A water surface dynamic obstacle trajectory prediction method based on multi-source data fusion

Abstract

This invention discloses a method for predicting the trajectory of dynamic obstacles on the water surface based on multi-source data fusion, belonging to the field of water surface target perception and motion prediction technology based on radio waves. The method first acquires radar observations and environmental field data, then constructs spatiotemporally consistent synchronous observation data and observation confidence levels through spatiotemporal registration and observation quality assessment. Next, it obtains estimated obstacle motion states and wind-driven drift coefficients through dynamic constraint state estimation. After motion state identification and prediction model switching, it outputs motion state labels and switches prediction models accordingly. Finally, it generates a sequence of predicted trajectory points with confidence intervals based on predicted trajectories and confidence intervals, and completes risk assessment and early warning output through risk grading and early warning output. This invention incorporates the physical constraints of floating object drift and the spatiotemporal fusion of multi-source heterogeneous data into a unified prediction framework, applicable to the prediction of future trajectories and the quantification of uncertainties of floating objects on flowing water surfaces under complex environmental conditions.

Description

Technical Field

[0001] This invention relates to the field of water surface target perception and motion prediction technology based on wireless waves, specifically a method for predicting the trajectory of dynamic obstacles on the water surface based on multi-source data fusion. Background Technology

[0002] Dynamic obstacles on the water surface pose a direct threat to the navigation of surface robots, the safety of waterway navigation, and water environment cleanup operations. Specifically, the dynamic obstacles on the water surface referred to in this invention are floating objects on flowing water, that is, floating objects that drift with the water flow and wind.

[0003] Existing methods for analyzing the motion of floating objects on the water surface mainly rely on processing water surface image sequences using a monocular camera: detecting the changing features of water surface ripple texture, performing a lookup match between the ripple features and a pre-set empirical motion state table, and outputting the current motion direction and speed level of the floating object.

[0004] The above scheme has the following limitations: First, the lookup table matching only outputs the qualitative motion state at the current moment and cannot provide a position prediction for future periods. Downstream obstacle avoidance and salvage tasks can only respond passively based on the current state. Second, the water surface ripple features only reflect the local water surface morphology and are not fused with surface flow and wind field data. Under low light, large waves, rain, and fog conditions, the ripple features degrade, and the output results become distorted. Third, once the floating object leaves the camera's field of view or is obscured by waves or debris, the scheme loses the target and cannot continue to provide the target's location. Fourth, the output results do not include uncertainty measurement, and downstream tasks cannot assess the reliability of the results. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for predicting the trajectory of dynamic obstacles on the water surface based on multi-source data fusion, which solves the problem of insufficient trajectory prediction in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the trajectory of dynamic obstacles on the water surface based on multi-source data fusion, the method comprising:

[0007] S1: First, the radar observation and acquisition unit continuously transmits radio waves to the monitored water area and receives the echo signals reflected by the targets in the water area. Obstacle detection and target association are performed on the echo signals to obtain obstacle location observation data. Then, the environmental field acquisition unit obtains surface flow data and wind field data.

[0008] S2: First, the spatiotemporal alignment unit unifies the obstacle location observation data, surface flow data and wind field data to the same sampling time to obtain synchronous observation data. Then, the spatial registration unit converts the obstacle location in the synchronous observation data from radar range and azimuth coordinates to geographic coordinates and removes invalid observations. Finally, the observation confidence evaluation unit calculates the radar observation confidence based on the signal quality index.

[0009] S3: Construct a floating motion mechanics model through the drift motion mechanics modeling unit, then obtain the obstacle motion state estimate and the filter residual sequence through the filter state estimation unit, and finally use the wind drag coefficient adaptive estimation unit to perform online estimation and correction of the wind drag drift coefficient using the filter residual sequence and echo intensity index to obtain the wind drag drift coefficient estimate.

[0010] S4: The state feature quantity construction unit performs sliding window statistics on the obstacle motion state estimate and the filtered residual sequence. Then, the state logic discrimination unit performs logical discrimination on the three feature quantities in a progressive order of first mutation discrimination, then static discrimination, and finally position verification. The motion state label is output as one of the following: free drift state, stranded static state, or collision disturbance mutation state. Finally, the prediction model switching unit switches the prediction model according to the motion state label.

[0011] S5: The loss determination unit performs a cumulative comparison with a threshold on the continuous unobserved count of each registered obstacle to determine whether the registered obstacle has entered the observation loss state. When it is determined that the observation loss state has been entered, the dynamic deduction unit performs dynamic deduction to obtain the deduced position sequence. When the registered obstacle is detected again, the re-observation correction unit performs correction on the deduced position sequence with the new obstacle position observation data.

[0012] S6: The prediction unit driven by the environmental forecast uses the prediction model to gradually extrapolate the surface flow forecast data and wind field forecast data in the future period to obtain the predicted trajectory point sequence. Then, the uncertainty quantification unit performs multiple parameter perturbation extrapolations on the predicted trajectory point sequence, statistically analyzes the location distribution of each predicted trajectory point, and generates the confidence interval of each predicted trajectory point.

[0013] S7: The drift destination analysis unit performs spatial intersection discrimination on the predicted trajectory point sequence with the confidence interval, shoreline data, channel range data, protected area range data, and water depth and topography data, and outputs the drift destination marker. Then, the risk level generation unit generates a risk level based on the drift destination marker, entry time, and confidence interval width, and encapsulates the risk level, drift destination marker, entry time, predicted trajectory point sequence, and confidence interval into a risk warning message for output.

[0014] This invention provides a method for predicting the trajectory of dynamic obstacles on the water surface based on multi-source data fusion. It has the following beneficial effects:

[0015] (1) This invention integrates obstacle location observation data obtained by radar observation with surface flow data and wind field data into the same spatiotemporal reference through spatiotemporal alignment and spatial registration to form synchronous observation data. The reliability of each data source is quantified by the observation confidence assessment, thereby determining the values ​​of observation noise and process noise in adaptive Kalman filtering. This overcomes the shortcomings of existing technologies that rely solely on the ripple features of a single water surface image and suffer from feature degradation and output distortion under low light and high wave conditions. It realizes the complementary fusion of multi-source heterogeneous data within the state estimation framework. Radar radio wave observation is not limited by lighting conditions and has strong adaptability to low visibility conditions such as night, rain, fog, and sea fog.

