Moving target track identification and prediction method based on computer vision

Abstract

The invention relates to the technical field of automatic driving, and discloses a moving target track identification and prediction method based on computer vision, which comprises the following steps of: firstly, acquiring observation track points and road data containing map error information, generating a plurality of passable paths and identifying common guide road sections among the passable paths; classifying the paths into non-separable path groups in the same direction according to the length of the shared road section, and calculating a structural confusion degree parameter by combining the separation degree between the groups and the number of groups; modulating a sampling weight by using the parameter, executing deterministic sampling consistency fitting in combination with a preset fixed sequence, and selecting an optimal motion model instance; and finally, performing weighted normalization on each path group according to a scoring result, and outputting a prediction trajectory.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and more specifically, to a method for recognizing and predicting the trajectory of moving targets based on computer vision. Background Technology

[0002] In autonomous driving systems, accurately predicting the future trajectories of surrounding traffic participants (such as vehicles and non-motorized vehicles) is the cornerstone of the planning and decision-making module's safety. Existing prediction schemes typically rely on high-precision maps or online maps generated in real time by onboard sensors to provide road distribution information. For example, the general Lanelet2 framework represents roads as directed graphs containing nodes and connections, uses routing graphs for reachability analysis, and determines the current lane of the target location through a matching module. In the trajectory fitting stage, the principle of random sample consistency is often used to remove outliers from noisy observation data, in order to select motion models that conform to the road direction through sampling and scoring mechanisms.

[0003] However, existing technical solutions face multiple serious challenges when dealing with complex road configurations such as intersections and ramp merging areas: First, there is systemic interference caused by errors in online map construction. Existing online mapping technologies often suffer from geometrical uncertainties when generating lane topology. Current prediction links typically ignore confidence information in this mapping process or treat it as ordinary random noise for filtering. However, research shows that geometrical errors in maps are not simply white noise, but a key variable in the prediction link. Ignoring this variable can lead to misjudgments by the system when determining whether a target has deviated from its lane.

[0004] Second, endpoint drift disrupts the stability of topological connections. In lane topology reasoning tasks, the endpoints connecting lane centerlines often drift. This drift disrupts the connectivity between lanes, making the reasoning logic based on graph connectivity unreliable, which in turn leads to downstream prediction modules incorrectly eliminating potential feasible paths.

[0005] Third, shared guide lines create difficulties in path identification. At intersections or ramps, multiple routes with drastically different future directions often share the same geometric guide line over a considerable distance. In this road layout, relying solely on observation data within a limited historical window makes it difficult for the system to distinguish which branch the target will ultimately enter. Without introducing deeper correlation indicators, traditional fitting methods often generate averaged compromise trajectories that fail to accurately reflect the target's true intentions.

[0006] Fourth, the output nondeterminism introduced by random sampling mechanisms. Traditional fitting methods based on the principle of random sample consistency involve random sampling in their core steps. This means that even with identical input observation data, the system's calculated output at different times may exhibit slight fluctuations or differences. This nondeterminism is unacceptable for mass-production control systems that require high stability and makes rigorous regression testing and verification difficult. Summary of the Invention

[0007] This invention provides a method for recognizing and predicting the trajectory of moving targets based on computer vision, which solves the technical problems mentioned in the background art.

[0008] This invention provides a method for recognizing and predicting the trajectory of a moving target based on computer vision, comprising: The observation trajectory points of the moving target and road grid data containing map error information are obtained to generate multiple passable paths; Based on the map error information, the spatial tolerance range is defined, the shared guiding road segment between the passable paths is identified, and the paths whose shared guiding road segment length exceeds the spatial tolerance range are classified as indivisible path groups in the same direction. Based on the number of groups in the same direction that are inseparable and the lateral position deviation of each group at the far end of the prediction, a confusion parameter is generated. The sampling weights are modulated using the confusion parameter, and a sampling consistency fitting operation is performed based on the sampling weights and a preset fixed sequence to select the optimal motion model instance. Based on the scoring results of the optimal motion model instance, each group of inseparable paths moving in the same direction is weighted and normalized to output the predicted trajectory of the moving target.

[0009] The beneficial effects of this invention include: by constructing a confusion parameter, the map error distribution information and the structural attributes of the shared road guidance segment are jointly quantified, which effectively solves the path identification problem caused by endpoint drift and overlapping guide lines at complex intersections, and significantly improves the robustness of the prediction system to environmental uncertainties; at the same time, by introducing a deterministic pseudo-random sampling sequence and a weight modulation mechanism based on the confusion parameter, while retaining the anti-interference capability of the sampling consistency algorithm, the output jitter and non-determinism caused by traditional random sampling are completely eliminated, and a stable trajectory output that meets the requirements of mass production-level control systems is achieved. Attached Figure Description

[0010] Figure 1 This is a flowchart of the computer vision-based moving target trajectory recognition and prediction method of the present invention. Detailed Implementation

[0011] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0012] like Figure 1 As shown, a computer vision-based method for recognizing and predicting the trajectory of a moving target includes: The observation trajectory points of the moving target and road grid data containing map error information are obtained to generate multiple passable paths; Based on the map error information, the spatial tolerance range is defined, the shared guiding road segment between the passable paths is identified, and the paths whose shared guiding road segment length exceeds the spatial tolerance range are classified as indivisible path groups in the same direction. Based on the number of groups in the same direction that are inseparable and the lateral position deviation of each group at the far end of the prediction, a confusion parameter is generated. The sampling weights are modulated using the confusion parameter, and a sampling consistency fitting operation is performed based on the sampling weights and a preset fixed sequence to select the optimal motion model instance. Based on the scoring results of the optimal motion model instance, each group of inseparable paths moving in the same direction is weighted and normalized to output the predicted trajectory of the moving target.

[0013] The observation trajectory points of a moving target are a set of discrete data points collected by onboard sensors that reflect the continuous movement position of the moving target (such as a vehicle or non-motorized vehicle).

[0014] The process of acquiring the observation trajectory points of a moving target includes: Vehicle-mounted cameras, LiDAR, and millimeter-wave radar were selected as sensing devices. The sensors underwent time synchronization and spatial coordinate calibration beforehand to ensure that the data collected by different sensors were aligned in both time and space. Sensors acquired raw data of moving targets in real time. The camera captured image frames containing the target, the LiDAR output 3D point cloud data of the target, and the millimeter-wave radar provided distance and velocity information. All data included a timestamp. Target detection was performed on the acquired raw data, identifying moving targets such as vehicles and non-motorized vehicles from the images and point clouds. A continuous tracking link was established using multi-target tracking algorithms (such as Kalman filtering and the Hungarian algorithm), and each target was assigned an identifier. Multi-sensor data for the same target were fused. The iterative nearest-point algorithm was used to match point cloud features with image features. Combined with the measurement models of each sensor, the 3D position coordinates of the target in the world coordinate system were calculated to form initial trajectory points. Reasonable position change and velocity thresholds were set to eliminate abnormal trajectory points caused by sensor noise and occlusion, retaining continuous and stable valid data points to ensure trajectory reliability. Valid trajectory points were arranged in timestamp order to form a complete sequence of observed trajectory points.

