Obstacle motion trail prediction method and device

Abstract

The invention provides a method for predicting a movement track of an obstacle. The method comprises the following steps: step 1, acquiring perceived obstacle data and priori map information data; 2, judging the type of the obstacle; step 3, extracting map features and object features; step 4, coding all the feature elements through a vectorization method, and uniformly splicing the feature elements into a set composed of broken lines; step 5, encoding each broken line in the vectorization scene into a local coordinate system; step 6, constructing a symmetric model according to the local coordinate system, so that the self-attention mechanism is decoupled from the global coordinate system; step 7, performing multi-group intention query through the symmetric model, and synchronously predicting future trajectories of a plurality of obstacles; and step 8, according to the future trajectories of the plurality of obstacles and the queried behavior intention, generating a future trajectory more conforming to the scene.

Description

Technical Field

[0001] The present application relates to the field of autonomous driving technology, and in particular to a method and device for predicting obstacle motion trajectories. Background Art

[0002] In the existing autonomous driving framework, trajectory prediction serves as a bridge between perception and decision-making. Autonomous driving trajectory prediction involves analyzing various data from other pedestrians and vehicles as they travel to predict their future trajectories. Then, based on the game-playing between the autonomous vehicle and other vehicles, timely decisions are made to avoid collisions or warn of potential collisions. Trajectory prediction, a key technology for understanding how intelligent agents interact with dynamic environments, is crucial for improving their autonomy, adaptability, and safety. It not only helps intelligent agents avoid collisions but also improves the efficiency and safety of path planning. However, as the complexity of application scenarios increases, traditional trajectory prediction methods are unable to meet the demand for adaptable and generalizable trajectory prediction technology. Summary of the Invention

[0003] In view of this, the present invention provides a method and device for predicting the motion trajectory of an obstacle to solve technical problems such as inaccurate trajectory prediction and unrealistic feedback of the obstacle's future trajectory in the prior art.

[0004] The present invention provides an obstacle motion trajectory prediction method, which is applied to an intelligent driving vehicle. The method includes: step 1, obtaining perceived obstacle data and prior map information data; step 2, determining the obstacle category based on the perceived obstacle data; step 3, extracting map features and object features based on the obstacle category and the prior map information data; step 4, encoding all feature elements using a vectorization method and uniformly splicing them into a set of broken lines; step 5, encoding each broken line in the vectorized scene into a local coordinate system, and the local coordinate system is centered on each broken line and is independent of the global coordinate system; step 6, constructing a symmetric model based on the local coordinate system to decouple the self-attention mechanism from the global coordinate system; step 7, performing multiple groups of intent queries through the symmetric model to synchronously predict the future trajectories of multiple obstacles; and step 8, generating a future trajectory that better conforms to the scene based on the future trajectories of the multiple obstacles and the queried behavioral intentions.

[0005] Furthermore, the obstacle categories include motor vehicles, non-motor vehicles and pedestrians.

[0006] Furthermore, the priori map information includes basic road element information.

[0007] Furthermore, step 5 includes: step 51, obtaining the current position and direction of the obstacle; step 52, obtaining the broken line corresponding to the map feature element, and the collective center and tangent direction of the broken line; step 53, using the current position and direction of the obstacle, and the collective center and tangent direction of the broken line as a reference system to establish a local coordinate system.

[0008] Furthermore, step 7 also includes: identifying the behavioral intention of the obstacle through global intention positioning to enhance overall efficiency.

[0009] Furthermore, step 7 also includes: after global intention positioning, performing local motion refinement to adaptively refine the predicted trajectory to improve accuracy.

