Water surface environment perception method based on bayesian-fuzzy fusion
By employing a Bayesian-fuzzy fusion-based water surface environment perception method, and utilizing multimodal information fusion and hierarchical reasoning models, the problem of insufficient robustness of unmanned waterborne platforms in environmental perception in complex waters was solved, achieving efficient and stable obstacle recognition and dynamic target tracking.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WUHU SHIPYARD CO LTD
- Filing Date
- 2026-03-31
- Publication Date
- 2026-07-14
AI Technical Summary
Existing unmanned water platforms lack a mechanism for hierarchical collaboration and unified modeling of Bayesian inference and fuzzy discrimination in complex and dynamic waters, resulting in limited robustness of environmental perception.
A water surface environment perception method based on Bayesian-fuzzy fusion is adopted. Through multimodal information fusion, Bayesian inference model and fuzzy inference model, the true probability of obstacles is identified and predicted. Bayesian inference is combined to continuously recursively estimate the target state, and fuzzy inference is used to flexibly distinguish complex features.
It improves the accuracy and stability of environmental perception, enhances the robustness of perception in complex scenarios, effectively identifies obstacles and floating objects, improves the continuity and reliability of dynamic target tracking, and reduces computational and energy consumption.
Smart Images
Figure CN122391801A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of environmental perception technology, and more specifically, this invention relates to a water surface environmental perception method based on Bayesian-fuzzy fusion. Background Technology
[0002] As unmanned surface vessels (USVs) are increasingly used in smart port inspections, waterway security, environmental monitoring, maritime emergency rescue, waterway patrols, and autonomous transportation, their autonomous perception and decision-making capabilities in complex and dynamic waters have become a key factor restricting the practical application and large-scale deployment of these systems. Environmental perception, as the foundation for USVs to achieve autonomous navigation, target identification, obstacle warning, and path planning, directly relates to platform operational safety, mission efficiency, and system robustness.
[0003] Because the aquatic environment presents not only "probabilistic uncertainties" such as sensor observation noise, but also "fuzzy uncertainties" such as blurred target boundaries, water surface reflection artifacts, and semantic ambiguity, existing methods typically focus on only one of statistical filtering (probabilistic) or rule-based discrimination (fuzzy), lacking a mechanism for hierarchical collaboration and unified modeling of the two, resulting in limited perception robustness in complex scenarios. Summary of the Invention
[0004] This invention provides a water surface environment perception method based on Bayesian-fuzzy fusion, which aims to solve at least one of the above-mentioned problems.
[0005] This invention is implemented as follows: a water surface environment perception method based on Bayesian-fuzzy fusion, the method being as follows:
[0006] (1) Collect multimodal environmental perception information, perform time and space alignment on the collected modal environmental perception information, and identify obstacles in the current environment based on the multimodal environmental perception information;
[0007] (2) Based on the Bayesian inference model and the fuzzy inference model, predict the probability P that the target obstacle in the current environment is a real obstacle, and identify the target obstacle with a large probability P as an obstacle.
[0008] Furthermore, the process of obtaining probability P is as follows:
[0009] (21) Obtain the observation vector of the target obstacle at the current time k. The observation vector at the current moment Input the Bayesian inference model, and the Bayesian inference model outputs the first probability that the target obstacle in the current environment is a real obstacle;
[0010] (22) Collect the current water surface environment parameters, input the current water surface environment parameters and the first probability that the target obstacle belongs to the real obstacle into the fuzzy inference model, and the fuzzy inference model infers the second probability that the target obstacle belongs to the real obstacle;
[0011] (23) The first probability and the second probability that the target obstacle belongs to the real obstacle are fused together to form the probability P that the target obstacle belongs to the real obstacle.
[0012] Furthermore, the observation vector This includes: state observation of the target obstacle, spatial geometric observation, visual semantic observation, and the observation state of the ship itself. Among them, the state observation of the target obstacle includes the relative distance, relative azimuth, and Doppler radial velocity of the target obstacle; the spatial geometric observation of the target obstacle includes the center coordinates of the three-dimensional bounding box of the target obstacle, its length and width dimensions, and the depth extreme points of the underwater contour; the visual semantic observation of the target obstacle includes the pixel coordinates of the center of the two-dimensional detection box of the target obstacle; and the observation state of the ship itself includes the ship's heading, speed, and three-axis attitude angles.
[0013] Furthermore, the water surface environmental parameters include: water surface reflection intensity, image edge continuity, radar echo stability, sonar profile integrity, visibility coefficient, and wave level parameters.