[0015] Road grid data containing map error information is descriptive data of the road environment, covering both road topology and geometric features, as well as error quantification information generated during map construction, which is used to support path search and uncertainty analysis.

[0016] The process of acquiring road grid data containing map error information includes: First, acquiring basic map data: Based on road data provided by high-precision maps or online map services, extracting geometric information such as lane boundaries, centerlines, and intersection structures, while also acquiring semantic information such as traffic rules (e.g., no-entry, turning restrictions). Second, using vehicle-mounted LiDAR to scan the road environment in real time, collecting 3D point cloud data of the road surface and surrounding facilities. This point cloud data is then registered and fused with the basic map to update road geometry information and compensate for the timeliness limitations of the basic map. Third, based on the updated road geometry information, constructing a directed routing graph model with lane intersections and lane endpoints as nodes, and single lane segments as edges. The direction of the edges is defined according to the actual driving direction, clarifying the connection relationships between lanes. Fourth, analyzing the deviation distribution during the registration process between the basic map and the online point cloud, calculating the registration deviation of key points on the centerline of each lane segment, and constructing a map geometric covariance matrix based on the deviation distribution. The trace of the matrix is ​​used to characterize the magnitude of the map geometric error at that location. The system associates a pre-defined traffic rule database with lane segment edges in a directed routing graph, assigning corresponding traffic rule constraint penalty values ​​to each lane segment edge to clarify the driving restrictions for different lane segments. It integrates road geometry, topological relationships, map geometric covariance matrix, and traffic rule attributes, and then performs grid partitioning of the road according to pre-defined grid division rules to form road grid data containing map error information.

[0017] Data validation and output: The constructed road grid data is validated to check the rationality of topological connections, the accuracy of error quantification, and the correctness of rule associations. Invalid or erroneous data is removed, and finally, complete road grid data is output.

[0018] Preferably, the observation trajectory points of the moving target and road grid data containing map error information are acquired to generate multiple passable paths, including: In the directed routing graph model constructed from the road grid data, a multi-path search algorithm is used to search for a preset number of routes starting from the current location. A collection of alternative routes The path cost function calculation formula used in the multi-path search algorithm is as follows: ; in, Indicate route Total agency value Indicate route The lane segment edges included This represents summing the edges of all lane segments in the route. Indicates lane segment edge geometric length, This represents the weighting coefficient of traffic rule constraints and penalties. Indicates lane segment edge The corresponding traffic rule constraint penalty value; select the total agency value. The smallest front The routes are defined as the passable paths.

[0019] Alternative routes are potential driving routes obtained by searching a directed routing graph model based on road grid data, starting from the current location of the moving target.

[0020] The feasible paths are a preset number of routes selected from the candidate routes that have the lowest path cost function value.

[0021] Lane edge is the basic road unit that makes up each route.

[0022] The geometric length of a lane segment edge is the length of the lane segment edge itself, which can be obtained by measuring with LiDAR or directly reading from a high-definition map.

[0023] The traffic rule constraint penalty weighting coefficient is a coefficient used to balance the lane segment edge geometric length and the traffic rule constraint penalty value. It is preferably between 0.8 and 1.2 to ensure that the geometric length and traffic rule constraint have a balanced proportion in the path cost calculation and avoid a single factor dominating route selection.

[0024] The traffic rule constraint penalty value is a quantitative value set for the traffic rule restriction corresponding to the lane segment edge. It can be obtained through a preset traffic rule priority lookup table. For example, the penalty value for a lane segment edge that prohibits passage is set to 10, and the penalty value for a lane segment edge that restricts turning is set to 3.

[0025] The preset number is the number of alternative routes and passable paths set in advance, preferably 8 to 12, in order to balance route coverage completeness and calculation efficiency. Too few numbers may miss feasible routes, while too many numbers will increase the calculation burden.

[0026] Multipath search algorithms are path search algorithms that can stably output a fixed set of routes. They prioritize strict algorithms, generating alternative routes iteratively and removing duplicates to ensure that the same input corresponds to the same output.

[0027] The path cost function uses a weighted sum of geometric length and traffic rule constraint penalties. Specifically, the total cost of each route equals the sum of the geometric lengths of all lane segments along the route, plus the sum of the products of the traffic rule constraint penalty values ​​and their corresponding weight coefficients for each lane segment. For example, a route may contain two lane segments: the first segment is 50 meters long with no traffic rule restrictions (penalty value 0), and the second segment is 30 meters long with turning restrictions (penalty value 3). With a weight coefficient of 1, the total cost would be 50 + 30 + 1 × 3 = 83.

[0028] Multi-path search algorithms, taking the strict algorithm as an example, first find the first route with the lowest cost, then generate candidate routes by removing edges from the route, and obtain the next optimal route after filtering. Repeat this process until a preset number of candidate routes are generated. In the process, duplicate routes need to be eliminated to ensure the uniqueness and stability of the route set.

[0029] The route selection rule is to select a preset number of routes with the smallest cost function value as the passable paths. For example, when the preset number is 8, all generated candidate routes are sorted from smallest to largest total cost, and the top 8 are selected as passable paths to avoid redundancy or omissions caused by fuzzy route selection criteria.

[0030] The multi-path search algorithm uses the strict algorithm, and the specific implementation steps are as follows: First, based on the directed routing graph model, the Dijkstra algorithm is used to find the first optimal route starting from the current position; Second, each edge in the first route is removed in turn, and the optimal route is searched again in the remaining graph structure to generate a set of candidate routes; Third, routes that are duplicates of the selected routes are eliminated from the candidate routes, and the best one is selected as the second traversable path; Fourth, steps two and three are repeated until a preset number of traversable paths are obtained. The search terminates when the number of generated traversable paths reaches the preset value.

[0031] The quantitative standards for traffic rule-based penalties are graded according to the strictness of the rules: a penalty of 10 is set for prohibitions on passage (such as red lights and wrong-way lanes), a penalty of 7 is set for strong restrictions (such as emergency lanes on highways and prohibitions on U-turns), a penalty of 3 is set for weak restrictions (such as yielding to turns and speed limits), and a penalty of 0 is set for no restrictions. This grading is based on the degree of impact of traffic rules on driving to ensure the rationality of the penalty values.

[0032] The value of the traffic rule constraint penalty weight coefficient needs to be adjusted according to the actual scenario. In the scenario of urban road intersections, traffic rules have a greater impact, and the weight coefficient can be taken as 1.2; in the scenario of highways, geometric length has a greater impact, and the weight coefficient can be taken as 0.8; in the scenario of ordinary urban roads, it is taken as 1.0 to ensure that the cost function calculation can reflect the actual advantages and disadvantages of the route in different scenarios.

