A target trajectory prediction method, computer equipment, machine-readable storage medium and motor vehicle

Abstract

The application discloses a target trajectory prediction method, computer equipment, a machine readable storage medium and a motor vehicle, and relates to the technical field of automatic driving, and comprises a modeling step, an interactive heat map is generated; a position prediction step, the position of a prediction target is determined; and a trajectory prediction step, the trajectory prediction step constructs driving force constraints and environmental force constraints, and predicts trajectory coordinates of a next time step to form a trajectory of the prediction target. The application greatly improves the accuracy of trajectory prediction.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, specifically to a target trajectory prediction method, computer equipment, machine-readable storage medium, and motor vehicle. Background Technology

[0002] In the field of autonomous driving, predicting the movement trends of targets around the vehicle is a crucial aspect supporting the operation of autonomous driving systems. In urban intersection traffic scenarios, when planning the optimal trajectory based on the vehicle's own state, it is necessary not only to consider the movement trajectories of surrounding vehicles but also to determine the movement trends of pedestrians and non-motorized vehicles, thereby effectively avoiding collisions.

[0003] Existing technologies transform the problem of predicting the trajectory of a behavioral target into a time-series prediction problem that infers the future trajectory point position using the target's past motion. Due to the randomness and uncertainty of the target's motion, trajectory prediction is shifting from simply predicting the future trajectory closest to the true value to a multimodal prediction of all possible future trajectories. However, such endpoint-oriented trajectory prediction methods rely heavily on the accuracy of the predicted endpoint; the deviation between the predicted endpoint and the actual trajectory endpoint significantly impacts the accuracy of the complete trajectory prediction. Existing kinematic model-based behavioral prediction methods contain excessively high model complexity, making them unable to fit large amounts of trajectory data, resulting in low prediction accuracy. Furthermore, most social interaction-based behavioral prediction methods do not adequately consider environmental information beyond interactions, neglecting effective information such as road structure, which can lead to predicted trajectories of targets around the vehicle deviating from the lane or passable area. Summary of the Invention

[0004] This invention aims to address one of the technical problems in related technologies to a certain extent. To this end, this invention provides a target trajectory prediction method that improves the accuracy of trajectory prediction.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A target trajectory prediction method is provided for predicting the trajectories of targets around a vehicle during autonomous driving operations in an urban intersection scenario. The target trajectory prediction method includes...

[0007] The modeling steps are as follows: obtain trajectory heatmaps based on satellite maps of the predicted target and urban intersection scene; perform semantic segmentation on the satellite map of the urban intersection scene according to the obstacle categories in the urban intersection scene to obtain scene semantic map; construct distance heatmaps based on the influence of other targets at different locations around the predicted target on the predicted target; filter other targets in the distance heatmap; and generate interaction heatmaps based on the filtered other targets.

[0008] The location prediction step generates an endpoint heatmap based on the trajectory heatmap, scene semantic map, and interaction heatmap of the predicted target, and determines the predicted location of the predicted target from the endpoint heatmap.

[0009] Trajectory prediction steps include:

[0010] Calculate the direction and magnitude of the expected velocity of the predicted target moving to the predicted position, and construct driving force constraints based on the direction and magnitude of the expected velocity;

[0011] Construct environmental force constraints based on the scene semantic graph;

[0012] A dynamic model under discrete conditions is established based on driving force constraints and environmental force constraints. The trajectory coordinates of the next time step are predicted based on the dynamic model under discrete conditions, thus forming the trajectory of the predicted target.

[0013] Optionally, the direction and magnitude of the expected velocity of the predicted target moving to the predicted position can be calculated using the following formula:

[0014]

[0015] in, Let be a scalar, representing the direction of the desired velocity. T is a scalar representing the magnitude of the desired velocity. fut For the future, X t The observed trajectory at the current time step, The predicted position is obtained by the position prediction step.

[0016] Optionally, the driving force constraint can be constructed based on the direction and magnitude of the desired velocity using the following formula:

[0017]

[0018] in, The driving force constraint at time step t. It is a vector, representing the desired velocity. W is a vector representing the velocity at the current time step. τ The network parameters for constructing driving force constraints; τ is the relaxation time, which is the time required for the predicted target to adjust its current velocity and reach the target again. The time interval; φ(·) represents the neural network.

[0019] Optionally, environmental force constraints can be constructed from the scene semantic graph using the following formula:

[0020]

[0021] in, Let E be the environmental force constraint at time step t, k be a natural number, and E be the environmental force constraint at time step t. tTo predict the position of stationary obstacles around the target, I env s is the segment of the scene semantic graph at time step t. env ×s env The image is of a certain size, ψ(·) is the image feature extraction network, ξ is the environmental force coefficient of the predicted stationary obstacles around the target, and W ξ Here are the network parameters for the image feature extraction network, and C is the number of obstacle categories.