[0016] (2) This invention uses the floating motion dynamics model composed of surface flow drift term and wind-induced drift term as the filter state equation. The wind-induced drift coefficient of each registered obstacle is estimated online by residual-driven identification and echo intensity index correction. The motion state label is identified in the progressive order of first sudden change discrimination, then static discrimination, and finally position verification, and the prediction model is adaptively switched. This overcomes the defect of the existing table lookup scheme that can only output the qualitative motion state at the current moment and cannot predict the future position. It realizes the prediction of the future trajectory of dynamic obstacles on the water surface under the constraints of physical laws.

[0017] (3) Under the condition of observation loss, the present invention continuously extrapolates the target position by combining the estimated value of wind drift coefficient with surface flow forecast data and wind field forecast data, and corrects the extrapolated trajectory and restores the filter when re-detecting, which solves the robustness defect of the prior art that the target is lost and the scheme fails once the obstacle is out of the field of vision or is blocked; at the same time, multiple parameter perturbation extrapolation is performed on the predicted trajectory points to generate confidence intervals, and the uncertainty of the prediction is output along with the risk warning message, which provides a reliable basis for the local obstacle avoidance of downstream water surface robots and global monitoring of shore-based platforms, as well as the pre-allocation of multiple salvage robot tasks. Attached Figure Description

[0018] Figure 1 Flowchart of the overall solution of this invention;

[0019] Figure 2 Flowchart of the spatiotemporal registration and observation quality assessment of this invention;

[0020] Figure 3 Flowchart of the filter state estimation and wind drag coefficient estimation of this invention;

[0021] Figure 4 Flowchart of motion state recognition of the present invention;

[0022] Figure 5 Flowchart of the observation loss inference and completion process of this invention;

[0023] Figure 6 Flowchart of the predicted trajectory and confidence interval generation of this invention;

[0024] Figure 7 The risk classification flowchart of this invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1: River monitoring scenario, such as Figure 1-7 As shown, this embodiment of the invention provides a method for predicting the trajectory of dynamic obstacles on the water surface based on multi-source data fusion. The dynamic obstacles are specifically floating objects on the flowing water surface. Dynamic obstacles entering the tracking process are referred to as registered obstacles, and each registered obstacle is assigned a unique target number. The method is deployed on an embedded computing platform of a shore-based monitoring station or an unmanned patrol boat. The method includes:

[0027] S1: First, the radar observation and acquisition unit continuously transmits radio waves to the monitored water area and receives the echo signals reflected by the targets in the water area. Obstacle detection and target association are performed on the echo signals to obtain obstacle location observation data. Then, the environmental field acquisition unit obtains surface flow data and wind field data.

[0028] The radar observation and acquisition process is as follows: The radar observation and acquisition unit includes a marine navigation radar and a signal processing program running on an embedded computing platform. For each scan, the echo signal is first processed to suppress the echo background using a sea clutter suppression method. Then, a constant false alarm rate (CFAR) detection method is used to extract candidate echo regions from the range and azimuth unit. For each candidate echo region, its range, azimuth, and centroid range and azimuth coordinates are calculated. At the same time, the echo intensity index of the region is calculated. The echo intensity index is the amount obtained after normalizing the radar echo power of the candidate echo region, and its value range is a closed interval [0,1]. Subsequently, the Hungarian algorithm is used to perform target association between candidate echo regions between adjacent scans. The inter-frame position displacement is used as the association cost. When the association cost is less than the association threshold, the current candidate echo region and the existing target are determined to be the same dynamic obstacle on the water surface and a unique target number is assigned. Each scan, the radar observation and acquisition unit outputs a set of obstacle position observation data. Each set of obstacle position observation data includes a sampling timestamp, target number, centroid range and azimuth coordinates, and echo intensity index.

[0029] In this embodiment, it should be specifically noted that the marine navigation radar is fixedly installed at a shore-based monitoring station or an unmanned patrol boat, operates in the X-band, has an antenna rotation period of two seconds, a range resolution of five meters, a maximum range of twelve nautical miles, and its beam covers the entire monitored water area. The area threshold is set according to the number of range azimuth cells corresponding to the smallest detectable target, and the correlation threshold is set to thirty meters. The sea clutter suppression method, the constant false alarm rate detection method, and the Hungarian algorithm are all conventional technical means, and this embodiment does not impose specific limitations on them.

[0030] The environmental field acquisition is as follows: The environmental field acquisition unit includes a measurement equipment interface and a forecast data interface. The measurement equipment interface is connected to an acoustic Doppler current profiler and an ultrasonic anemometer deployed in the monitored water area. The acoustic Doppler current profiler outputs a surface flow velocity vector every five minutes. The environmental field acquisition unit performs format parsing and timestamp annotation on the received data to obtain surface flow data and wind field data. The forecast data interface receives surface flow forecast data and wind field forecast data for the monitored water area issued by hydrological forecasting departments and meteorological forecasting departments.

[0031] In this embodiment, it should be specifically noted that the surface flow velocity vector includes an eastward component and a northward component, with the unit being meters per second. The ultrasonic anemometer outputs the wind speed and direction at a height of 10 meters above the water surface once per minute. The forecast data is given in grid form with a spatial grid spacing of 50 meters, a time resolution of one minute, and a forecast lead time of 30 minutes.