[0033] Preferably, the spatial tolerance range is defined by combining the map error information, and the shared guiding road segments between the passable paths are identified, including: For each of the observed trajectory points Calculate the single-point error scale The average value of all observation points is calculated as the spatial tolerance range. The calculation formula is as follows: ; ; in, This represents the preset scale coefficient. Represents the trace operation of a matrix. Indicates the first The perceptual covariance matrix of each observed trajectory point Indicates the first Projected positions of observation trajectory points The geometric covariance matrix of the map at that location, This represents the total number of observed trajectory points; For any two of the aforementioned traversable paths and splicing their centerlines into a continuous curve and Calculate the length of the shared guide road segment. The calculation formula is as follows: ; in, This indicates the supremum operation. Represents the arc length distance variable. This represents the arc length parameter along the centerline. Indicates the length from the starting point. The arc length interval, Indicates the arc length of two continuous curves The geometric Euclidean distance at [location] This indicates the spatial tolerance range.

[0034] The observation trajectory point index is a sequence number used to distinguish different observation trajectory points.

[0035] The perception covariance matrix is ​​a matrix that describes the distribution of positioning errors of observed trajectory points. It can be calculated and obtained through a positioning algorithm that fuses camera and lidar.

[0036] The arc length of the projection position is the arc distance from the observation trajectory point projected onto the reference line.

[0037] The map geometric covariance matrix is ​​a matrix that describes the distribution of geometric errors at corresponding locations in road grid data. It can be obtained through LiDAR point cloud matching error analysis during online map construction.

[0038] The preset scale coefficient is a coefficient used to convert the sum of the traces of the covariance matrix into the actual error scale. It is preferably 1.0 to 1.5 so that the calculated single-point error scale fits the perception and map error levels in the actual scene.

[0039] The single-point error scale is the comprehensive error quantification value of a single observed trajectory point. It is obtained by multiplying the square root of the sum of the trace of the perception covariance matrix and the trace of the map geometric covariance matrix by a preset scale coefficient.

[0040] The total number of observation trajectory points is the total number of observation trajectory points involved in the calculation, which can be obtained by counting the number of valid trajectory points collected by the sensors.

[0041] Spatial tolerance range is the arithmetic mean of the single-point error scales of all observed trajectory points, used to determine whether two paths are within an indistinguishable geometric range.

[0042] Any two passable paths are a pair of paths selected from the filtered passable paths, used to calculate the length of the shared guide road segment between them.

[0043] A continuous curve is a smooth curve formed by sequentially splicing together the centerlines of all segments of a single passable path.

[0044] Arc length is a variable used to measure the length of a continuous curve, and its value ranges from 0 to 0.

[0045] The length of the shared guide segment is the maximum arc length between any two continuous curves of passable paths, where the geometric Euclidean distance is consistently less than or equal to the spatial tolerance range.

[0046] The calculation logic for spatial tolerance range involves the fusion and quantification of multi-source errors, rather than a simple summation. Specifically, first, the trace of the perceptual covariance matrix corresponding to each observation trajectory point is calculated, reflecting the degree of error dispersion in the perceptual data. Then, the trace of the map geometric covariance matrix at the projected location of that point is calculated, reflecting the degree of error dispersion in the map data. The two traces are added together and the square root is taken to obtain the unscaled comprehensive error. After multiplying by a preset scaling factor, the single-point error scale of a single observation trajectory point is obtained. Finally, the arithmetic mean of the single-point error scales of all observation trajectory points is calculated, which is the spatial tolerance range. For example, if the trace of the perceptual covariance matrix of an observation trajectory point is 0.04, the trace of the map geometric covariance matrix is ​​0.01, and the preset scaling factor is 1.0, then the single-point error scale is √(0.04 + 0.01) = 0.2236. If there are 10 similar observation points, the average spatial tolerance range is approximately 0.22.

[0047] The shared guide segment length is used to identify areas where the distance is consistently less than or equal to the spatial tolerance range. First, the centerlines of each segment of the two passable paths are spliced ​​together in the order of travel, using a Bézier curve for a smooth transition to ensure continuous curves without breaks. Then, starting from the starting point, the path is advanced synchronously along the two continuous curves, calculating the geometric Euclidean distance point by point. When the distance at a point exceeds the spatial tolerance range, advancement stops. The maximum arc length from the starting point to that point is the shared guide segment length. For example, after splicing the two paths, if the Euclidean distance between all points within the first 80 meters is less than 0.3 meters, and the distance at the 81st meter is 0.35 meters, with a spatial tolerance range of 0.3 meters, then the shared guide segment length is 80 meters.

[0048] The perceptual covariance matrix is ​​obtained using a camera and LiDAR fusion positioning algorithm. The steps are as follows: the camera acquires image data and identifies feature points, and the LiDAR acquires 3D point cloud data; the two types of data are synchronized in time and calibrated in space; the feature points and point cloud are matched by the iterative nearest point algorithm; the residual distribution in the matching process is calculated, and the perceptual covariance matrix is ​​constructed based on the residuals.

[0049] The geometric covariance matrix of a map originates from the online map building process, which specifically involves: LiDAR collecting 3D point cloud data of the road environment; registering and matching the point cloud data with an existing map; calculating the deviation between the registered point cloud and the corresponding location on the map; and constructing the geometric covariance matrix of the map based on the deviation distribution. The larger the deviation, the larger the trace of the covariance matrix.

[0050] The value of the preset scale coefficient needs to be adjusted according to the scenario: for urban road intersection scenarios, where the perception and map errors are relatively large, the coefficient is set to 1.3 to 1.5; for highway scenarios, where the errors are relatively small, the coefficient is set to 1.0 to 1.2; and for ordinary urban road scenarios, the coefficient is set to 1.2 to ensure the rationality of the spatial tolerance range under different scenarios.

[0051] The splicing of continuous curves adopts a third-order Bézier curve transition, specifically: selecting the endpoints of the center lines of two adjacent road segments and one control point before and after the endpoints; by adjusting the positions of the control points, the first derivative of the spliced ​​curve is made continuous; ensuring that the curve is smooth without abrupt changes, and avoiding the impact of splicing breakpoints on the calculation of the length of the shared guide road segment.

[0052] Preferably, paths whose shared guide road segment length exceeds the spatial tolerance range are classified into a group of inseparable paths in the same direction, including: Set the historical observation window length value For any two of the aforementioned passable paths and Determine the length of the shared guide road segment. The historical observation window length value The size relationship; if the following relationship is satisfied: ; Then, it is determined that the two passable paths belong to the same indivisible path group in the same direction, denoted as... Perform the above determination on all the said passable paths to obtain the number of groups. The indivisible path group in the same direction .

[0053] The length of the historical observation window is a pre-set value used to measure the range of historical trajectory observation. It is preferably 2 to 3 seconds or 15 to 20 meters, so as to cover enough historical trajectory information to reflect the movement trend, without causing redundant calculations or information lag due to the excessive range.

[0054] The length of the shared guide segment is the maximum arc length between any two continuous curves of passable paths, where the geometric Euclidean distance is consistently less than or equal to the spatial tolerance range.

[0055] A set of inseparable paths in the same direction is a set of two or more passable paths that satisfy the condition that the length of the shared guiding road segment is greater than or equal to the length of the historical observation window.

