CN-121973812-A - Track prediction method and related device

CN121973812ACN 121973812 ACN121973812 ACN 121973812ACN-121973812-A

Abstract

The invention discloses a track prediction method and a related device, and relates to the technical field of track prediction, wherein the track prediction method comprises the steps of obtaining target scene data and prediction reference data required by current prediction moment prediction, wherein the target scene data at least comprises historical observation track data of a moving object in a target scene, and the prediction reference data at least comprises a track control point determined by the last prediction moment for the moving object; the method comprises the steps of predicting target parameter deviation of a moving object according to target scene data and prediction reference data, determining the change amount of parameters of a future track compared with the previous prediction time, determining a future track control point of the moving object according to the target parameter deviation amount of the moving object, and generating the future track data of the moving object according to the future track control point of the moving object. The track prediction method disclosed by the invention has higher prediction stability, and can meet the requirement on track prediction stability in actual application scenes.

Inventors

WANG DONGSHENG
SUN PEIKUN
HU JINSHUI
GUO TAO

Assignees

科大讯飞股份有限公司

Dates

Publication Date: 20260505
Application Date: 20260306

Claims (16)

1. A track prediction method, comprising: Acquiring target scene data and prediction reference data required by prediction at the current prediction moment, wherein the target scene data comprises related data of a plurality of semantic instances in a target scene, the plurality of semantic instances at least comprise a plurality of moving objects of a track to be predicted, the related data at least comprise historical observation track data of each moving object, and the prediction reference data at least comprise track control points determined for each moving object at the last prediction moment; predicting a target parameter deviation amount of each moving object according to the target scene data and the prediction reference data, wherein the target parameter deviation amount is a change amount of a target parameter compared with the previous prediction time, and the target parameter is a parameter for determining a future track; determining a track control point of each moving object in a future period according to the target parameter deviation amount of each moving object; and generating track data of each moving object in a future period according to the track control point of each moving object in the future period.
2. The trajectory prediction method according to claim 1, wherein the prediction reference data further includes an inertial reference control point of each of the moving objects; the inertial reference control point is a control point of an inertial reference track, and the inertial reference track is a future track which is calculated by continuing the motion state of the current prediction moment.
3. The trajectory prediction method according to claim 1 or 2, characterized in that the target parameter deviation amount is a control point deviation amount, which is a change amount of a future trajectory control point compared to the previous prediction time; the determining a track control point of each moving object in a future period according to the target parameter deviation of each moving object comprises the following steps: And determining the track control point of each moving object in a future period according to the control point deviation amount of each moving object and the track control point determined for each moving object at the last prediction moment.
4. The trajectory prediction method according to claim 2, wherein the target parameter deviation amount is a control point residual deviation amount, the control point residual deviation amount is a variation amount of a control point residual compared with a previous prediction time, and the control point residual is a residual of a future trajectory control point to be determined and an inertial reference control point; the determining a track control point of each moving object in a future period according to the target parameter deviation of each moving object comprises the following steps: And determining the track control point of each moving object in a future period according to the control point residual error deviation amount of each moving object, the control point residual error determined for each moving object at the previous prediction time and the inertia reference control point acquired for each moving object at the current prediction time.
5. The trajectory prediction method according to claim 2, wherein acquiring the inertial reference control point for each of the moving objects comprises: For each of the moving objects: According to the speed of the moving object at the current prediction moment, estimating the track of the moving object in a future period through a uniform speed model to obtain an inertial reference track of the moving object; And converting the inertia reference track of the moving object into a track control point to obtain the inertia reference control point of the moving object.
6. The track prediction method according to claim 5, wherein the predicting the track of the moving object in the future period according to the speed of the moving object at the current prediction time by using a uniform speed model to obtain the inertial reference track of the moving object includes: The speed of the moving object at the current predicted moment is subjected to noise disturbance for a plurality of times, so that a plurality of disturbed speeds are obtained; according to the speeds after the disturbance, a plurality of tracks of the moving object in a future period are estimated through a uniform speed model, so that a plurality of inertial reference tracks of the moving object are obtained; The step of converting the inertia reference track of the moving object into a track control point to obtain the inertia reference control point of the moving object comprises the following steps: And respectively converting the plurality of inertial reference tracks into track control points to obtain a plurality of groups of inertial reference control points of the moving object.
7. The trajectory prediction method according to claim 1, wherein the determining a trajectory control point of each of the moving objects in a future period according to the target parameter deviation amount of each of the moving objects includes: normalizing the target parameter deviation of each moving object by utilizing a predetermined maximum target parameter deviation and a predetermined minimum target parameter deviation to obtain a normalized target parameter deviation of each moving object; and determining the track control point of each moving object in a future period according to the normalized target parameter deviation amount of each moving object.