[0010] The present invention also provides a device for predicting obstacle motion trajectories, which includes: a vehicle sensor component for acquiring perceived obstacle data and prior map information data to determine the obstacle category; a symmetric modeling module connected to the vehicle sensor component for extracting map features and object features based on the obstacle category and prior map information data, encoding all feature elements through a vectorization method, and uniformly splicing them into a set of broken lines, encoding each broken line in the vectorized scene into a local coordinate system, and the local coordinate system is centered on each broken line and is independent of the global coordinate system, and a symmetric model is constructed based on the local coordinate system; a joint motion decoder connected to the symmetric modeling module for performing multiple groups of intent queries through the symmetric model to synchronously predict the future trajectories of multiple obstacles; an intention query module connected to the joint motion decoder for generating a future trajectory that is more consistent with the scene based on the future trajectories of multiple obstacles and the queried behavioral intention.

[0011] Furthermore, the intention query module is also used to identify the behavioral intention of obstacles through global intention positioning to enhance overall efficiency.

[0012] Furthermore, the intention query module is also used to: perform local motion refinement after global intention positioning, and adaptively refine the predicted trajectory to improve accuracy.

[0013] Furthermore, the step 3 extracts map features and object features through a lossy compression algorithm.

[0014] The present invention provides a method and device for predicting the motion trajectory of an obstacle. This technical solution predicts the motion trajectory of an obstacle through symmetrical modeling and intent query. It is mainly used to solve the problems of low detection rate and high false detection rate due to existing methods, as well as the inaccurate prediction of the long-term "terminal trajectory" of obstacles by the prediction module in the autonomous driving system and the "binding failure" of the predicted trajectories of some traffic participants who do not act in accordance with traffic rules. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of an obstacle motion trajectory prediction method provided by the present invention;

[0016] Figure 2 1 is a flow chart of a method for encoding each polyline in a vectorized scene into a local coordinate system, provided by the present invention;

[0017] Figure 3 This is a flow chart of another obstacle motion trajectory prediction method provided by the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] Example 1

[0020] The present invention provides a method and device for predicting obstacle motion trajectory, which is applied to intelligent driving vehicles. The device includes a vehicle sensor component, a symmetric modeling module, a joint motion decoder, and an intention query module. The device operates through the following method steps, such as Figure 1 shown.

[0021] Step 1: Obtain perceived obstacle data and prior map information data;

[0022] Step 2: Determine the obstacle category based on the perceived obstacle data;

[0023] Step 3: Extract map features and object features based on obstacle categories and prior map information data;

[0024] Step 4: Encode all feature elements using a vectorization method and concatenate them into a set of polylines;

[0025] Step 5: Encode each polyline in the vectorized scene into a local coordinate system centered on each polyline and independent of the global coordinate system.

[0026] Step 6: Construct a symmetric model based on the local coordinate system to decouple the self-attention mechanism from the global coordinate system.

[0027] Step 7: Use the symmetric model to perform multiple intent queries and simultaneously predict the future trajectories of multiple obstacles.

[0028] In step 8, a future trajectory that better matches the scenario is generated based on the future trajectories of multiple obstacles and the query behavior intention.

[0029] The present invention provides a method and device for predicting the motion trajectory of an obstacle. This technical solution predicts the motion trajectory of an obstacle through symmetrical modeling and intent query. It is mainly used to solve the problems of existing methods such as low detection rate, high false detection rate, and inability to accurately predict long-term trajectory information. The inability to feedback the true future trajectory of the obstacle, which in turn affects the decision-making and planning of autonomous vehicles, resulting in uneven vehicle decision-making and planning results or significant safety risks, especially in the decision-making of non-motor vehicles and pedestrians, which poses a significant safety hazard.

[0030] Example 2

[0031] The present invention provides a method and device for predicting obstacle motion trajectory, which is applied to intelligent driving vehicles. The device includes a vehicle sensor component, a symmetric modeling module, a joint motion decoder, and an intention query module. The device operates through the following method steps, such as Figure 1 shown.

[0032] Step 1: Obtain perceived obstacle data and prior map information data;

[0033] The priori map information includes basic road element information.

[0034] Step 2: Determine the obstacle category based on the perceived obstacle data;

[0035] The obstacle categories include motor vehicles, non-motor vehicles and pedestrians.