[0056] The number of groups is the number of indivisible path groups in the same direction that are ultimately formed after all passable paths have been processed.

[0057] The path determination relationship is a logical relationship used to determine whether two passable paths belong to the same indivisible path group in the same direction. That is, when the length of the shared guide segment of the two paths is greater than or equal to the length of the historical observation window, the two paths have this determination relationship.

[0058] The core logic for path grouping is to correlate the length of the shared guide segment with the length of the historical observation window, rather than simply relying on the geometric similarity of the paths. Specifically, the length of the historical observation window defines the range of currently available valid historical trajectories. If the length of the shared guide segment of two paths exceeds this range, it means that, based on the existing historical observation data, their geometric trajectories cannot be distinguished, and therefore they are grouped together. For example, if the historical observation window length is set to 2 seconds (corresponding to 15 meters), and the length of the shared guide segment of two paths is 18 meters, which is greater than 15 meters, it means that the existing historical trajectories are insufficient to distinguish their future directions, and they are determined to be the same inseparable path group in the same direction.

[0059] When handling complex relationships between multiple paths, transitivity rules are followed. If path A and path B meet the criteria, and path B and path C meet the criteria, then path A and path C are automatically grouped into the same group. For example, if the shared guide segment length of paths A and B is 20 meters, the shared guide segment length of paths B and C is 19 meters, and the historical observation window length is 15 meters, all three conditions are met, ultimately forming a group of inseparable paths in the same direction containing the three paths.

[0060] The length of the historical observation window needs to be dynamically adjusted according to the scenario: In urban road intersection scenarios, the traffic environment is complex and the trajectory changes rapidly, so a value of 2.5 to 3 seconds (corresponding to 18 to 20 meters) should be preferred; in highway scenarios, the trajectory is relatively stable, so a value of 2 to 2.5 seconds (corresponding to 15 to 18 meters) should be preferred; in ordinary urban road scenarios, a value of 2 seconds (corresponding to 15 meters) should be used to ensure that the observation range matches the actual needs in different scenarios.

[0061] The specific steps for grouping multiple paths are as follows: traverse all passable paths and generate all path pairs; for each path pair, determine whether the length of the shared guiding road segment is greater than or equal to the length of the historical observation window; mark the path pairs that meet the conditions as associated; perform clustering based on the association relationship, and group all mutually associated paths into one group; count the number of groups after clustering, which is the final number of groups.

[0062] The anomaly handling mechanism includes: if the number of groups is 0, it means that the length of the shared guiding road segment of all path pairs is less than the length of the historical observation window, and each path is treated as a separate group; if the number of groups is 1, it means that all passable paths are related to each other and are grouped into the same group, and subsequent processes are calculated based on this single group.

[0063] Preferably, based on the number of groups in the same-direction inseparable path group and the lateral position deviation of each group at the prediction far end, a confusion parameter is generated, including: The path centerline with the minimum global path cost function is selected as the reference centerline. For each of the aforementioned indivisible path groups in the same direction The path with the smallest path cost function within the group is selected as the representative path, and its centerline is... ; Calculation in the prediction time domain Corresponding predicted distance At this point, the lateral positional deviation of the representative path relative to the reference center line is a numerical value. The calculation formula is as follows: ; in, This represents the normal vector at the current moment. Indicates the arc length at the current position. Indicates the predicted distance; Calculate the confusion parameter The calculation formula is as follows: ; ; ; in, This indicates information about categories that are not easily distinguishable. This indicates the number of groups in the indivisible path group in the same direction. Represents the natural logarithm operation. The variance value represents the lateral positional deviation. This represents the arithmetic mean of the lateral positional deviations of all groups.

[0064] The reference centerline is the centerline corresponding to the path with the minimum global path cost function among all traversable paths.

[0065] A set of inseparable paths in the same direction is a set of two or more passable paths that satisfy the condition that the length of the shared guiding road segment is greater than or equal to the length of the historical observation window.

[0066] The representative path is the path with the minimum path cost function within each group of inseparable paths moving in the same direction.

[0067] The centerline of the path represents the centerline of the road corresponding to the path.

[0068] The prediction time domain is a pre-defined time range for predicting the future trajectory of a moving target, preferably 3 to 5 seconds, to balance the foresight of the prediction and the real-time nature of the calculation. Too short a time cannot reflect long-term trends, while too long a time will increase uncertainty.

[0069] The predicted distance is the distance that a moving target may travel within the prediction time domain, and is obtained by multiplying the robust velocity estimate within the window with the prediction time domain value.

[0070] The arc length at the current position is the arc distance corresponding to the current position of the moving target on the reference center line.

[0071] The normal vector at the current moment is a vector perpendicular to the tangent direction of the reference centerline at the current position, and is used to calculate the lateral position deviation.

[0072] Lateral position deviation represents the lateral offset of the path at the predicted far end position relative to the reference centerline.

[0073] The arithmetic mean of the lateral positional deviations is the average value of the lateral positional deviations of all inseparable paths in the same direction.

[0074] The variance of lateral positional deviation is a numerical value that reflects the degree of dispersion of lateral positional deviations for all inseparable path groups in the same direction.

[0075] Category indistinguishability information is a value obtained by taking the natural logarithm of the number of groups of indivisible paths in the same direction, and is used to quantify the degree of indistinguishability of paths.

[0076] The confusion parameter is the product of the class indistinguishability information and the variance of the lateral positional bias, and is used to comprehensively quantify the degree of path indistinguishability and the cost of misclassification.

[0077] The robust speed estimation within the window is based on the stable speed of the moving target calculated from trajectory data within the historical observation window. The median speed is used first to avoid interference from extreme speed values.

[0078] The logic behind the synthesis of the confusion parameter combines the indistinguishability of paths with the cost of misclassification, rather than quantifying it from a single dimension. Specifically, the amount of class indistinguishable information is determined by the number of groups; the more groups, the more indistinguishable path categories there are under existing historical data, and the larger this value. The variance of the lateral positional deviation is determined by the dispersion of the lateral offsets of each group; the larger the variance, the higher the degree of separation between different path groups at the prediction distance, and the greater the trajectory error after misclassification. The confusion parameter obtained by multiplying the two values ​​simultaneously reflects these two key characteristics. For example, if there are four indistinguishable path groups in the same direction, the natural logarithm of the number of groups is approximately 1.386, the lateral positional deviations of each group are 0.5 meters, 0.8 meters, 1.2 meters, and 1.5 meters, the arithmetic mean is 1.0 meter, and the variance is 0.145, then the confusion parameter is 1.386 multiplied by 0.145, which is approximately equal to 0.201.

[0079] The selection rules for representative paths and reference centerlines have clear priorities. The reference centerline selects the path with the lowest global cost, ensuring a unified benchmark for lateral deviation calculation. The representative path selects the path with the lowest cost within a group, ensuring the typicality of the path characteristics in each group. This two-level selection method avoids deviation calculation distortion caused by inconsistent benchmarks. For example, if there are three paths in a group with costs of 80, 85, and 90 respectively, the path with a cost of 80 is selected as the representative path to ensure accurate representation of the characteristics of that group.