[0032] S2: First, the spatiotemporal alignment unit unifies the obstacle location observation data, surface flow data, and wind field data to the same sampling time to obtain synchronous observation data. Then, the spatial registration unit converts the obstacle locations in the synchronous observation data from radar range-azimuth coordinates to geographic coordinates and removes invalid observations. Finally, the observation confidence assessment unit calculates the radar observation confidence based on signal quality indicators and the environmental data confidence based on the timeliness of the data source. The registration and assessment process for this step is described in [link to documentation]. Figure 2 ;

[0033] The spatiotemporal alignment is as follows: the sampling periods for obstacle location observation data, wind field data, and surface flow data are different, namely two seconds, one minute, and five minutes, respectively. The spatiotemporal alignment unit sets a unified sampling period of two seconds and first performs clock synchronization on each device using the network time protocol to ensure that the time base of each data source is consistent. For each unified sampling moment, the spatiotemporal alignment unit performs linear interpolation on the surface flow data and wind field data in the time domain. Using two adjacent measured samples before and after that moment as interpolation nodes, the surface flow velocity vector, wind speed, and wind direction at that moment are calculated by time-weighted average. After alignment is completed, the spatiotemporal alignment unit outputs synchronized observation data. The synchronized observation data at each unified sampling moment includes the obstacle centroid distance and azimuth coordinates, surface flow velocity vector, wind speed, and wind direction.

[0034] In addition, when the interval between two adjacent measured samples from the same data source exceeds the preset interpolation window, the spatiotemporal alignment unit marks the moment as missing environmental data, and the filtering state estimation unit uses the most recent valid surface flow data and wind field data to perform state prediction for the moment marked as missing environmental data. In this embodiment, the interpolation window is fifteen minutes.

[0035] The spatial registration is as follows: The spatial registration unit first reads the radar antenna installation position coordinates obtained by the positioning device and the antenna azimuth obtained by the azimuth reference system. Then, it converts the centroid distance-azimuth coordinates of each uniform sampling time in the synchronous observation data into geographic coordinates according to the distance-azimuth relationship. A local rectangular coordinate system is established with the southwest corner of the monitored water area as the origin, the east as the x-axis, and the north as the y-axis, with the unit being meters. After the conversion is completed, the spatial registration unit performs validity verification on each set of geographic coordinates. When the geographic coordinates fall outside the monitored water area, or when the geographic coordinates fall on the land side according to the shoreline data, the corresponding position observation is marked as invalid and removed from subsequent processing.

[0036] In this embodiment, it should be specifically noted that the installation location coordinates are taken from a fixed point on the shoreline where the geographic coordinates can be measured. When the radar observation and acquisition unit is installed on a mobile unmanned patrol boat, the spatial registration unit also reads the pitch angle, roll angle and heading angle measured by the inertial measurement unit, performs compensation on the platform attitude and then performs coordinate transformation.

[0037] The observation confidence assessment is as follows: The observation confidence assessment unit calculates two types of confidence: radar observation confidence and environmental data confidence. For radar observation confidence, the observation confidence assessment unit extracts three signal quality indicators from the currently scanned echo signal: The first is the signal-to-noise ratio (SNR) indicator, which is the SNR of the candidate echo region divided by a preset SNR benchmark value and normalized; if it exceeds one, it is truncated to one. The second is the echo stability indicator, which is the normalized fluctuation of the echo intensity in the candidate echo region within the observation window, minus the normalized value; the greater the fluctuation, the lower the value. The third is the interference level indicator, which is the normalized residual interference level within the current scanning range, minus the normalized value; the stronger the interference, the lower the value. The radar observation confidence score is calculated using the weighted sum of three normalized indices: signal-to-noise ratio (SNR) 0.4, echo stability 0.3, and interference level 0.3. The radar observation confidence score is calculated within the closed interval [0,1]. For environmental data confidence, the observation confidence assessment unit determines the confidence score based on the source and timeliness of the surface flow and wind field data: when both surface flow and wind field data originate from actual measurement equipment and the timeliness does not exceed ten minutes, the environmental data confidence score is 1.0; when both surface flow and wind field data originate from numerical weather prediction products, the environmental data confidence score is categorized according to forecast timeliness: 0.9 for forecast timeliness not exceeding ten minutes, 0.7 for forecast timeliness exceeding ten minutes but not exceeding twenty minutes, and 0.5 for forecast timeliness exceeding twenty minutes.

[0038] It should be specifically noted in this embodiment that under sea clutter interference or heavy rainfall conditions, the signal-to-noise ratio index decreases, and the confidence of radar observation decreases significantly. This change will be transmitted to the fusion weights via the observation noise covariance in step S3.

[0039] S3: First, a floating motion mechanics model is constructed using the drift motion mechanics modeling unit. Then, an adaptive Kalman filter is performed using the floating motion mechanics model as the state equation by the filtering state estimation unit to obtain the obstacle motion state estimate and the filtering residual sequence. Finally, the wind drag coefficient adaptive estimation unit uses the filtering residual sequence and echo intensity index to perform online estimation and correction of the wind drag drift coefficient to obtain the wind drag drift coefficient estimate. The filtering and coefficient estimation process for this step is described in [link to relevant documentation]. Figure 3 ;

[0040] The drift kinetic modeling is as follows: The drift kinetic modeling unit treats dynamic obstacles on the water surface as passive floating objects. The drift velocity of the passive floating object consists of two parts: surface current drift term and wind-induced drift term. The drift kinetic modeling unit constructs a drift kinetic model of the floating object, and its velocity expression is:

[0041] ;

[0042] Among them, u cγ is the surface flow velocity vector obtained by bilinear interpolation of surface flow data or surface flow forecast data at the obstacle location. w Let |u be the dimensionless wind drift coefficient. a | represents the magnitude of the wind speed vector in the wind field data, e w Let be the unit vector representing the direction of wind-induced drift, which is the direction after deflecting the downstream direction by a preset deflection angle. Based on this, the drift kinetic modeling unit constructs the state equations of the floating object drift kinetic model. The state variables include eastward coordinates, northward coordinates, eastward velocity components, and northward velocity components. The state equations consist of two parts: position equations and velocity equations. The position equation states that the coordinates at the next moment are equal to the sum of the products of the coordinates at the current moment, the corresponding velocity component at the current moment, and the filtering step size. The velocity equation adopts a relaxed form, and its expression is:

[0043] ;

[0044] Among them, v k Let v be the velocity component at the current moment. th,k τ represents the theoretical drift velocity component at the current moment, Δt is the filtering step size, and τ is the theoretical drift velocity component at the current moment. v The relaxation time constant;

[0045] In this embodiment, it should be specifically noted that the preset deflection angle is 15 degrees, the initial value of the wind drift coefficient is 0.03, and the value range is limited to the closed interval [0.01, 0.06]. The relaxation time constant is 30 seconds, and the filtering step size is 2 seconds. The relaxation form makes the velocity state asymptotically converge to the theoretical drift velocity under the constraint of environmental data, avoiding the rigid assignment of velocity.