8. The trajectory prediction method according to claim 1, wherein the historical observation trajectory data of any one of the moving objects is historical observation trajectory data in a local reference coordinate system centered on a current position of the moving object; the plurality of semantic instances further includes a plurality of static scene elements in the target scene; The related data further comprises map data of each static scene element and relative pose data among the plurality of semantic instances, wherein the map data of any static scene element is map data in a local reference coordinate system with key points of the static scene element as centers.
9. The trajectory prediction method according to claim 8, wherein predicting the target parameter deviation amount of each of the moving objects from the target scene data and the prediction reference data includes: Extracting characteristics from the data of each semantic instance in the target scene data to obtain instance characteristics of each semantic instance; extracting features from the relative pose data among the plurality of semantic instances to obtain relative pose features among the plurality of semantic instances; Enhancing the instance characteristics of each semantic instance by utilizing the prediction reference characteristics, the instance characteristics of each semantic instance and the relative pose characteristics among the plurality of semantic instances to obtain instance enhancement characteristics of each semantic instance; predicting the target parameter deviation amount of each moving object according to the instance enhancement characteristic of each moving object in the plurality of semantic instances.
10. The trajectory prediction method according to claim 1, wherein the trajectory control point of any one of the moving objects in the future period is a bezier control point; Generating track data of each moving object in a future period according to the track control point of each moving object in the future period, wherein the track data comprises the following steps: For each of the moving objects: Fitting a Bezier curve based on a Bezier basis function and Bezier control points of the moving object in a future period of time to serve as a future track curve of the moving object; Obtaining future speed data of the moving object by solving a first derivative of the future track curve, determining future course angle data of the moving object according to the future speed data, and obtaining future acceleration data of the moving object by solving a second derivative of the future track curve; The future track curve, the future speed data, the future heading angle data and the future acceleration data are taken as future track data of the moving object.
11. The trajectory prediction method according to claim 1, wherein predicting the target parameter deviation amount of each of the moving objects from the target scene data and the prediction reference data includes: Predicting the target parameter deviation of each moving object by taking the target scene data and the prediction reference data as prediction basis based on a target parameter deviation prediction model; The target parameter deviation prediction model is obtained by training a training sample marked with real target parameter deviation and real future track data, the training sample comprises a scene data sample and a prediction reference data sample, the training target of the target parameter deviation prediction model is that the target parameter deviation predicted according to the training sample is enabled to be consistent with the real target parameter deviation, and the future track data determined based on the predicted target parameter deviation is enabled to be consistent with the real future track data.
12. The trajectory prediction method of claim 11, wherein the plurality of semantic instances further comprises a plurality of static scene elements in the target scene, the related data further comprising map data for each of the static scene elements and relative pose data between the plurality of semantic instances; the target parameter deviation prediction model comprises an example encoder, a relative pose encoder, a control point encoder, a characteristic fusion module and a target parameter deviation prediction module; the predicting the target parameter deviation amount of each moving object based on the target parameter deviation amount prediction model by taking the target scene data and the prediction reference data as prediction basis comprises the following steps: Based on the instance encoder, encoding the data of each semantic instance in the target scene data to obtain instance characteristics of each semantic instance; based on the relative pose encoder, encoding the relative pose data among the plurality of semantic instances to obtain the relative pose characteristics among the plurality of semantic instances; encoding control points contained in the prediction reference data based on the control point encoder to obtain prediction reference characteristics; Based on the feature fusion module, enhancing the instance features of each semantic instance by utilizing the prediction reference features, the instance features of each semantic instance and the relative pose features among the plurality of semantic instances to obtain instance enhanced features of each semantic instance; and based on the target parameter deviation amount prediction module, predicting the target parameter deviation amount of each moving object by taking the instance enhancement characteristic of each moving object in the plurality of semantic instances as a prediction basis.
13. An electronic device comprising at least one processor and a memory coupled to the processor, wherein: The memory is used for storing a computer program; The processor is configured to execute the computer program to enable the electronic device to implement the trajectory prediction method according to any one of claims 1 to 12.
14. A computer storage medium carrying one or more computer programs which, when executed by an electronic device, enable the electronic device to implement a trajectory prediction method as claimed in any one of claims 1 to 12.
15. A computer program product comprising computer readable instructions which, when run on an electronic device, cause the electronic device to implement the trajectory prediction method of any one of claims 1 to 12.
16. An automatic driving vehicle is characterized by comprising an environment sensing unit, a track prediction unit and a decision planning unit; The environment sensing unit is used for acquiring surrounding environment data; The track prediction unit is configured to predict track data of a moving object around the vehicle in a future period according to the surrounding environment data by using the track prediction method according to any one of claims 1 to 12; and the decision planning unit is used for planning the driving path of the own vehicle according to the track data of the moving objects around the own vehicle in the future period.