[0014] Furthermore, obstacle information of obstacles in the current environment is stored in a record table. The obstacle information includes obstacle category, obstacle identifier, and obstacle observation status. The obstacle observation status includes obstacle state observation, spatial geometric observation status, and visual semantic observation status.
[0015] Furthermore, the process of extracting the target obstacle is as follows:
[0016] Based on the obstacle information extracted from the previous frame, detect whether there are any newly added obstacles or unstable obstacles in the current frame. If the detection result is positive, then the newly added obstacles and unstable obstacles will be taken as target obstacles.
[0017] Furthermore, the difference between the Doppler radial velocity of the obstacle in the current frame and the Doppler radial velocity of the obstacle in the previous frame is calculated. If the difference is greater than or equal to a set difference threshold, the state of the corresponding obstacle is considered unstable. If the difference is less than the set difference threshold, the state of the corresponding obstacle is considered stable.
[0018] Furthermore, after step (2), the following is also included:
[0019] The grid state of the target obstacle in the two-dimensional grid map is determined based on probability P. The grid state includes: occupied, empty, and unknown.
[0020] Furthermore, the process of marking the raster state is as follows:
[0021] If the probability P is greater than the first probability threshold, the corresponding target obstacle is determined to be a real obstacle, and the grid marker of the corresponding target obstacle is marked as occupied. If the probability is less than the second probability threshold, the corresponding target obstacle is determined not to be an obstacle, and the grid marker of the corresponding target obstacle is marked as idle. The grid to which the target obstacle belongs, which is between the first probability threshold and the second probability threshold, is marked as unknown. The first probability threshold is greater than the second probability threshold.
[0022] This invention addresses both probabilistic and fuzzy uncertainties in complex aquatic environments through a hierarchical coupling of Bayesian inference and fuzzy discrimination, thereby improving the accuracy and stability of environmental perception results and enhancing the robustness of perception in complex scenarios. Furthermore, because the Bayesian module performs continuous recursive estimation of the target state, and the fuzzy module performs flexible discrimination of complex features such as reflection artifacts, boundary ambiguity, and semantic unclearness, it can more effectively identify obstacles, floating objects, and channel boundaries, and improve the continuity and reliability of dynamic target tracking. Attached Figure Description
[0023] Figure 1 A flowchart of a water surface environment perception method based on Bayesian-fuzzy fusion provided for embodiments of the present invention. Detailed Implementation
[0024] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, so as to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention.
[0025] Figure 1 A flowchart of a water surface environment perception method based on Bayesian-fuzzy fusion provided in this embodiment of the invention is shown below.
[0026] (1) Collect multimodal environmental perception information, perform time and space alignment on the collected modal environmental perception information, and identify obstacles in the current environment based on the multimodal environmental perception information;
[0027] In this embodiment of the invention, the multimodal environmental awareness information includes:
[0028] The environmental perception information collected by the visible light camera, the environmental point cloud collected by the millimeter-wave radar, the ship's heading, speed, and three-axis attitude angles provided by the inertial navigation unit (IMU) and GNSS, and the sonar data collected by the multi-beam sonar are all processed sequentially, followed by time synchronization and spatial synchronization. Time synchronization and spatial synchronization are referred to as spatiotemporal synchronization, which maps various types of observation information to the same spatiotemporal reference frame. The preprocessing of the environmental perception information includes: performing dehazing, brightness equalization, reflection suppression, and edge enhancement on the environmental images; performing clutter suppression and point filtering on the point cloud data; performing denoising and contour enhancement on the sonar echo (sonar data); and performing outlier removal and geometric clustering on the environmental point cloud, thereby constructing a standardized multidimensional observation vector to provide a unified input for subsequent fusion inference.