[0046] The filtered state estimation is as follows: The filtered state estimation unit performs state estimation using an adaptive Kalman filter algorithm. The observations are taken from the geographic coordinates in the synchronous observation data. The observation matrix is ​​used to select the position components from the state variables as the observations. The state equation adopts the floating motion mechanics model. The observation noise covariance is the baseline observation noise covariance multiplied by an amplification factor negatively correlated with the radar observation confidence. The amplification factor increases as the radar observation confidence decreases. The baseline observation noise covariance is set according to the radar range and azimuth measurement error. The process noise covariance is the baseline process noise covariance multiplied by an amplification factor negatively correlated with the environmental data confidence. The filtered state estimation unit performs two steps of prediction and update at each uniform sampling time: First, it performs prediction on the state variables and covariance according to the state equation. Then, if there is a valid position observation at the current time, it calculates the Kalman gain and performs observation update. If the position observation is missing or marked as invalid at the current time, it only performs prediction. After each update, the filtered state estimation unit records the difference between the observed value and the predicted value of the position component as the filtering residual. The filtering residuals at all times constitute the filtering residual sequence.

[0047] In this embodiment, it should be specifically noted that the reference observation noise covariance is a diagonal matrix with diagonal elements of one square meter, and the reference process noise covariance is a diagonal matrix with diagonal elements of 0.01. When the confidence of radar observation decreases, the observation noise covariance increases. In the state update, the filter reduces the fusion weight of radar observation and increases the fusion weight of the floating motion mechanics model driven by surface flow data and wind field data. The fusion of obstacle position observation data with surface flow data and wind field data is completed within the filtering framework.

[0048] In this embodiment, it should be further explained that the filter initialization uses the first three sets of effective position observations: the centroid of the three sets of position observations is used as the initial position, and the displacement difference between adjacent position observations divided by the time interval is used as the initial velocity; the filter state estimation unit outputs the obstacle motion state estimation value and the filter residual sequence, and the obstacle motion state estimation value includes the filter position and the filter velocity.

[0049] The adaptive estimation of the wind drag coefficient is as follows: Different dynamic obstacles on the water surface have varying shapes, masses, and sizes above water, resulting in different wind drag coefficients. Using a fixed coefficient would introduce systematic bias. The adaptive estimation unit performs online estimation and correction of the wind drag coefficient through two paths: residual-driven identification and echo intensity index correction. The residual-driven identification path reads the filtered residual sequence output by the filtered state estimation unit, projects each filtered residual onto a unit vector in the wind-induced drift direction to obtain the projected residual, and then performs regression on the projected residual sequence using a recursive least squares method with a forgetting factor of 0.98 to obtain the wind-induced drift direction. The system bias estimate is then used to update the wind drift coefficient according to the update formula, which is the sum of the current wind drift coefficient and the product of the updated gain and the system bias estimate. The echo intensity index correction path reads the echo intensity index carried by the obstacle position observation data and corrects the wind drift coefficient according to the correction formula, which is the sum of the wind drift coefficient and 0.5 plus the echo intensity index. After the two paths are completed, the wind drift coefficient adaptive estimation unit limits the wind drift coefficient to the closed interval [0.01, 0.06] to obtain the wind drift coefficient estimate.

[0050] In this embodiment, it should be specifically noted that the update gain is 0.5, which is a dimensionless damping factor used to suppress overcorrection caused by observation noise; the identification window duration is 60 seconds, the same as the update cycle; and the lower limit threshold for wind speed is 2 meters per second. The update of the residual-driven identification path is subject to two preconditions: first, the radar observation confidence is higher than 0.6; second, the current motion state label is in a free drift state. If the preconditions are not met, the update is paused, and the current wind drift coefficient estimate is maintained. The stronger the echo intensity, the larger the target size and the larger the water surface area, the larger the wind-receiving area, and the stronger the wind effect. The update cycle of the wind drift coefficient adaptive estimation unit is 60 seconds, and the wind drift coefficient estimate is stored according to the target number for subsequent prediction and observation loss prediction of the registered obstacle.

[0051] S4: First, the state feature quantity construction unit performs sliding window statistics on the obstacle motion state estimate and the filtered residual sequence. Then, the state logic discrimination unit performs logical discrimination on the three features in a progressive order of first mutation discrimination, then static discrimination, and finally position verification, outputting a motion state label of one of the following: free drift state, stranded static state, or collision disturbance mutation state. Finally, the prediction model switching unit switches the prediction model according to the motion state label and transmits the motion state label and the switched prediction model to step S6. The discrimination process of this step is as follows: Figure 4 ;

[0052] The state feature quantity is constructed as follows: The state feature quantity construction unit performs sliding window statistics on the obstacle motion state estimate and the filtered residual sequence to construct three features. The first is the velocity statistical feature quantity, which includes the mean velocity and standard deviation of the filtered velocity sequence within the window; the second is the position-shoreline distance feature quantity, which takes the shortest distance from the current filtering position to the shoreline polyline. The shoreline data is stored in polyline form. The state feature quantity construction unit calculates the point-to-segment distance from the filtering position to each shoreline segment and takes the minimum value, in meters; the third is the residual cumulative sum feature quantity, which is calculated using the cumulative sum method, and its recursive formula is:

[0053] ;

[0054] Among them, s k s is the cumulative sum variable at the current moment. k-1 r is the cumulative sum variable from the previous time step. k Let ε be the filtered residual at the current time. r The residual cumulative sum characteristic is taken as the value of the cumulative sum variable at the current time, and the initial value of the cumulative sum variable is zero;

[0055] In this embodiment, it should be specifically noted that the window length is 30 seconds, corresponding to 15 samples at a sampling interval of 2 seconds; the residual allowable deviation is twice the standard deviation of the observation noise; the residual cumulative sum and characteristic quantity are sensitive to persistent small deviations and are used to capture state changes caused by collision disturbance events.