[0022] Optionally, a dynamic model under discrete conditions can be established based on the driving force constraints and environmental force constraints using the following formula:

[0023]

[0024] Among them, F goal As a driving force constraint, F env The constraint is the environmental force, and Δt is the time interval.

[0025] Optionally, the trajectory coordinates for the next time step can be predicted based on the dynamic model under discrete states using the following formula:

[0026]

[0027] in, The trajectory coordinates for the next time step. The coordinates of the trajectory at the current time step.

[0028] Optionally, the target trajectory prediction method further includes a multi-stage training step, which employs a teacher forcing training strategy and includes: training the position prediction step, training the neural network φ(·), and jointly training the neural network φ(·) and the image feature extraction network ψ(·).

[0029] The technical solution provided by this invention fully considers effective information such as environmental information and road structure in addition to interaction, avoiding deviation of the predicted trajectory from the lane or passable area. It also combines deep learning model with social force model to efficiently utilize the predicted endpoint information and scene semantic information, thereby improving the accuracy of trajectory prediction.

[0030] This invention employs a multi-stage model training approach, reducing the dependency between various computational modules. This allows each computational module to improve performance through module replacement, solving the problem of high coupling between model modules in existing prediction methods, which leads to optimization difficulties. It fully learns all the effective information of the traffic scene, avoids the predicted endpoint from deviating from the true location, and improves the accuracy of the final trajectory prediction.

[0031] Furthermore, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the target trajectory prediction method described in any of the above-mentioned embodiments.

[0032] Meanwhile, a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the target trajectory prediction method described in any one of the above claims.

[0033] In addition, the present invention also provides a motor vehicle with an autonomous driving function. When the motor vehicle is operating the autonomous driving function, in an urban intersection scenario, the trajectory prediction method described in any one of the foregoing claims is used to predict the trajectory of targets around the vehicle.

[0034] These features and advantages of the present invention will be disclosed in detail in the following specific embodiments and accompanying drawings. The preferred embodiments or means of the present invention will be shown in detail in conjunction with the accompanying drawings, but are not intended to limit the technical solutions of the present invention. In addition, each of these features, elements and components appearing in the following text and drawings is a plurality of, and different symbols or numbers are used for convenience of representation, but all represent parts with the same or similar construction or function. Attached Figure Description

[0035] The present invention will be further described below with reference to the accompanying drawings:

[0036] Figure 1 This is a schematic diagram illustrating the construction of an interactive heatmap in an embodiment of the present invention;

[0037] Figure 2 This is a schematic diagram of location prediction in an embodiment of the present invention;

[0038] Figure 3 This is a schematic diagram of trajectory prediction in an embodiment of the present invention. Detailed Implementation

[0039] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described are intended to explain the present invention and should not be construed as limiting the invention.

[0040] The terms "an embodiment," "example," or "trademark" used in this specification refer to a particular feature, structure, or characteristic described in connection with the embodiment itself that may be included in at least one embodiment disclosed in this patent. The phrase "in an embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment.

[0041] Example:

[0042] This embodiment provides a target trajectory prediction method for predicting the trajectory of targets around a vehicle during autonomous driving in an urban intersection scenario. In short, the target trajectory prediction method provided in this embodiment includes three main steps: a modeling step, a location prediction step, and a trajectory prediction step. Each step is described in detail below.

[0043] The modeling process includes the following sub-steps:

[0044] Trajectory heatmaps are obtained based on the distances between each pixel and the predicted target in the satellite map of the urban intersection scene. The observed trajectory sequence of the predicted target at the starting time is then transformed into a trajectory heatmap sequence. It should be noted here that, in this embodiment, the difference between a heatmap and a heatmap sequence is that a heatmap sequence is the superposition of heatmaps from several time steps. In other words, a heatmap is the heatmap corresponding to a specific time step in a heatmap sequence, while a heatmap sequence is the collection of all heatmaps within the prediction time step.

[0045] The predicted trajectory sequence of the target at the initial time is as follows:

[0046]

[0047] Among them, X tg Let t0 be the observed trajectory sequence of the target predicted at the start time, tg be the predicted target, and T be the predicted target. obs For historical time steps, (x tg ,y tg () represents the coordinates of the observed trajectory. It should be noted that in this embodiment, the superscripts in the upper right corner of all parameters are not exponents or powers, but rather refer to the dimension of the data, and the symbols... The meanings of number field and time step are both general technical meanings in this field.