[0036] Step 3: Extract map features and object features based on obstacle categories and prior map information data;

[0037] Map features and object features are extracted using a lossy compression algorithm. This lossy compression algorithm, PolylineEncoder, is a lossy compression algorithm used to encode a series of coordinates into a single string. It is used to encode polylines and Bing point compression using the PHPTraits algorithm. Its advantage is that its output string consists only of URL-safe characters, so it can be used as a query parameter without URL encoding. Furthermore, the floating-point precision is configurable, meaning the encoding accuracy can be adjusted as needed. This encoding method is particularly suitable for representing geographic paths in mapping applications.

[0038] Step 4: Encode all feature elements using a vectorization method and concatenate them into a set of polylines;

[0039] Step 5: Encode each polyline in the vectorized scene into a local coordinate system centered on each polyline and independent of the global coordinate system.

[0040] Assume that there are N a target agents, and their historical state sequences are expressed as Where g represents the global coordinate system corresponding to the target agent in the scene, T h represents the historical observation time, C a The length of the sequence representing the historical motion state, which consists of historical motion information such as position, motion direction, and speed. h The sequence will be completed using Zero-Padding. The road graph will be represented as where N m represents the number of polylines in the map, n represents the number of coordinate points contained in each polyline, C m Represents the attributes of each point, such as the point's position in the global coordinate system and the road type.

[0041] A coordinate transformation function T(·) is added here to transform the global coordinate system of the polyline feature into a local coordinate system centered on the polyline. Specifically, the position and driving direction of each agent at the last observation time (i.e., the current time) are used as the reference system of each agent's local coordinate system, and for the polylines corresponding to the map elements, the collective center of these polylines and the tangent direction are used as the reference system of the local coordinate system. The encoded features are recorded as as well as These two features capture the agent's historical motion information and the map's geometric information respectively.

[0042] Step 6: Construct a symmetric model based on the local coordinate system to decouple the self-attention mechanism from the global coordinate system.

[0043] Step 7: Use the symmetric model to perform multiple intent queries and simultaneously predict the future trajectories of multiple obstacles.

[0044] In step 8, a future trajectory that better matches the scenario is generated based on the future trajectories of multiple obstacles and the query behavior intention.

[0045] The present invention provides a method and device for predicting the motion trajectory of an obstacle. This technical solution predicts the motion trajectory of an obstacle through symmetrical modeling and intent query. It is mainly used to solve the problems of existing methods such as low detection rate, high false detection rate, and inability to accurately predict long-term trajectory information. The inability to feedback the true future trajectory of the obstacle, which in turn affects the decision-making and planning of autonomous vehicles, resulting in uneven vehicle decision-making and planning results or significant safety risks, especially in the decision-making of non-motor vehicles and pedestrians, which poses a significant safety hazard.

[0046] Example 3

[0047] The present invention provides a method and device for predicting obstacle motion trajectory, which is applied to intelligent driving vehicles. The device includes a vehicle sensor component, a symmetric modeling module, a joint motion decoder, and an intention query module. The device operates through the following method steps, such as Figure 3 shown.

[0048] Step 1: Obtain perceived obstacle data and prior map information data;

[0049] Step 2: Determine the obstacle category based on the perceived obstacle data;

[0050] Step 3: Extract map features and object features based on obstacle categories and prior map information data;

[0051] Step 4: Encode all feature elements using a vectorization method and concatenate them into a set of polylines;

[0052] Step 5: Encode each polyline in the vectorized scene into a local coordinate system centered on each polyline and independent of the global coordinate system.