[0056] The state logic discrimination is as follows: the state logic discrimination unit performs progressive discrimination on the motion state of dynamic obstacles on the water surface in a fixed order of first mutation discrimination, then static discrimination, and finally position verification. The first step is mutation discrimination, which determines whether the residual accumulation and feature quantity exceed the mutation threshold, and whether the velocity change within the window exceeds the velocity change threshold. If both conditions are met, the motion state label is output as collision disturbance mutation state, and the current discrimination ends. If either condition is not met, the second step is entered. The second step is static discrimination, which determines whether the velocity standard deviation is lower than the stranding velocity threshold and the velocity mean is lower than the grounding velocity threshold. If the speed is below the grounding speed threshold, the motion status label is output as free drift if the judgment is negative, and proceeds to the third step if the judgment is positive. The third step is position verification, which reads the position-shoreline distance feature and water depth topography data, and determines whether the position-shoreline distance feature is less than the shoreline buffer distance and whether the water depth at the current filtering position is less than one meter. If both conditions are met, the motion status label is output as grounded and stationary. If the conditions are not met, the stationary time is accumulated. If the stationary time exceeds 120 seconds, the motion status label is still output as grounded and stationary. Otherwise, the motion status label of the previous moment is maintained.

[0057] In this embodiment, it should be specifically noted that the mutation threshold is five times the allowable deviation of the residual, the velocity change threshold is 0.3 meters per second, the grounding velocity threshold is 0.05 meters per second, and the shoreline buffer distance is three meters. The reason for placing mutation discrimination first is that collision disturbance events cause abrupt changes in velocity and filter residuals, while stationary and grounding are gradual processes. Performing mutation discrimination first can prevent mutation events from being masked by gradual discrimination logic.

[0058] It should be further explained in this embodiment that the state logic discrimination unit is also equipped with a hysteresis mechanism: the motion state label is only updated when the discrimination result remains unchanged for five consecutive seconds, and the motion state label of the previous moment is maintained when the discrimination result changes.

[0059] The prediction model switching is as follows: The prediction model switching unit maintains three prediction models: free drift prediction model, stranded and stationary prediction model, and disturbance transition prediction model. When the motion state label is free drift, it switches to the free drift prediction model, and the position is deduced according to the velocity expression and position equation of the floating object drift kinetic model. The values ​​of surface flow and wind are replaced by surface flow data and wind field data for future time-based surface flow forecast data and wind field forecast data. When the motion state label is stranded and stationary, it switches to the stranded and stationary prediction model, and the predicted position remains unchanged at the current filtered position. When the motion state label is collision disturbance abrupt change, it switches to the disturbance transition prediction model. The predicted velocity decays from the current filtered velocity to the theoretical drift velocity. The predicted velocity during the transition period is obtained by adding two parts: the first part is the current filtered velocity multiplied by the power of the deduction step number of the decay factor, and the second part is the theoretical drift velocity multiplied by one and the power of the deduction step number of the decay factor. The theoretical drift velocity is the drift velocity calculated based on the surface flow forecast data and wind field forecast data.

[0060] In this embodiment, it should be specifically noted that the attenuation factor is 0.9, the transition period is 60 seconds, and the model automatically switches to the free drift prediction model after the transition period ends. When switching prediction models, the prediction model switching unit uses the current filtering state as the initial value for the new model to maintain the continuity of the trajectory.

[0061] S5: This step works in parallel with step S3. The loss determination unit performs a cumulative count and threshold comparison on the continuous unobserved count of each registered obstacle to determine whether the registered obstacle has entered the observation loss state. When the observation loss state is determined, the dynamics deduction unit performs dynamics deduction to obtain the deduced position sequence. When the registered obstacle is detected again, the re-observation correction unit corrects the deduced position sequence with the new obstacle position observation data, and the correction result is sent back to step S3 to restore filtering. At the same time, the deduced position sequence is transmitted to step S6 to maintain trajectory continuity. The deduction and correction process of this step is as follows: Figure 5 ;

[0062] The loss determination is as follows: the loss determination unit maintains a continuous unobserved count for each registered obstacle. In each sampling period, when the radar observation and acquisition unit does not detect or associate with the registered obstacle, the loss determination unit increments the continuous unobserved count of the registered obstacle by one. When the radar observation and acquisition unit detects and successfully associates with the obstacle, the loss determination unit clears the continuous unobserved count to zero. When the continuous unobserved count exceeds the loss determination threshold, the loss determination unit determines that the registered obstacle has entered the observation loss state and notifies the dynamics deduction unit.

[0063] In this embodiment, it should be specifically noted that the loss determination threshold is adaptively set based on the radar observation confidence level: when the radar observation confidence level is higher than 0.3, the loss determination threshold is set to five; when the radar observation confidence level is lower than 0.3, the detection capability decreases, and the loss determination threshold is increased to fifteen to avoid false loss caused by sea clutter interference and increased missed detection rate under rain and snow conditions; the loss determination ensures that dynamic obstacles on the water surface are not directly discarded during the period when they are briefly covered by sea clutter or interference, but are tracked by subsequent inference;

[0064] The dynamic extrapolation is as follows: The dynamic extrapolation unit receives the registered obstacle that has entered the observation loss state, reads the last output of the obstacle motion state estimate, the wind drift coefficient estimate of the registered obstacle stored according to the target number, and the surface flow forecast data and wind field forecast data. Taking the last filtered position as the extrapolation starting point and two seconds as the extrapolation step size, the unit performs step-by-step extrapolation for each future moment according to the velocity expression and position equation of the floating object drift dynamics model to obtain the extrapolated position sequence.