[0048] The trajectory heatmap sequence has T obs One channel, expressed as:

[0049]

[0050] Where P is the trajectory heatmap sequence; and the trajectory heatmap is:

[0051]

[0052] Among them, P t Let be the expression for the trajectory heatmap at time t. For each pixel (p) in the satellite map x ,p y ) to the location of the predicted target tg The distance, where i and j are natural numbers, and max d tThe location of each pixel in the satellite map relative to the predicted target (tg). The maximum distance, w is the width of the trajectory heatmap, and h is the height of the trajectory heatmap.

[0053] Based on obstacle categories in urban intersection scenarios, semantic segmentation is performed on satellite maps of urban intersection scenes to classify scene constraints. Obstacle categories in urban intersection scenes include: roads, zebra crossings, sidewalks, buildings, greenery, and other obstacles. If the predicted target is a vehicle or non-motorized vehicle, its future trajectory is highly likely to lie within the passable area of ​​the road. If the predicted target is a pedestrian, when walkable areas such as sidewalks and zebra crossings exist, the pedestrian's walking range will also be fixed within a limited area. Pedestrian targets may be attracted to buildings and enter their interiors, or they may walk parallel to the outer edge of buildings; when there is greenery or obstacles in the pedestrian's path, they will choose to change their path to actively avoid them. Therefore, adding scene constraint information can effectively narrow the predicted range of the endpoint and improve accuracy. Semantic segmentation is performed using a U-Net network structure, which is existing technology and will not be elaborated here. The resulting scene semantic map... Where C represents the number of obstacle categories. Not all intersections contain the aforementioned six types of obstacles; therefore, different intersections may contain different types of obstacles. The resulting scene semantic map S1 corresponds to the trajectory heatmap, specifically, the scene semantic map S1 and the trajectory heatmap share the same pixel.

[0054] The social interaction between the predicted target tg and its surrounding targets is a crucial factor influencing downstream trajectory prediction. Therefore, the predicted target tg constantly monitors the movements of surrounding targets within its field of vision during its movement to plan its route in a timely manner and avoid collisions. Thus, this embodiment models social interaction based on three elements: time, location, and field of vision. The location and number of surrounding targets of the predicted target tg change over time. Based on this, a distance heatmap is constructed according to the influence of targets at different locations around the predicted target within historical observation time steps. The distance heatmap has a side length of S2 and the number of channels is T. obs One, the expression is:

[0055]

[0056] Where, N o The distance heatmap sequence provides the influence of targets at different locations around the predicted target tg on the predicted target tg during the historical observation period. Let be the distance heatmap at time t, and α be a user-defined hyperparameter, which is a parameter that will not be iterated in subsequent training steps. The setting of the hyperparameter is open to those skilled in the art and is not limited here. The hyperparameter α is used to control the influence of targets surrounding the predicted target tg.

[0057] Surrounding targets that influence the future trajectory of the predicted target tg are often located within the field of view. Therefore, it is necessary to screen all surrounding targets present in the intersection scene. In urban intersection scenes, the high density of traffic participants increases the number of surrounding targets. Therefore, a threshold m is set for the maximum number of surrounding targets. Then, the targets within the observation field of view are sorted by distance, and at most m targets at different locations around the predicted target are selected as valid targets. Based on the location of the selected targets and their influence on the predicted target, a field of view mask is generated using the following formula:

[0058]

[0059] Among them, M t For the view mask, ∈{0,1}, r M X represents the influence of the selected target on the predicted target. k The location of the selected target.

[0060] Then as Figure 1 As shown, based on the view mask, targets are selected again in the distance heatmap. Based on the selected targets, targets with negligible influence are filtered out, and interaction information of surrounding targets is used to generate an interaction heatmap. The formula for generating the interaction heatmap is:

[0061]

[0062] Where N is the interactive heatmap sequence, N t This is the interactive heatmap at time t. Hadamard product

[0063] At this point, the modeling process is complete. This embodiment models the predicted target and its surrounding environment as trajectory heatmaps, interaction heatmaps, and scene semantic maps, respectively. This improves the accuracy of predicting the future trajectories of targets around the vehicle, solves the problem of incomplete utilization of surrounding environmental information in existing behavior prediction methods, provides more effective input feature information for the model, and fully utilizes social interaction, environmental perception, and historical motion information to achieve multimodal future trajectory prediction of targets around the vehicle.