[0080] The prediction time domain value needs to be dynamically adjusted according to the scenario: In urban road intersection scenarios, the traffic environment is complex and the trajectory changes rapidly, so a value of 4 to 5 seconds is preferred; in highway scenarios, the trajectory is relatively stable, so a value of 3 to 4 seconds is preferred; in ordinary urban road scenarios, a value of 3 seconds is preferred to ensure that the prediction range matches the scenario requirements.

[0081] The steps for calculating the normal vector at the current moment are as follows: Select two adjacent points before and after the current position on the reference center line; calculate the vector formed by these two points as the tangent direction vector of the current position; rotate the tangent direction vector counterclockwise by 90 degrees to obtain the normal vector at the current moment, ensuring that the direction is consistent with the left or right side of the road.

[0082] The calculation method for robust speed estimation within the window is as follows: extract the instantaneous speed corresponding to each observation point within the historical observation window; remove outliers in the instantaneous speed that exceed the normal driving speed range, such as removing values ​​greater than 60 km / h or less than 0 km / h on urban roads; calculate the median of the remaining effective instantaneous speeds as the robust speed estimate within the window.

[0083] The method for calculating the lateral position deviation is as follows: determine the position point of the reference centerline at the far end of the prediction, that is, the point corresponding to the sum of the current position arc length and the prediction distance; determine the position point of the representative path centerline at the far end of the prediction; calculate the distance between these two position points in the direction of the normal vector at the current moment, which is the lateral position deviation. The normal vector points to one side and is negative on the other side.

[0084] Preferably, modulating the sampling weights using the confusion parameter includes: The confusion parameter is obtained by using a linear mapping function. Convert to topology information temperature parameters The calculation formula is as follows: ; in, and It is a preset fixed constant; Calculate the length of a typical shared guide road segment The calculation formula is as follows: ; in, This indicates the median operation. Indicates the length of the shared guide segment between any two passable paths; Regarding the first For each observed trajectory point, calculate its topological separation sensitive weight. The calculation formula is as follows: ; ; in, Indicates the first The arc length values ​​along the line for each observation trajectory point This indicates exponentiation. This represents the total number of observed trajectory points. These are the initial weight values ​​before normalization.

[0085] The confusion parameter is a parameter that comprehensively quantifies the indistinguishability of inseparable paths in the same direction and the cost of misclassification.

[0086] The topological information temperature parameter is a parameter obtained by transforming the confusion parameter through a linear mapping function, and is used to modulate the weight of the observed trajectory points.

[0087] The linear mapping function is a function that converts the confusion parameter to the topology information temperature parameter. The core is to ensure that the higher the confusion, the larger the topology information temperature parameter.

[0088] The preset fixed constant τ0 is the reference value of the topology information temperature parameter, preferably 0.5 to 1.0, so that the initial range of the temperature parameter fits the weighted modulation requirements and avoids modulation failure caused by the initial value being too high or too low.

[0089] The preset fixed constant α is the adjustment coefficient of the topology information temperature parameter, preferably 0.8 to 1.2, in order to balance the influence of confusion on the temperature parameter and ensure that the temperature change matches the confusion.

[0090] The typical shared guide segment length is the median length of the shared guide segment for all paired accessible paths, and is used to characterize the overall typical length of shared path segments.

[0091] The arc length along the observation trajectory point is the arc distance of each observation trajectory point along the reference center line, which can be obtained by cumulatively calculating after projecting the observation points onto the reference center line.

[0092] The initial weights are the weight values ​​obtained by exponential calculation for each observed trajectory point, without normalization.

[0093] The topology separation sensitive weight is the weight obtained after normalizing the initial weight, and is used to guide the selection probability of observation points during the sampling process.

[0094] The total number of observation trajectory points is the number of observation trajectory points that participate in the weight calculation, which can be obtained by counting the total number of valid trajectory points collected by the sensors.

[0095] The specific implementation of the linear mapping function is that the topological information temperature parameter is equal to the product of a preset fixed constant τ0 multiplied by 1 and the preset fixed constant α and the confusion parameter. For example, if the confusion parameter is 0.2, τ0 is 0.8, and α is 1.0, then the topological information temperature parameter is 0.8×(1+1.0×0.2)=0.96, ensuring that the temperature parameter increases synchronously when the confusion increases.

[0096] The median rather than the average length of typical shared guide segments is chosen because the median avoids interference from extreme values ​​and better reflects the characteristics of shared segment lengths for most routes. For example, the shared guide segment lengths for five paired routes are 60 meters, 70 meters, 80 meters, 90 meters, and 150 meters, with an average of 90 meters and a median of 80 meters, which better reflects the typical shared segment length.

[0097] The initial weights are calculated by dividing the difference between the arc length along the observation trajectory point and the length of the typical shared guide road segment by the topological information temperature parameter, followed by an exponential operation. This gives higher weights to observation points closer to the end of the typical shared road segment (near the separation point). For example, if an observation point has an arc length of 85 meters, a typical shared guide road segment length of 80 meters, a temperature parameter of 0.96, and a difference of 5 meters, its initial weight after the exponential operation is higher than that of an observation point with an arc length of 70 meters, thus strengthening the role of separation point information in sampling.

[0098] Normalization involves summing the initial weights of all observation points, and then dividing each initial weight by the sum to obtain the topology separation sensitive weights. This ensures that the total weights are equal to 1, which meets the probability distribution requirements. For example, if the initial weights of three observation points are 2, 3, and 5, with a total of 10, the normalized weights will be 0.2, 0.3, and 0.5, respectively.

[0099] The specific form of the linear mapping function is that the topological information temperature parameter is equal to τ0 multiplied by 1 plus α multiplied by the confusion parameter. This linear relationship can guarantee the stability and interpretability of the conversion and avoid the uncontrollable modulation caused by complex functions.

[0100] The reference centerline is used as the benchmark for calculating the arc length of the observation trajectory points. The specific steps are as follows: take the starting point of the reference centerline as the starting point of the arc length, project the observation trajectory points onto the reference centerline, calculate the cumulative arc length distance from the starting point to the projection point, and ensure that the arc length calculation benchmark of all observation points is consistent.

[0101] The base of the exponential operation is fixed as the natural constant, which is approximately 2.718. This base can make the weight changes smooth, avoid the weight distribution being too concentrated or dispersed, and ensure the balance of the sampling process.

[0102] The normalization process uses L1 normalization, which means that the initial weight of each observation point is divided by the sum of all initial weights. The specific steps are: calculate the sum of all initial weights, divide each initial weight by the sum, and retain the calculation result as the topology separation sensitive weight to ensure the effectiveness and consistency of the weights.