[0065] It should be further explained in this embodiment that each simulation position is marked with a simulation mark during the simulation process. The simulation position is not used as an observation to participate in the filtering of other registered obstacles. The dynamic simulation unit is set to a maximum simulation time of 300 seconds. If the registered obstacle is not detected again after the simulation time reaches 300 seconds, the dynamic simulation unit cancels the registered obstacle, terminates the simulation and outputs a trajectory termination mark.

[0066] The re-observation correction is as follows: When the radar observation and acquisition unit re-detects a dynamic obstacle on the water surface, the re-observation correction unit first performs association confirmation, calculates the distance between the newly detected position and the current inferred position of the registered obstacle, and confirms that the newly detected target and the registered obstacle in the inferred state are the same dynamic obstacle on the water surface when the distance is less than the association confirmation threshold. After successful confirmation, the re-observation correction unit uses the new obstacle position observation data as the observation to perform filtering update on the inferred position, re-initializes the state covariance of the adaptive Kalman filter to restore the sensitivity of the filter to the observation, and returns the registered obstacle to the normal filtering process in step S3.

[0067] In this embodiment, it should be specifically noted that the correlation confirmation threshold is set to ten meters; for trajectory jumps that occur at the junction between the deduced position and the correction position, the observation and correction unit performs weighted smoothing on the trajectory points in the deduced interval using a backward sliding window with a window length of five seconds to maintain trajectory continuity.

[0068] S6: First, the environmental forecast-driven prediction unit uses a prediction model to progressively extrapolate the surface flow forecast data and wind field forecast data for future periods to obtain a sequence of predicted trajectory points. Then, the uncertainty quantification unit performs multiple parameter perturbation extrapolations on the predicted trajectory point sequence, statistically analyzes the location distribution of each predicted trajectory point, and generates a confidence interval for each predicted trajectory point. The prediction and quantification process for this step is described in [link to documentation]. Figure 6 ;

[0069] The environmental forecast-driven prediction is as follows: The environmental forecast-driven prediction unit receives the motion state tag and prediction model identifier, and simultaneously reads the current obstacle motion state estimate and wind-induced drift coefficient estimate of the registered obstacle, as well as surface flow forecast data and wind field forecast data. The deduction is performed according to the type of prediction model: Under the free drift prediction model, the theoretical drift velocity at the next moment is calculated according to the velocity expression, where the surface flow term is obtained by bilinear interpolation from the surface flow forecast data at the corresponding future moment, and the wind-induced drift term uses the wind-induced drift coefficient estimate of the registered obstacle. The position is then updated according to the position equation, and the predicted trajectory point sequence is obtained after iterative execution; Under the stranded and stationary prediction model, the predicted trajectory points with the same position as the current filtered position are output; Under the disturbance transition prediction model, the deduction is performed according to the attenuation transition method set in step S4, and the deduction continues according to the free drift prediction model after the transition period ends.

[0070] In this embodiment, it should be specifically noted that the forecast lead time is 30 minutes, the prediction duration is set to 15 minutes, the extrapolation step size is 30 seconds, and the total number of predicted trajectory points is 30.

[0071] The uncertainty quantification is as follows: the uncertainty quantification unit performs quantification on the uncertainty of the predicted trajectory point sequence by using parameter perturbation set extrapolation. One hundred sampling extrapolations are performed in each prediction period. Each sampling first performs perturbation sampling on three types of parameters: the first type is the initial velocity, which is sampled according to the filtered velocity covariance output by the filtered state estimation unit; the second type is the wind drift coefficient, which is uniformly sampled within the positive and negative 20% range of the estimated wind drift coefficient; the third type is the environmental forecast velocity, where the perturbation amplitude is determined according to the confidence level of the environmental data, with a larger perturbation amplitude for lower environmental data confidence. Then, using the sampled parameters as input, a complete extrapolation process of the environmental forecast-driven prediction unit is executed to obtain a sample trajectory. After collecting one hundred sample trajectories, for each predicted trajectory point, a sample set is constructed using the positions of the one hundred sample trajectories at that point. The 5% and 95% quantiles of the east and north coordinates are taken respectively to form the 90% confidence interval for that predicted trajectory point.

[0072] In this embodiment, it should be specifically noted that the disturbance amplitude of the environmental forecast flow velocity is taken as the product of 1.5 minus the difference in the confidence level of the environmental data and 0.1, in meters per second; the uncertainty quantification unit outputs the predicted trajectory point sequence and the confidence interval of each predicted trajectory point. The width of the confidence interval increases with the extension of the prediction time, and the downstream module directly reads the uncertainty level of the prediction based on the width of the confidence interval.

[0073] S7: First, the drift destination analysis unit performs spatial intersection discrimination on the predicted trajectory point sequence with confidence intervals, shoreline data, channel range data, protected area range data, and water depth and topography data, outputting drift destination markers. Then, the risk level generation unit generates a risk level based on the drift destination markers, entry time, and confidence interval width, and encapsulates the risk level, drift destination markers, entry time, predicted trajectory point sequence, and confidence interval into a risk warning message for output. The grading process for this step is described in [link to relevant documentation]. Figure 7 ;

[0074] The drift destination analysis is as follows: The drift destination analysis unit reads the predicted trajectory point sequence and the confidence interval of each predicted trajectory point, and simultaneously reads the shoreline data, channel range data, protected area range data, and water depth and topography data of the monitored water area. The channel range data and protected area range data are stored in polygon form. The drift destination analysis unit performs spatial intersection discrimination in the following order: The first step is channel entry discrimination. The predicted trajectory points are traversed in chronological order. When there is a predicted trajectory point whose confidence interval overlaps with the channel range polygon, the drift destination is marked as channel entry, and the time corresponding to the first predicted trajectory point that meets the condition is marked as entry time. The second step is grounding and landing discrimination. When there is a predicted trajectory point located within the shoreline buffer zone and the water depth at that location is less than one meter, the drift destination is marked as grounding and landing. The third step is protected area entry discrimination. When there is a predicted trajectory point whose confidence interval overlaps with the protected area polygon, the drift destination is marked as protected area entry. If none of the three discriminations are met, the drift destination is marked as continuous drifting.

