operations on the obstacle information to obtain obstacle avoidance scene clusters under different obstacle avoidance scene types includes: The K-means algorithm is used to cluster the relative motion parameters corresponding to each obstacle to obtain obstacle avoidance scenario clusters under different obstacle avoidance scenario types; wherein, the obstacle avoidance scenario types include encounter situation type, left diagonal crossing type, right diagonal crossing type, and overtaking situation type.

4. The dynamic obstacle avoidance method for an amphibious platform integrating scene cognition according to claim 3, characterized in that, The calculation of the collision risk index corresponding to each obstacle avoidance scenario cluster includes: Calculate the collision risk score for each obstacle in the obstacle avoidance scenario cluster, and sum the collision risk scores of all obstacles in the obstacle avoidance scenario cluster to obtain the collision risk index; wherein, the collision risk score is: ， , , , , These represent the nearest encounter time risk membership degree, nearest encounter distance risk membership degree, relative distance risk membership degree, relative azimuth angle risk membership degree, and speed ratio risk membership degree of the j-th obstacle in the i-th obstacle avoidance scenario cluster, respectively. , , , , These are the nearest encounter time risk membership weight, nearest encounter distance risk membership weight, relative distance risk membership weight, relative azimuth angle risk membership weight, and speed ratio risk membership weight, respectively, for the j-th obstacle in the i-th obstacle avoidance scenario cluster. i=1,2,3,4, j=1,2,3,…,N, where N is the number of obstacles in the i-th obstacle avoidance scenario cluster.

5. The dynamic obstacle avoidance method for an amphibious platform integrating scene cognition according to claim 4, characterized in that, The step of determining the obstacle avoidance mode based on the obstacle information in the activated cluster and the obstacle avoidance scene type corresponding to the activated cluster includes: The obstacle with the highest collision risk score in the activated cluster is selected as the target obstacle; Calculate the target collision risk change rate and the target's closest encounter time corresponding to the target obstacle; Obtain the preset risk threshold and preset encounter time threshold corresponding to this obstacle avoidance scenario type; The obstacle avoidance mode is determined based on the target collision risk change rate, the preset risk threshold, the target's most recent encounter time, and the preset encounter time threshold.

6. The dynamic obstacle avoidance method for an amphibious platform integrating scene cognition according to claim 5, characterized in that, The trajectory generation model is a fifth-order polynomial trajectory sampling model based on the Frenet coordinate system.

7. The dynamic obstacle avoidance method for an amphibious platform integrating scene cognition according to claim 6, characterized in that, The preset kinematic constraints include: ， ， Where s is the cumulative longitudinal arc length of the amphibious platform along the global reference path. Let be the instantaneous curvature of the amphibious platform at arc length s. , These are the first and second derivatives of the lateral displacement of the amphibious platform with respect to the arc length s, respectively. The maximum curvature for the amphibious platform's turning. Let be the lateral acceleration of the amphibious platform at time t. The maximum acceleration of the amphibious platform is given. , , , These are the coefficients of the lateral trajectory model in the fifth-order polynomial trajectory sampling model.

8. The dynamic obstacle avoidance method for an amphibious platform integrating scene cognition according to claim 1, characterized in that, The cost function is: ， in, , , These are the safety cost weight, the tracking cost weight, and the smoothing cost weight, respectively. Let k be the k-th candidate trajectory in the candidate set, where k = 1, 2, 3, ..., N, and N is the number of candidate trajectories in the candidate set. , , These are the safety cost, the tracking cost, and the smoothing cost for the k-th candidate trajectory, respectively.

9. The dynamic obstacle avoidance method for an amphibious platform integrating scene cognition according to claim 8, characterized in that, The safety cost weight, tracking cost weight, and smoothing cost weight are obtained in the following way: The basic weights of safety cost, tracking cost, and smoothing cost are determined using an offline hierarchical method. The basic weights of the safety cost, the basic weights of the tracking cost, and the basic weights of the smoothing cost are adjusted based on the collision risk score of the target obstacle.

10. The dynamic obstacle avoidance method for an amphibious platform integrating scene cognition according to claim 9, characterized in that, The correction of the basic weights of the safety cost, the basic weights of the tracking cost, and the basic weights of the smoothing cost based on the collision risk score of the target obstacle includes: ， ， ， ， ， ， in, , , These are the basic weights for the safety cost, the basic weights for the tracking cost, and the basic weights for the smoothing cost, respectively. , , These are the safety weight surge factor, the tracking weight decay factor, and the smoothing weight fine-tuning factor, respectively. The collision risk score is given to the target obstacle.

Images