[0103] Preferably, a sampling consistency fitting operation is performed based on the sampling weights and a preset fixed sequence to select the optimal motion model instance, including: Set a fixed number Candidate motion model instances; for the first Given several candidate motion model instances, generate a deterministic pseudo-random numerical sequence. The calculation formula is as follows: ; in, This represents the Sobol sequence generating function. Indicates the index of the smallest sample point. This represents the hash seed calculated using the unique identifier of the moving target; Using the aforementioned topology to separate sensitive weights The cumulative distribution function is selected based on the corresponding observation point index. : ; Fit polynomial motion curves to the extracted samples Using the aforementioned confusion parameter Modulation residual standard deviation : ; in This is a preset constant; Calculate the first Likelihood score for each instance , and select The instance corresponding to the maximum is the instance of the optimal motion model: ; ; in, The value of the fitted residual is... Here is the interior point weight constant. is the background noise constant.

[0104] The fixed number of candidate motion model instances is the pre-set number of motion models participating in the fitting, preferably 64 to 128, in order to balance the fitting effect and computational efficiency. Too few instances may miss the optimal model, while too many instances will increase the computational delay.

[0105] The candidate motion model instance index is a sequence number that distinguishes different candidate motion model instances.

[0106] The identifier of a moving target is a number used to uniquely identify each moving target, which can be obtained through camera image recognition or lidar target tracking technology.

[0107] The hash map value is a numerical value obtained by hashing the identifier of the moving target, and is used to initialize the fixed sequence generator.

[0108] The Sobol sequence generator is a function that generates uniformly distributed pseudo-random numbers and is used to generate deterministic pseudo-random numerical sequences.

[0109] The index of the minimum sample point is the index that distinguishes each observation point in the minimum sample set.

[0110] A pseudo-random numerical sequence is a deterministic numerical sequence with values ​​between 0 and 1, generated by the Sobol sequence generating function.

[0111] Topology separation sensitive weights are normalized weights used to guide the probability of selecting observation points.

[0112] The cumulative distribution function is a function built on topological separation sensitive weights, used to map pseudo-random numerical values ​​to observation point indices.

[0113] The observation point index is the sequence number of the corresponding observation trajectory point, used to extract the minimum sample set from all observation points.

[0114] The minimum sample set is the minimum set of observations required to fit a motion model.

[0115] The minimum sample set size is the number of observation points that constitute the minimum sample set, preferably 4, so as to meet the fitting requirements of the cubic polynomial motion curve and ensure that the model parameters are solved completely.

[0116] Polynomial motion curves are curves that describe the longitudinal and lateral motion patterns of a moving target, including longitudinal cubic polynomials and lateral linear polynomials.

[0117] Longitudinal motion parameters are parameters in polynomial motion curves that describe the longitudinal movement of a target along a road.

[0118] Lateral motion parameters are parameters in polynomial motion curves that describe the lateral offset of a target along the road.

[0119] The confusion parameter is a parameter that comprehensively quantifies the indistinguishability of inseparable paths in the same direction and the cost of misclassification.

[0120] The preset standard deviation of the longitudinal residual is the initial standard deviation of the longitudinal fitting residual, preferably 0.1 to 0.3 meters, to match the longitudinal positioning error level in a typical road scenario.

[0121] The preset standard deviation of the lateral residual is the initial standard deviation of the lateral fitting residual, preferably 0.05 to 0.2 meters, to match the lateral positioning error level in a typical road scenario.

[0122] The modulation coefficient is a constant used to adjust the influence of the confusion parameter on the residual standard deviation, preferably 0.8 to 1.2, in order to balance the matching relationship between confusion and residual scale.

[0123] The standard deviation of the modulated longitudinal residual is the standard deviation of the longitudinal residual after expansion modulation by the confusion parameter.

[0124] The standard deviation of the modulated transverse residual is the standard deviation of the transverse residual after expansion modulation by the confusion parameter.

[0125] The observation trajectory point index is a sequence number that distinguishes different observation trajectory points.

[0126] The fitting residual is the deviation between the actual value of the observed trajectory point and the predicted value of the polynomial motion curve, which is normalized by the standard deviation of the residual.

[0127] The in-point weight constant is the weighting coefficient of the in-point contribution in the likelihood score, preferably between 0.7 and 0.9, to highlight the dominant role of in-points in the model score.

[0128] The background noise constant is the baseline contribution value of noise points in the likelihood score, preferably between 1e-6 and 1e-4, to avoid noise points significantly interfering with the model score.

[0129] The likelihood score is a numerical measure of how well a candidate motion model instance fits the model; a higher score indicates a better fit.

[0130] The optimal motion model instance is the model instance with the highest likelihood score among all candidate motion model instances.

[0131] The core of the deterministic pseudo-random sampling mechanism is to use the hash mapping value of the identifier of a moving target as a seed to ensure that the same input corresponds to the same pseudo-random numerical sequence. For example, if the unique identifier of a moving target is 123 and the hash mapping value is 456, the Sobol sequence generator is initialized with this, and the pseudo-random numerical sequence generated each time is completely consistent, thus solving the non-deterministic problem of traditional random sampling.

[0132] The cumulative distribution function mapping logic for topology-separated sensitive weights is as follows: first, calculate the cumulative sum of weights for all observation points; then, compare the pseudo-random value with the cumulative sum one by one, and find the first observation point whose cumulative sum is greater than or equal to the pseudo-random value as the sampling point. For example, if the weights of three observation points are 0.2, 0.3, and 0.5, the cumulative sums are 0.2, 0.5, and 1.0, and the pseudo-random value is 0.4, then the second observation point is selected.

[0133] The expansion modulation of the residual standard deviation dynamically adjusts the judgment scale of the residuals through the confusion parameter. The higher the confusion, the larger the residual standard deviation, and the higher the tolerance for fitting bias. For example, if the preset benchmark value of the longitudinal residual standard deviation is 0.2 meters, the modulation coefficient is 1.0, and the confusion parameter is 0.3, then the modulated longitudinal residual standard deviation is 0.2 × (1 + 1.0 × 0.3) = 0.26 meters, avoiding the misjudgment of small deviations under the shared prefix as outliers.

[0134] The likelihood score is calculated by summing the log-likelihood transformations of the fitting residuals for each observation point. The inlier weight constant determines the contribution percentage of each inlier. For example, if the fitting residual for an observation point is 0.1, the inlier weight constant is 0.8, and the background noise constant is 1e-5, then the likelihood contribution for that point is log(0.8×e...). (-0.1 / 2) The sum of the contributions of all points (+0.2×1e-5) is the likelihood score.

[0135] The Sobol sequence generation function is configured as follows: the sequence dimension is set to the product of the fixed number of candidate motion model instances and the minimum sample set size; the initial state is set to 0; and the period is set to the minimum prime number greater than the number of candidate motion model instances to ensure the uniform distribution characteristics of the sequence.

[0136] The minimum sample set size needs to be matched with the order of the motion model. If a combination of vertical cubic polynomial and horizontal linear polynomial is used, the minimum sample set size is fixed at 4. The vertical cubic polynomial requires more than 3 observation points to determine 4 parameters, and the horizontal linear polynomial requires more than 2 observation points to determine 2 parameters. 4 observation points can simultaneously meet the fitting requirements of both.