[0053] Each polyline in the vectorized scene is encoded into its own local coordinate system centered on the polyline, that is, the position and driving direction of each target at the last observation moment (i.e. the current moment) are used as the reference system of each target's local coordinate system, and for the polylines corresponding to the map elements, the collective center and tangent direction of these polylines are used as the reference system of the local coordinate system. Figure 2 As shown, step 5 includes:

[0054] Step 51, obtaining the current position and direction of the obstacle;

[0055] Step 52: Obtain the polyline corresponding to the map feature element, and the center and tangent direction of the polyline;

[0056] Step 53: Using the current position and direction of the obstacle, as well as the center and tangent direction of the polyline as reference systems, a local coordinate system is established.

[0057] This method introduces the Transformer encoder-decoder structure to process complex environmental contexts and combines it with the local self-attention mechanism to effectively encode complex scene context information to predict multimodal future motion.

[0058] Step 6: Construct a symmetric model based on the local coordinate system to decouple the self-attention mechanism from the global coordinate system.

[0059] Since these broken line features are encoded in their respective local coordinate systems, they are all independent of the global coordinate system. The token features input into the model are all decoupled from the global coordinate system, so that the association of tokens can be symmetrically modeled in subsequent links, and the self-attention mechanism can be decoupled from the global coordinate system. The attention mechanism imitates the internal process of biological observation behavior, that is, a mechanism that aligns internal experience and external sensation to increase the observation precision of some areas. The attention mechanism can quickly extract important features of sparse data, and is therefore widely used in natural language processing tasks, especially machine translation. The self-attention mechanism is an improvement of the attention mechanism, which reduces dependence on external information and is better at capturing the internal correlation of data or features. For example, assuming that the i-th token is used as the query, first transform the coordinates and directions of all tokens to the local coordinate system of the i-th token, and obtain:

[0060]

[0061] where i∈{1,...,N a +N m}, and j∈Ω(i) represents the neighborhood tokens of querytoken. and Represents the relative position and direction of the jth token in the local coordinate system of the ith token, and is then calculated as follows

[0062]

[0063] Step 7: Use the symmetric model to perform multiple intent queries and simultaneously predict the future trajectories of multiple obstacles.

[0064] Step 7 also includes identifying the obstacle's behavioral intent through global intention localization to enhance overall efficiency. Following global intention localization, local motion refinement is then performed to adaptively refine the predicted trajectory to improve accuracy. This method improves both prediction accuracy and efficiency through two key processes: global intention localization and local motion refinement.

[0065] In step 8, a future trajectory that better matches the scenario is generated based on the future trajectories of multiple obstacles and the query behavior intention.

[0066] The method ultimately outputs multimodal future trajectories for multiple targets that are not only accurate but also meet the actual needs of the scenario. It demonstrates enhanced performance and efficiency in predicting accurate multimodal future trajectories for multiple targets, especially as the number of targets increases.

[0067] The present invention provides a method and device for predicting the motion trajectory of an obstacle. This technical solution predicts the motion trajectory of an obstacle through symmetrical modeling and intent query. It is mainly used to solve the problems of existing methods such as low detection rate, high false detection rate, and inability to accurately predict long-term trajectory information. The inability to feedback the true future trajectory of the obstacle, which in turn affects the decision-making and planning of autonomous vehicles, resulting in uneven vehicle decision-making and planning results or significant safety risks, especially in the decision-making of non-motor vehicles and pedestrians, which poses a significant safety hazard.

[0068] In summary, the embodiments of the present invention provide a method and device for predicting obstacle motion trajectories. This technical solution can further improve the problem of inaccurate long-term obstacle trajectory prediction, thereby greatly improving the smoothness and ride comfort of autonomous driving vehicles; at the same time, it can also improve the robustness and generalization ability of the obstacle trajectory prediction algorithm, reducing the manpower and material costs of optimizing obstacle trajectory prediction results through traditional "parameter adjustment" methods; it can also predict the future trajectories of multiple targets through a single encoding and decoding model, thereby improving the real-time performance of the prediction algorithm; in addition, it can also improve the accuracy of obstacle trajectory prediction in complex scenarios (such as congested intersections), improve the overall safety and reliability of the autonomous driving system, and has advantages in inference speed and memory efficiency. In particular, when the model processes a large number of targets of interest, it can process data faster and use less computing resources.