Description

Track prediction method and related device Technical Field The present application relates to the field of track prediction technologies, and in particular, to a track prediction method and a related device. Background In some fields, it is required to predict future track data of moving objects, for example, in the field of automatic driving, it is required to predict future track data of moving objects (such as vehicles, pedestrians, etc.) around an own vehicle according to surrounding environment data sensed by an environment sensing unit of the automatic driving vehicle, so as to provide decision basis for the own vehicle. Most of the current track prediction schemes are track prediction methods based on discrete point coordinate regression, and the track prediction methods directly regress a series of three-dimensional track coordinate points at future discrete moments according to historical observation track data (such as track data perceived by an environment perception unit) of a moving object of a track to be predicted at each prediction moment. Although the track prediction scheme of the target can predict the future track of the moving object, the track prediction stability is poor, and the requirement on the track prediction stability in the actual application scene cannot be met. Disclosure of Invention In view of the above, the present application provides a track prediction method and related apparatus, which are used to solve the problem of poor track prediction stability in the current track prediction scheme, and the technical scheme is as follows: the first aspect of the present application provides a track prediction method, including: Acquiring target scene data and prediction reference data required by prediction at the current prediction moment, wherein the target scene data comprises related data of a plurality of semantic instances in a target scene, the plurality of semantic instances at least comprise a plurality of moving objects of a track to be predicted, the related data at least comprise historical observation track data of each moving object, and the prediction reference data at least comprise track control points determined for each moving object at the last prediction moment; predicting a target parameter deviation amount of each moving object according to the target scene data and the prediction reference data, wherein the target parameter deviation amount is a change amount of a target parameter compared with the previous prediction time, and the target parameter is a parameter for determining a future track; determining a track control point of each moving object in a future period according to the target parameter deviation amount of each moving object; and generating track data of each moving object in a future period according to the track control point of each moving object in the future period. In a possible implementation, the prediction reference data further includes an inertial reference control point for each of the moving objects; the inertial reference control point is a control point of an inertial reference track, and the inertial reference track is a future track which is calculated by continuing the motion state of the current prediction moment. In one possible implementation, the target parameter deviation amount is a control point deviation amount, and the control point deviation amount is a change amount of a future track control point compared to the previous prediction time; the determining a track control point of each moving object in a future period according to the target parameter deviation of each moving object comprises the following steps: And determining the track control point of each moving object in a future period according to the control point deviation amount of each moving object and the track control point determined for each moving object at the last prediction moment. In one possible implementation manner, the target parameter deviation amount is a control point residual deviation amount, where the control point residual deviation amount is a change amount of a control point residual compared with the previous prediction time, and the control point residual is a residual of a future track control point to be determined and an inertial reference control point; the determining a track control point of each moving object in a future period according to the target parameter deviation of each moving object comprises the following steps: And determining the track control point of each moving object in a future period according to the control point residual error deviation amount of each moving object, the control point residual error determined for each moving object at the previous prediction time and the inertia reference control point acquired for each moving object at the current prediction time. In one possible implementation manner, acquiring an inertial reference control point of each moving object includes: For each of the movi