[0075] In this embodiment, it should be specifically noted that the shoreline buffer zone is the area within three meters of the shoreline; when multiple judgment results are established at the same time, they are selected according to priority: entering the waterway takes priority over running aground and landing on shore, and running aground and landing on shore takes priority over entering the protected area.

[0076] The risk level is generated as follows: The risk level generation unit generates a risk level based on the drift destination marker, entry time, and confidence interval width. The risk level is divided into three levels: Level 1 risk, Level 2 risk, and Level 3 risk. Level 1 risk is defined as when the drift destination marker indicates entry into the channel and the entry time is no more than 15 minutes from the current time. Level 2 risk is defined as when the drift destination marker indicates entry into the channel and the entry time is more than 15 minutes from the current time, or when the drift destination marker indicates entry into the protected area. Level 3 risk is defined as when the drift destination marker indicates grounding or continuous drifting. Based on this, the risk level generation unit performs confidence interval correction: when the confidence interval width of the predicted trajectory point corresponding to the drift destination marker exceeds the width threshold, the prediction uncertainty is relatively large, and the risk level is increased by one level according to the conservative principle. Level 1 risk remains unchanged and is not increased further.

[0077] In this embodiment, it should be specifically noted that the width threshold is 30 meters; risk warning messages are sent to the surface robot for it to perform local obstacle avoidance path planning, and at the same time to the shore-based platform for it to perform global monitoring and display and pre-allocation of tasks to multiple salvage robots; risk classification and warning output are executed once every 60 seconds for each registered obstacle;

[0078] Steps S1 to S7 are executed cyclically within each prediction update cycle, forming a closed loop of multi-source data acquisition, spatiotemporal registration, state estimation, motion state identification, observation loss inference and completion, trajectory prediction and risk warning.

[0079] Example 2: This embodiment of the invention provides a calculation example, where the data is exemplary constructed data used to illustrate the implementation process of the method. The prediction process is as follows: The monitored water area is 800 meters long and 300 meters wide. In step S1, the radar observation and acquisition unit detects a dynamic obstacle on the water surface at a certain moment. The target number is T01. The geographic coordinates of the centroid distance azimuth coordinates after conversion by the spatial registration unit in step S2 are 120 meters east and 45 meters north. At the same time, the surface flow data are 0.4 meters per second east and 0.1 meters per second north, the wind speed is 5 meters per second, the wind direction is due east, and the radar observation confidence level is 0.82. In step S3, filtering is started with an initial value of 0.03 for the wind drift coefficient. The filtering converges within 30 seconds. The wind drift coefficient adaptive estimation unit corrects the estimated value of the wind drift coefficient of T01 to 0.041 based on the echo intensity index of 0.35. In step S4, the state logic... The discrimination unit determines that the cumulative residual and characteristic quantity do not exceed the mutation threshold, and the velocity standard deviation is higher than the grounding velocity threshold, outputting the motion state label as free drift state. In step S6, the environmental forecast-driven prediction unit outputs thirty predicted trajectory points. In step S7, the drift destination analysis unit determines that the predicted trajectory will enter the channel at the eighth minute, with a confidence interval width of twelve meters. The risk level generation unit outputs a level one risk. During execution, T01 was briefly obscured by sea clutter for ten seconds. In step S5, the loss determination unit determines that it has entered an observation loss state. The dynamics extrapolation unit continues extrapolation, and T01 is re-detected at the sixth second. The re-observation correction unit completes the correction with a 1.8-meter deviation between the extrapolated position and the re-detected position, maintaining trajectory continuity. This process demonstrates that each step is sequentially connected under real disturbances, and the data transfer between filtering, extrapolation completion, and correction remains consistent.

[0080] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A water surface dynamic obstacle trajectory prediction method based on multi-source data fusion, characterized in that, The method includes: S1: First, the radar observation and acquisition unit continuously transmits radio waves to the monitored water area and receives the echo signals reflected by the targets in the water area. Obstacle detection and target association are performed on the echo signals to obtain obstacle location observation data. Then, the environmental field acquisition unit obtains surface flow data and wind field data. S2: First, the spatiotemporal alignment unit unifies the obstacle location observation data, surface flow data and wind field data to the same sampling time to obtain synchronous observation data. Then, the spatial registration unit converts the obstacle location in the synchronous observation data from radar range and azimuth coordinates to geographic coordinates and removes invalid observations. Finally, the observation confidence evaluation unit calculates the radar observation confidence based on the signal quality index. S3: Construct a floating motion mechanics model through the drift motion mechanics modeling unit, then obtain the obstacle motion state estimate and the filter residual sequence through the filter state estimation unit, and finally use the wind drag coefficient adaptive estimation unit to perform online estimation and correction of the wind drag drift coefficient using the filter residual sequence and echo intensity index to obtain the wind drag drift coefficient estimate. S4: The state feature quantity construction unit performs sliding window statistics on the obstacle motion state estimate and the filtered residual sequence. Then, the state logic discrimination unit performs logical discrimination on the three feature quantities in a progressive order of first mutation discrimination, then static discrimination, and finally position verification, and outputs a motion state label of one of the following: free drift state, stranded static state, or collision disturbance mutation state. Finally, the prediction model switching unit switches the prediction model to free drift prediction model, stranded static prediction model, or disturbance transition prediction model according to the motion state label. S5: The loss determination unit performs a cumulative comparison with a threshold on the continuous unobserved count of each registered obstacle to determine whether the registered obstacle has entered the observation loss state. When it is determined that the observation loss state has been entered, the dynamic deduction unit performs dynamic deduction to obtain the deduced position sequence. When the registered obstacle is detected again, the re-observation correction unit performs correction on the deduced position sequence with the new obstacle position observation data. S6: The prediction unit driven by the environmental forecast uses the prediction model to gradually extrapolate the surface flow forecast data and wind field forecast data in the future period to obtain the predicted trajectory point sequence. Then, the uncertainty quantification unit performs multiple parameter perturbation extrapolations on the predicted trajectory point sequence, statistically analyzes the location distribution of each predicted trajectory point, and generates the confidence interval of each predicted trajectory point. S7: The drift destination analysis unit performs spatial intersection discrimination on the predicted trajectory point sequence with the confidence interval, shoreline data, channel range data, protected area range data, and water depth and topography data, and outputs the drift destination marker. Then, the risk level generation unit generates a risk level based on the drift destination marker, entry time, and confidence interval width, and encapsulates the risk level, drift destination marker, entry time, predicted trajectory point sequence, and confidence interval into a risk warning message for output.