[0029] (2) Based on the Bayesian inference model and the fuzzy inference model, predict the probability P that the target obstacle in the current environment is a real obstacle, identify the target obstacle with a large probability P as an obstacle, and mark it on the map;
[0030] In this embodiment of the invention, the probability that the target obstacle is a real obstacle is... The acquisition process is as follows:
[0031] (21) Obtain the observation vector of the target obstacle at the current time k. The observation vector at the current moment Input a Bayesian inference model, and the Bayesian inference model outputs the first probability that the target obstacle in the current environment is a real obstacle. ;
[0032] This invention utilizes Bayesian inference to recursively estimate the target's position, velocity, motion trend, and environmental state. Let the target state vector at time k be... The observation vector is Then, under the linear Gaussian condition, the state equation is established. With observation equation ,in, Here is the state transition matrix. For the observation matrix, For process noise, To observe the noise, For a moment The state vector is used to recursively calculate the posterior distribution of the target through a prediction and update process, achieving continuous estimation of obstacle positions and dynamic target trajectories. The Bayesian inference module mainly undertakes the functions of observation uncertainty quantification, target dynamic modeling, and continuous state update. The Bayesian inference module operates in a recursive manner of "prior prediction - observation correction - posterior update". For the ... At time 1, the posterior state of the previous time step is first estimated. and its error covariance matrix Combined with the state transition matrix Perform a priori prediction on the target state at the current moment to obtain the priori state estimate at the current moment. Simultaneously calculate the prior error covariance matrix at the current time. ,in, The process noise covariance matrix is used to characterize the uncertainties caused by target maneuvering changes, wave disturbances, and platform motion compensation residuals on the state propagation process. Subsequently, the multi-source observation vectors, after spatiotemporal alignment and normalization, are... Input the observation model and calculate And based on the prior error covariance matrix Covariance matrix of observation noise Calculate the covariance matrix Based on this, the Kalman gain is further calculated. The gain is then used to correct the prior state to obtain the posterior state estimate. Simultaneously update the posterior error covariance matrix. ,in, Using the identity matrix, through the above recursive process, the system can fuse the prior state of the previous moment with the observation evidence of the current moment in real time during the continuous observation process, so as to continuously estimate the target's position, velocity, heading change trend and local environmental state.
[0033] In this embodiment of the invention, the observation vector This includes: state observation of the target obstacle, spatial geometric observation, visual semantic observation, and the observation state of the ship itself. Among them, the state observation of the target obstacle includes the relative distance, relative azimuth, and Doppler radial velocity of the target obstacle, provided by millimeter-wave radar; the spatial geometric observation of the target obstacle includes the center coordinates of the three-dimensional bounding box of the target obstacle, its length and width dimensions, and the depth extreme points of the underwater contour, extracted by clustering lidar point cloud and sonar data; the visual semantic observation of the target obstacle includes the pixel coordinates of the center of the two-dimensional detection box of the target obstacle, output by target detection from environmental images acquired by the visible light camera; the observation state of the ship itself includes the ship's heading, speed, and three-axis attitude angles, acquired by the inertial navigation unit (IMU) and GNSS.
[0034] (22) Collect the current water surface environment parameters and the first probability that the target obstacle is a real obstacle. Inputting the fuzzy inference model, the fuzzy inference model infers the second probability that the target obstacle is a real obstacle. ;
[0035] In this embodiment of the invention, the fuzzy inference model includes a membership function and a fuzzy rule base. The membership function combines water surface environmental parameters and the first probability of belonging to a real obstacle. The process is transformed into a fuzzy semantic description. Based on the fuzzy rules stored in the fuzzy rule base, the fuzzy state corresponding to the fuzzy semantic description is output, which is the second probability that the target obstacle belongs to a real obstacle. .
[0036] In this embodiment of the invention, the water surface environmental parameters include: water surface reflection intensity, image edge continuity, radar echo stability, sonar profile integrity, visibility coefficient, and wave level parameters.
[0037] The water surface reflection intensity is calculated by normalizing the proportion of pixels in the bright areas, the local brightness variance, and the area of the specular reflection region in the visible light image (environmental image); the image edge continuity is obtained by statistically analyzing the edge pixel connectivity length, boundary closure degree, and edge direction consistency within the target candidate region; the radar echo stability is calculated by the fluctuation of target echo amplitude between adjacent time points, the continuous existence rate of point traces, and the smoothness of radial velocity changes; the sonar contour integrity is extracted by the echo contour closure degree, point cloud clustering integrity rate, and depth boundary continuity; the visibility coefficient is comprehensively estimated by the image contrast attenuation degree, the clarity of distant targets, and meteorological sensor auxiliary information; the wave level parameter can be jointly characterized by the pitch, roll, heave amplitudes, and water surface texture undulation intensity output by the platform's inertial navigation system. All of the above parameters are normalized before entering the fuzzy inference model to map them to a unified numerical range for membership division and rule matching. For situations where sensor performance is inconsistent or target characteristics are unstable under different operating conditions, the fuzzy inference model can effectively handle features that are difficult to describe with precise mathematical models, thus compensating for the shortcomings of pure probabilistic models in the semantic analysis of complex water surfaces.