[0137] The preset benchmark values ​​for longitudinal residual standard deviation and lateral residual standard deviation need to be adjusted according to the scenario: In the scenario of urban road intersections, where the error is relatively large, the longitudinal benchmark value should be 0.2 to 0.3 meters and the lateral benchmark value should be 0.1 to 0.2 meters; In the scenario of highways, where the error is relatively small, the longitudinal benchmark value should be 0.1 to 0.2 meters and the lateral benchmark value should be 0.05 to 0.1 meters.

[0138] The logic for setting the inlier weight constant and the background noise constant is as follows: the inlier weight constant is set according to the expected proportion of inliers in the scene, and is 0.8 for normal scenes. The background noise constant is 1e-5 to ensure that the likelihood contribution of noise points is much smaller than that of inliers and does not affect the model score ranking.

[0139] The details of the fitting residual calculation are as follows: the longitudinal residual is the difference between the arc length of the observation point along the reference center line and the arc length predicted by the model, and the lateral residual is the difference between the lateral offset of the observation point relative to the reference center line and the offset predicted by the model. The two are divided by the square of the corresponding standard deviation of the modulated residual and then summed to obtain the normalized fitting residual.

[0140] Preferably, the predicted trajectory of the moving target is output by weighting and normalizing each group of inseparable paths in the same direction based on the scoring results of the optimal motion model instance, including: For each of the aforementioned indivisible path groups Calculate the path group score. The calculation formula is as follows: ; in, This represents the likelihood score of the optimal motion model instance. This represents the preset penalty weight constant. This indicates the lateral positional deviation of the group. This represents the arithmetic mean of the lateral positional deviations of all groups. This indicates the temperature parameter of the topology information; Calculate the normalized weight values ​​for each group. The calculation formula is as follows: ; in, This indicates the number of groups in the indivisible path group in the same direction. Represents the natural constant. This indicates summing over all groups; Generate intra-group mapping trajectory The predicted trajectory of the moving target is calculated using weighted averages. The calculation formula is as follows: ; ; in, This represents the centerline of the representative path for this group. and The parameters represent the longitudinal and lateral motion parameters of the optimal motion model instance in the road alignment coordinate system. Represents the centerline normal vector. Indicates the predicted time.

[0141] A set of inseparable paths in the same direction is a set of two or more passable paths that satisfy the condition that the length of the shared guiding road segment is greater than or equal to the length of the historical observation window.

[0142] The path group score is a numerical measure of the fitness of each inseparable path group in the same direction. It is obtained by subtracting the separation penalty term from the likelihood score of the optimal motion model instance.

[0143] The optimal motion model instance is the model instance with the highest likelihood score among all candidate motion model instances.

[0144] Likelihood score is a numerical measure of how well a candidate motion model instance fits the model.

[0145] The separation penalty term is a numerical value used to constrain the dispersion of the lateral positional deviation of the path group. It is proportional to the square of the difference between the lateral positional deviation and the mean, and inversely proportional to the topology information temperature parameter.

[0146] The penalty weight constant is a constant that adjusts the strength of the separation penalty term, preferably between 0.8 and 1.2, in order to balance the model fitting accuracy and the rationality of path group separation, and avoid over-penalization or under-penalization.

[0147] Lateral position deviation represents the lateral offset of the path at the predicted far end position relative to the reference centerline.

[0148] The arithmetic mean of the lateral positional deviations is the average value of the lateral positional deviations of all inseparable paths in the same direction.

[0149] The topological information temperature parameter is a parameter obtained by transforming the confusion parameter through a linear mapping function, and is used to modulate the weight of the observed trajectory points.

[0150] Normalized weights are weights obtained by performing exponential operations and normalization on the scores of all path groups, and are used to weighted fuse the mapped trajectories of each group.

[0151] The number of groups for indivisible paths in the same direction is the final number of groups formed after all passable paths have been processed.

[0152] The intra-group mapped trajectory is the trajectory generated by mapping the longitudinal and lateral motion parameters of the optimal motion model instance to the center line of the corresponding path group representative path.

[0153] The representative path centerline is the centerline corresponding to the path with the minimum path cost function within each group of inseparable paths in the same direction.

[0154] Longitudinal motion parameters are parameters in polynomial motion curves that describe the longitudinal movement of a target along a road.

[0155] Lateral motion parameters are parameters in polynomial motion curves that describe the lateral offset of a target along the road.

[0156] The centerline normal vector is a vector perpendicular to the direction of the tangent at the corresponding position representing the centerline of the path, and is used to calculate the lateral offset mapping.

[0157] The prediction time is any time between the current time and the current time plus the prediction time domain.

[0158] The predicted trajectory is the final trajectory obtained by linearly weighting and summing all the mapped trajectories within the group, reflecting the future driving path of the moving target.

[0159] The design logic of path group scoring combines model fitting performance with path separation characteristics, rather than solely relying on model scores. Specifically, the likelihood score reflects the model's accuracy in fitting historical trajectories, while the separation penalty term constrains the degree of separation between path groups and other groups, avoiding the selection of path groups with excessively large separation. For example, if the optimal model likelihood score is 100, the difference between the lateral position deviation of a certain path group and the mean is 0.5 meters, the topological information temperature parameter is 0.9, and the penalty weight constant is 1.0, then the separation penalty term is 1.0 × (0.5)² / 0.9 ≈ 0.278, and the path group score is 100 - 0.278 ≈ 99.722.

[0160] The normalized weights are calculated by performing exponential operations followed by summation and normalization. That is, the score of each path group is first subjected to exponential operations, and then all the exponential operation results are summed. The normalized weight is obtained by dividing the exponential operation result of each path group by the sum.

[0161] The core of generating the intra-group mapped trajectory is to combine the motion parameters of the optimal model with the centerline representing the path group. The longitudinal motion parameters determine the distance traveled along the centerline, the lateral motion parameters determine the offset relative to the centerline, and the centerline normal vector ensures the correct lateral offset direction. For example, if the longitudinal motion parameters predict a travel of 50 meters along the centerline at a certain moment, the lateral motion parameters predict an offset of 0.3 meters, and the normal vector points to the right side of the road, then the mapped trajectory at that moment will be a point 50 meters from the centerline, offset 0.3 meters to the right.

[0162] The logic of generating predicted trajectories through linear weighted summation is to multiply the mapped trajectory within each group by the corresponding normalized weight and then sum them to obtain a single trajectory.

[0163] The value of the penalty weight constant needs to be dynamically adjusted according to the scenario: in urban road intersection scenarios, where path separation is complex, the value is 1.0 to 1.2; in highway scenarios, where path separation is simple, the value is 0.8 to 1.0; and in ordinary urban road scenarios, the value is 1.0 to ensure that the penalty intensity is adapted to different scenarios.

[0164] The steps for calculating the normal vector of the representative path centerline are as follows: First, select two adjacent points before and after the position corresponding to the prediction time on the representative path centerline; second, calculate the vector formed by these two points as the tangent direction vector at that position; third, rotate the tangent direction vector counterclockwise by 90 degrees to obtain the centerline normal vector, ensuring that the direction is consistent.