2. The water surface dynamic obstacle trajectory prediction method based on multi-source data fusion according to claim 1, characterized in that, In S1, the obstacle position observation data also carries an echo intensity index, which is the amount obtained after normalizing the radar echo power of the candidate echo region; the target association uses the inter-frame position displacement as the association cost, and when the association cost is less than the association threshold, it is determined that the current candidate echo region and the existing target are the same dynamic obstacle on the water surface and a unique target number is assigned.

3. The water surface dynamic obstacle trajectory prediction method based on multi-source data fusion according to claim 1, characterized in that, In S2, the spatial registration unit reads the radar's installation location coordinates and the antenna's azimuth reference, and converts the distance and azimuth coordinates of the obstacle's position in the synchronous observation data to geographic coordinates according to the distance-azimuth relationship; Observations that fall outside the monitored water area or on land after conversion are marked as invalid and removed; when the interval between two adjacent measured samples of surface flow data and wind field data exceeds the preset interpolation window, the spatiotemporal alignment unit marks the moment as missing environmental data, and the filtering state estimation unit uses the most recent valid surface flow data and wind field data to perform state prediction for the moment marked as missing environmental data.

4. The water surface dynamic obstacle trajectory prediction method based on multi-source data fusion according to claim 2, characterized in that, In step S3, the wind drag coefficient adaptive estimation unit projects the filtered residual sequence onto the wind-induced drift direction to obtain the projection residual. It then performs regression on the projection residual using the recursive least squares method to obtain a system bias estimate and updates the wind drag drift coefficient based on the system bias estimate. The wind drag coefficient adaptive estimation unit also reads the echo intensity index carried by the obstacle position observation data, corrects the updated wind drag drift coefficient based on the echo intensity index, and limits the correction result to a preset coefficient range to obtain the wind drag drift coefficient estimate. The online estimation is only performed when the radar observation confidence level is higher than the confidence threshold and the motion state label is in a free drift state.

5. The water surface dynamic obstacle trajectory prediction method based on multi-source data fusion according to claim 1, characterized in that, In S2, the observation confidence assessment unit uses the weighted sum of the signal-to-noise ratio, echo stability, and interference level as the radar observation confidence, and determines the environmental data confidence based on the source and timeliness of the surface flow and wind field data. In S3, the filter state estimation unit determines the observation noise covariance using the radar observation confidence and the process noise covariance using the environmental data confidence. The lower the radar observation confidence, the larger the observation noise covariance, which reduces the fusion weight of radar observations in the state update by the adaptive Kalman filter.

6. The water surface dynamic obstacle trajectory prediction method based on multi-source data fusion according to claim 1, characterized in that, In S4, the state logic discrimination unit performs discrimination in a progressive order: first mutation discrimination, then static discrimination, and finally position verification. In mutation discrimination, when the residual cumulative sum feature exceeds the mutation threshold and the velocity change exceeds the velocity change threshold, the collision disturbance mutation state is output. In static discrimination, when the velocity statistical feature is lower than the stranding velocity threshold, the position verification is initiated; otherwise, the free drift state is output. In position verification, when the position-shoreline distance feature is less than the shoreline buffer distance and the current water depth is less than the preset water depth, the stranding static state is output; otherwise, the motion state label of the previous moment is maintained.

7. The water surface dynamic obstacle trajectory prediction method based on multi-source data fusion according to claim 1, characterized in that, In S4, the prediction speed of the perturbation transition prediction model transitions from the current filtering speed to the theoretical drift speed in an exponential decay manner. The decay factor and the transition period are both preset. After the transition period ends, it automatically switches to the free drift prediction model. When switching prediction models, the prediction model switching unit uses the current filtering state as the initial value for the new model.

8. The water surface dynamic obstacle trajectory prediction method based on multi-source data fusion according to claim 1, characterized in that, In S5, the loss determination threshold is adaptively set according to the radar observation confidence level: when the radar observation confidence level is higher than the low confidence threshold, the loss determination threshold is the first threshold; when the radar observation confidence level is lower than the low confidence threshold, the loss determination threshold is the second threshold, which is greater than the first threshold; when the simulation time of the dynamic simulation unit exceeds the maximum simulation time and no registered obstacle is detected again, the registered obstacle is deregistered and a trajectory termination mark is output.

9. The water surface dynamic obstacle trajectory prediction method based on multi-source data fusion according to claim 1, characterized in that, In S6, the uncertainty quantification unit performs perturbation sampling on three types of parameters: initial velocity, estimated wind drift coefficient, and environmental forecast flow velocity. The sampled parameters are used to repeatedly execute the deduction process of the environmental forecast-driven prediction unit to obtain multiple sampling trajectories. The quantiles of the corresponding positions of multiple sampling trajectories are taken for each prediction trajectory point to generate a confidence interval.

10. The water surface dynamic obstacle trajectory prediction method based on multi-source data fusion according to claim 1, characterized in that, In S7, the drift destination marker is one of the following: grounding on shore, entering a waterway, entering a protected area, or continuing to drift; the risk level is divided into Level 1 risk, Level 2 risk, and Level 3 risk; when the drift destination marker is entering a waterway and the entry time does not exceed the preset entry time limit, it is classified as Level 1 risk; when the drift destination marker is entering a waterway and the entry time exceeds the preset entry time limit, or when the drift destination marker is entering a protected area, it is classified as Level 2 risk; When the drift destination is marked as grounding or continued drifting, it is classified as a level three risk; when the confidence interval width of the corresponding predicted trajectory point exceeds the width threshold, the risk level is upgraded by one level.

Images