[0038] To improve the engineering feasibility and scenario adaptability of the fuzzy rule base, the membership functions and rule thresholds corresponding to each input parameter in this invention are determined primarily through the following methods: First, based on measured data, historical navigation data, and sensor calibration results in typical water scenarios, statistical analysis is performed on each parameter to obtain its value distribution range under different operating conditions. Second, combined with domain experience knowledge, the parameters are divided into linguistic variables such as "low, medium, high" or "weak, medium, strong," and corresponding triangular, trapezoidal, or Gaussian membership functions are established. Finally, through offline experimental comparison, simulation testing, and iterative correction of sample scenarios, the boundaries of each membership interval and the rule trigger threshold are optimized, so that the fuzzy rules can not only reflect the changing patterns of the water environment but also adapt to the perception characteristics under different platforms and sensor combinations. In other words, the fuzzy rule parameters are derived from a joint determination method of "statistical characteristics of sensor measurement data + scenario experience knowledge + experimental calibration correction," rather than simply subjective settings.
[0039] (23) The first probability that the target obstacle is a real obstacle and the second probability The probability of a target obstacle being a real obstacle is calculated by merging the data. .
[0040] The first probability output by the Bayesian inference model that the target obstacle is a real obstacle. The second probability inferred by the fuzzy inference model that the target obstacle is a real obstacle Perform dynamic weighted fusion to obtain the probability that the target obstacle is a real obstacle. The specific calculation formula is as follows:
[0041] ;
[0042] in, These are the weighting coefficients, and In scenes with strong reflection, high blur, and low visibility, increase the weighting coefficient. It can enhance the second probability of the fuzzy inference model outputting a real obstacle. The effect is to increase the weighting coefficient in scenarios with stable observations, clear boundaries, and low noise. Improve the first probability of the Bayesian inference model outputting a real obstacle. This enhances the adaptability and stability of the overall perception results.
[0043] After extracting obstacles in the current environment based on the multimodal environment perception information of the current frame, the obstacle information is stored in the record table. The obstacle information includes obstacle category, obstacle identifier and obstacle observation state. The obstacle observation state includes obstacle state observation, spatial geometric observation state and visual semantic observation state.
[0044] To balance recognition accuracy and real-time performance, the system detects whether new or unstable obstacles exist in the current frame based on obstacle information extracted from the previous frame. If the detection result is positive, the new or unstable obstacles are designated as target obstacles. This invention evaluates the stability of an obstacle's state by measuring the fluctuation of its Doppler radial velocity. Specifically, it calculates the difference between the Doppler radial velocity of an obstacle in the current frame and that of an obstacle in the previous frame. If this difference is greater than or equal to a set threshold, the obstacle is considered unstable; otherwise, it is considered stable.
[0045] Each newly added obstacle is assigned a unique obstacle identifier. When a new obstacle identifier appears in the current frame, it is determined that a new obstacle has appeared in the current frame compared to the previous frame. If the obstacles extracted based on the multimodal environmental perception information of the current frame are neither newly added obstacles nor unstable obstacles compared to the previous frame, it is determined that there is no target obstacle in the current frame. The Bayesian inference model and fuzzy inference model are used to perform realism detection on the obstacles, which greatly reduces the execution cost. When there is a target obstacle in the current frame, the Bayesian inference model and fuzzy inference model are used to perform realism detection on the obstacles identified as target obstacles, which greatly improves the detection accuracy of obstacles with a low execution cost.
[0046] The map in this invention is a two-dimensional grid map. After determining the probability P that a target obstacle in the current environment belongs to a real obstacle, if the probability P is greater than a first probability threshold, the corresponding target obstacle is determined to be a real obstacle, and the grid marker of the corresponding target obstacle is marked as occupied. If the probability P is less than a second probability threshold, the corresponding target obstacle is determined not to be an obstacle, and the grid marker of the corresponding target obstacle is marked as idle. The grid to which the target obstacle belongs, which is between the first probability threshold and the second probability threshold, is marked as unknown. The first probability threshold is greater than the second probability threshold.
[0047] The environment perception method based on Bayesian-fuzzy fusion provided by this invention has the following beneficial technical effects:
[0048] (1) By layered coupling of Bayesian inference and fuzzy discrimination, the problem of probability uncertainty and fuzzy uncertainty in complex water environment can be addressed at the same time, thereby improving the accuracy and stability of environmental perception results and the perception robustness in complex scenarios. In addition, since the Bayesian module continuously recursively estimates the target state and the fuzzy module flexibly discriminates complex features such as reflection artifacts, boundary ambiguity and semantic ambiguity, it can more effectively identify obstacles, floating objects and channel boundaries, and improve the continuity and reliability of dynamic target tracking.