[0069] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting obstacle motion trajectories, applied to intelligent driving vehicles, characterized in that: The method comprises: Step 1: Obtain perceived obstacle data and prior map information data; Step 2: Determine the obstacle category based on the perceived obstacle data; Step 3: Extract map features and object features based on obstacle categories and prior map information data; Step 4: Encode all feature elements using a vectorization method and concatenate them into a set of polylines; Step 5: Encode each polyline in the vectorized scene into a local coordinate system according to the following formula. The local coordinate system is centered on each polyline and is independent of the global coordinate system. in, are the object features encoded, is the coded map feature, T(·) is the coordinate transformation function, is the historical state representation of the obstacle, is the map feature representation, g represents the global coordinate system corresponding to the target agent in the scene, T h represents the historical observation time, C a Indicates the length of the sequence of historical motion states, N m represents the number of polylines in the map, n represents the number of coordinate points contained in each polyline, C m Represents the attributes of each point; Step 6: Construct a symmetric model based on the local coordinate system to decouple the self-attention mechanism from the global coordinate system. Step 7: Use the symmetric model to perform multiple intent queries and simultaneously predict the future trajectories of multiple obstacles. In step 8, a future trajectory that better matches the scenario is generated based on the future trajectories of multiple obstacles and the query behavior intention.

2. The obstacle trajectory prediction method according to claim 1, characterized in that: The obstacle categories include motor vehicles, non-motor vehicles and pedestrians.

3. The obstacle motion trajectory prediction method according to claim 1, characterized in that: The priori map information includes basic road element information.

4. The obstacle motion trajectory prediction method according to claim 1, characterized in that: The step 5 comprises: Step 51, obtaining the current position and direction of the obstacle; Step 52: Obtain the polyline corresponding to the map feature element, and the center and tangent direction of the polyline; Step 53: Using the current position and direction of the obstacle, as well as the center and tangent direction of the polyline as reference systems, a local coordinate system is established.

5. The obstacle motion trajectory prediction method according to claim 1, characterized in that: The step 7 also includes: identifying the behavioral intention of the obstacle through global intention positioning to enhance overall efficiency.

6. The obstacle motion trajectory prediction method according to claim 5, characterized in that: The step 7 also includes: after global intention localization, performing local motion refinement to adaptively refine the predicted trajectory to improve accuracy.

7. A device for implementing the obstacle motion trajectory prediction method according to claims 1-6, characterized in that: The device comprises: Vehicle sensor components, used to obtain perceived obstacle data and prior map information data to determine obstacle categories; The symmetry modeling module is connected to the vehicle sensor assembly and is used to extract map features and object features based on obstacle categories and prior map information data. All feature elements are encoded using a vectorization method and unified into a set of polylines. Each polyline in the vectorized scene is encoded into a local coordinate system. This local coordinate system is centered on each polyline and is independent of the global coordinate system. A symmetric model is constructed based on the local coordinate system. The joint motion decoder, connected to the symmetric modeling module, is used to perform multiple sets of intent queries using the symmetric model and simultaneously predict the future trajectories of multiple obstacles; The intention query module is connected to the joint motion decoder and is used to generate a future trajectory that is more consistent with the scenario based on the future trajectories of multiple obstacles and the queried behavioral intention.

8. The device for predicting obstacle motion trajectory according to claim 7, characterized in that: The intention query module is also used to identify the behavioral intention of obstacles through global intention positioning to enhance overall efficiency.

9. The device for predicting obstacle motion trajectory according to claim 7, characterized in that: The intention query module is also used to perform local motion refinement after global intention positioning, and adaptively refine the predicted trajectory to improve accuracy.

10. The obstacle motion trajectory prediction method according to claim 7, characterized in that: The symmetric modeling module extracts map features and object features through a lossy compression algorithm.

Images