[0165] The time step for prediction is set to 0.1 seconds, meaning that starting from the current time, a prediction time is taken every 0.1 seconds until the current time is added to the prediction time domain, ensuring the continuity and refinement of the prediction trajectory. For example, when the prediction time domain is 3 seconds, a total of 31 prediction times are included.

[0166] Numerical stability is achieved using an exponential shift technique. This involves subtracting the maximum score from each path group's score before performing the exponential calculation, thus preventing exponential overflow caused by excessively large scores. For example, if three path groups have scores of 100, 99.8, and 99.5, subtracting the maximum score of 100 yields scores of 0, -0.2, and -0.5, respectively. The exponential calculation results are approximately 1, 0.819, and 0.607, which are then normalized to ensure computational stability.

[0167] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A method for recognizing and predicting the trajectory of a moving target based on computer vision, characterized in that, include: The observation trajectory points of the moving target and road grid data containing map error information are obtained to generate multiple passable paths; Based on the map error information, the spatial tolerance range is defined, the shared guiding road segment between the passable paths is identified, and the paths whose shared guiding road segment length exceeds the spatial tolerance range are classified as indivisible path groups in the same direction. Based on the number of groups in the same direction that are inseparable and the lateral position deviation of each group at the far end of the prediction, a confusion parameter is generated. The sampling weights are modulated using the confusion parameter, and a sampling consistency fitting operation is performed based on the sampling weights and a preset fixed sequence to select the optimal motion model instance. Based on the scoring results of the optimal motion model instance, each group of inseparable paths moving in the same direction is weighted and normalized to output the predicted trajectory of the moving target.

2. The method for recognizing and predicting the trajectory of a moving target based on computer vision according to claim 1, characterized in that, The observed trajectory points of the moving target and road grid data containing map error information are acquired to generate multiple passable paths, including: In the directed routing graph model constructed from the road grid data, a multi-path search algorithm is used to search for a preset number of alternative routes starting from the current location; wherein, the path cost function used by the multi-path search algorithm is composed of a weighted sum of the geometric length of the road segment and the traffic rule constraint penalty; The N candidate routes with the smallest path cost function values ​​are selected as the passable paths.

3. The method for recognizing and predicting the trajectory of a moving target based on computer vision according to claim 1, characterized in that, Based on the map error information, the spatial tolerance range is defined, and the shared guiding road segments between the passable paths are identified, including: For each observation trajectory point, the trace A of the perception covariance matrix corresponding to the observation trajectory point and the trace B of the map geometric covariance matrix of the corresponding position in the road grid data are calculated. The sum of trace A and trace B is taken as the square root and multiplied by a preset scaling factor to obtain the single-point error scale. The arithmetic mean of the single-point error scales of all observation trajectory points is calculated to obtain the spatial tolerance range. For any two of the passable paths, the center lines of the corresponding road segments are spliced ​​together to form a continuous curve. The maximum arc length distance is calculated from the starting point, provided that the geometric Euclidean distance between the two continuous curves is continuously less than or equal to the spatial tolerance range. The maximum arc length distance is taken as the length of the shared guide road segment between the passable paths.

4. The computer vision-based moving target trajectory recognition and prediction method according to claim 3, characterized in that, Paths whose shared guide segment length exceeds the spatial tolerance range are classified into a group of inseparable paths in the same direction, including: Set a historical observation window; Execution of determination process: For any two of the said passable paths, determine whether the length of the shared guiding road segment is greater than or equal to the length of the historical observation window; if so, determine that the two passable paths belong to the same group of inseparable paths in the same direction; Perform a determination process on all the passable paths, and count the number of groups of the final indivisible path groups in the same direction.

5. The method for recognizing and predicting the trajectory of a moving target based on computer vision according to claim 1, characterized in that, Based on the number of groups in the indivisible path group and the lateral position deviation of each group at the far end of the prediction, a confusion parameter is generated, including: The path centerline with the minimum global path cost function value is selected as the reference centerline. For each group of inseparable paths in the same direction, the path with the smallest path cost function within that group is selected as the representative path. Calculate the lateral positional deviation of the representative path relative to the reference centerline at the far-end time of prediction; calculate the variance of the lateral positional deviation of all the same-direction inseparable path groups; calculate the natural logarithm of the number of groups of the same-direction inseparable path groups as the category indistinguishable information. The confusion parameter is obtained by multiplying the category-indistinguishable information by the variance.

6. The method for recognizing and predicting the trajectory of a moving target based on computer vision according to claim 1, characterized in that, Modulating the sampling weights using the confusion parameter includes: The confusion parameter is converted into a topology information temperature parameter using a linear mapping function, such that the higher the confusion, the larger the topology information temperature parameter. Calculate the median length of the shared guide segment for all paired paths, and use it as the typical shared guide segment length; For each observation trajectory point, the difference between the arc length along the trajectory point and the length of the typical shared guide road segment is calculated. This difference is divided by the topology information temperature parameter and then exponentially calculated to obtain the initial weight of the observation trajectory point. The initial weights of all observation trajectory points are normalized to obtain the topology separation sensitive weight.

7. The method for recognizing and predicting the trajectory of a moving target based on computer vision according to claim 6, characterized in that, Based on the sampling weights and a preset fixed sequence, a sampling consistency fitting operation is performed to select the optimal motion model instance, including: Set a fixed number of candidate motion model instances; For each candidate motion model instance, the hash mapping value of the identifier corresponding to the motion target is used as a seed to initialize a preset fixed sequence generator and generate a pseudo-random numerical sequence. The pseudo-random numerical sequence is mapped to an observation point index using the cumulative distribution function of the topological separation sensitive weights, thereby extracting the minimum sample set; Fit a polynomial motion curve to the minimum sample set; The standard deviation of the residuals is expanded and modulated using the confusion parameter to calculate the fitting residuals of all observation points relative to the polynomial motion curve. The likelihood score is calculated based on the fitting residuals, and the candidate motion model instance with the highest likelihood score is selected as the optimal motion model instance.

8. The method for recognizing and predicting the trajectory of a moving target based on computer vision according to claim 1, characterized in that, Based on the scoring results of the optimal motion model instance, each group of inseparable paths moving in the same direction is weighted and normalized to output the predicted trajectory of the moving target, including: For each of the indivisible path groups in the same direction, a corresponding path group score is calculated; wherein the path group score is obtained by subtracting a separation penalty term from the likelihood score of the optimal motion model instance, wherein the separation penalty term is proportional to the square of the difference between the lateral position deviation of the indivisible path group in the same direction and the mean, and inversely proportional to the topology information temperature parameter. The path group scores of all the aforementioned inseparable path groups in the same direction are subjected to exponential calculation and normalization to obtain the normalized weight of each group. The longitudinal and lateral motion parameters of the optimal motion model instance are mapped to the center line of the representative path of each of the same direction inseparable path groups to generate the intra-group mapping trajectory. The normalized weights are used to perform a linear weighted summation of all the intra-group mapping trajectories to obtain the final predicted trajectory of the moving target.

Images