[0049] (2) By identifying target obstacles in obstacles, the perception accuracy and response speed are guaranteed, while effectively reducing unnecessary computational consumption and energy consumption, making it more suitable for deployment on unmanned water platforms with limited computing power and endurance.
[0050] The present invention has been described by way of example. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other occasions without modification, are all within the protection scope of the present invention.
Claims
1. A water surface environment perception method based on Bayesian-fuzzy fusion, characterized in that, The method is as follows: (1) Collect multimodal environmental perception information, perform time and space alignment on the collected modal environmental perception information, and identify obstacles in the current environment based on the multimodal environmental perception information; (2) Based on the Bayesian inference model and the fuzzy inference model, predict the probability P that the target obstacle in the current environment is a real obstacle, and identify the target obstacle with a large probability P as an obstacle.
2. The water surface environment perception method based on Bayesian-fuzzy fusion as described in claim 1, characterized in that, The process of obtaining probability P is as follows: (21) Obtain the observation vector of the target obstacle at the current time k. The observation vector at the current moment Input the Bayesian inference model, and the Bayesian inference model outputs the first probability that the target obstacle in the current environment is a real obstacle; (22) Collect the current water surface environment parameters, input the current water surface environment parameters and the first probability that the target obstacle belongs to the real obstacle into the fuzzy inference model, and the fuzzy inference model infers the second probability that the target obstacle belongs to the real obstacle; (23) The first probability and the second probability that the target obstacle belongs to the real obstacle are fused together to form the probability P that the target obstacle belongs to the real obstacle.
3. The water surface environment perception method based on Bayesian-fuzzy fusion as described in claim 2, characterized in that, Observation vector This includes: state observation of the target obstacle, spatial geometric observation, visual semantic observation, and the observation state of the ship itself. Among them, the state observation of the target obstacle includes the relative distance, relative azimuth, and Doppler radial velocity of the target obstacle; the spatial geometric observation of the target obstacle includes the center coordinates of the three-dimensional bounding box of the target obstacle, its length and width dimensions, and the depth extreme points of the underwater contour; the visual semantic observation of the target obstacle includes the pixel coordinates of the center of the two-dimensional detection box of the target obstacle; and the observation state of the ship itself includes the ship's heading, speed, and three-axis attitude angles.
4. The water surface environment perception method based on Bayesian-fuzzy fusion as described in claim 2, characterized in that, Water surface environmental parameters include: water surface reflection intensity, image edge continuity, radar echo stability, sonar profile integrity, visibility coefficient, and wave level parameters.
5. The water surface environment perception method based on Bayesian-fuzzy fusion as described in claim 1, characterized in that, The obstacle information of the obstacles in the current environment is stored in the record table. The obstacle information includes the obstacle category, obstacle identifier, and obstacle observation status. The obstacle observation status includes the obstacle state observation, spatial geometric observation status, and visual semantic observation status.
6. The water surface environment perception method based on Bayesian-fuzzy fusion as described in claim 1, characterized in that, The process of extracting target obstacles is as follows: Based on the obstacle information extracted from the previous frame, detect whether there are any newly added obstacles or unstable obstacles in the current frame. If the detection result is positive, then the newly added obstacles and unstable obstacles will be taken as target obstacles.
7. The water surface environment perception method based on Bayesian-fuzzy fusion as described in claim 6, characterized in that, Calculate the difference between the Doppler radial velocity of the obstacle in the current frame and the Doppler radial velocity of the obstacle in the previous frame. If the difference is greater than or equal to a set difference threshold, the state of the corresponding obstacle is considered unstable. If the difference is less than the set difference threshold, the state of the corresponding obstacle is considered stable.
8. The water surface environment perception method based on Bayesian-fuzzy fusion as described in claim 1, characterized in that, The process after step (2) also includes: The grid state of the target obstacle in the two-dimensional grid map is determined based on probability P. The grid state includes: occupied, empty, and unknown.
9. The water surface environment perception method based on Bayesian-fuzzy fusion as described in claim 8, characterized in that, The process of marking the grid state is as follows: If the probability P is greater than the first probability threshold, the corresponding target obstacle is determined to be a real obstacle, and the grid marker of the corresponding target obstacle is marked as occupied. If the probability is less than the second probability threshold, the corresponding target obstacle is determined not to be an obstacle, and the grid marker of the corresponding target obstacle is marked as idle. The grid to which the target obstacle belongs, which is between the first probability threshold and the second probability threshold, is marked as unknown. The first probability threshold is greater than the second probability threshold.