[0064] Then, the position prediction step is performed. This embodiment provides a special generative adversarial network as the endpoint prediction model, combining U-Net and PatchGAN models to construct a generator G and a discriminator D, predicting the endpoints of multiple possible future target trajectories. This provides additional target motion intent / trend information to the downstream trajectory prediction system, assisting the prediction system in predicting more accurate multimodal future trajectories. This ensures that the predicted endpoints are not limited to the constraints of the ground truth, exhibiting better multimodality and providing accurate data information for downstream trajectory prediction, thus guaranteeing the accuracy of downstream trajectory prediction. Figure 2 As shown, the location prediction step in this embodiment includes the following sub-steps:

[0065] The generator produces an endpoint heatmap based on the trajectory heatmap of the predicted target, the scene semantic map, and the interaction heatmap. The generator is a fully convolutional network with skip connections, and the formula for generating the endpoint heatmap is:

[0066]

[0067] Among them, O G The endpoint heatmap represents the probability distribution of the predicted endpoint of the target tg's future trajectory. env The input to the generator is W, where concat is the cascading operation. g Let G be the network parameters of the generator. In the above formula, G only represents the mapping from the generator's input and network parameters to the endpoint heatmap. The specific mapping relationship and the generator's network parameters are general technical meanings in this field and will not be elaborated here.

[0068] Compared to 2D trajectory sequence data, multi-channel heatmap data offers more intuitive sharing of underlying information between input and output, and the shared information maps to a wider range of features. Passing this shared information within the model effectively preserves the spatial details of intermediate feature maps. This embodiment uses a U-Net structure with skip connections as the generator to pass these shared features.

[0069] Specifically, in this embodiment, 2D convolution, activation functions, and pooling layers are the main components of the generator, and the generator's input... Features are extracted via convolutional layers, and their spatial scale is reduced via pooling layers until the bottleneck layer is reached. The decoder then upsamples the features from the bottleneck layer to reconstruct the prediction result. The entire process is as follows:

[0070] O G =G(H env W G ).

[0071] Skip connections are added to the encoder-decoder structure to pass processed features before and after the bottleneck layer. As shown in the formula below, shallow features of the i'-th layer decoder and the (n-i'+1)-th layer encoder are connected by a concat concatenation operation.

[0072]

[0073] Where n is the number of decoder layers, G enc G dec To correspond to the outputs of the encoding and decoding layers, U represents the feature transfer process, and H... skip This indicates the output of the jump connection module.

[0074] The discriminator determines the predicted location of the target based on the endpoint heatmap, as follows:

[0075] Due to the endpoint heatmap O G Since specific future endpoint coordinates cannot be directly provided, this embodiment is based on the endpoint heatmap O. G Given the probability distribution, K points are sampled using the TTST method as the predicted locations for the final output.

[0076] Based on the probability of the predicted location distribution, at least 10,000 locations are sampled from the interactive heatmap, and the location information of the sampled locations is extracted.

[0077] The extracted location information was used to obtain K-1 cluster centers using the K-means clustering algorithm;

[0078] Combining the global coordinate normalization method, the cluster centers of K modes are output as the predicted locations according to the following formula:

[0079]

[0080] in, For the predicted location, softar and gmax are used for global coordinate normalization, and Kmeans is the K-means clustering algorithm used to calculate the mean of location coordinates based on the probability distribution of the interactive heatmap.

[0081] In addition, to further improve the accuracy of location prediction, this embodiment also includes a training sub-step in the location prediction step.

[0082] In this embodiment, the discriminator and the generator are adversarial, improving their ability to distinguish between real and fake data through adversarial training, thereby synchronously constraining the generator's generation performance. Simultaneously, during training, the network parameters of both the generator and the discriminator are updated synchronously, achieving adversarial training and improving the generator's reliability. In this embodiment, only the trained generator network is run. In the training sub-step, as shown in the following formula, in this embodiment, the discriminator uses a fully convolutional structure to extract features from the input image using convolution operations and attempts to classify the authenticity of each K×K patch. The discriminator determines the authenticity of the K×K regions according to the following formula:

[0083] O D =D(concat(P,O) G );W D )

[0084] Among them, O D W is the output for the true / false judgment. D These are the network parameters for the generator. In this embodiment, pixel loss is used to compare the pixel-level differences between the generated results and the ground truth data during implementation, thereby controlling the endpoint heatmap O. G The low-frequency characteristics. Complementing this, the discriminator is used for the endpoint heatmap O. G Its high-frequency characteristics allow for control over global differences in images.

[0085] The loss function for game training is:

[0086]

[0087] Where L is the loss function, L GAN For the endpoint heatmap O G Loss function for mid-to-high frequency characteristics, L BCE For the endpoint heatmap O G The loss function based on low-to-medium frequency characteristics jointly guides the predicted endpoint heatmap O. G To the true probability graph Y gd The convergence makes the predicted endpoint distribution approximate the actual endpoint distribution. gd To predict the future T of the target fut The time-step motion trajectory sequence is a Gaussian distribution heatmap with mean and standard deviation of 4. Predict the target's future T. fut The motion trajectory sequence of the time step is:

[0088]

[0089] This embodiment improves the utilization of input environment features and enhances the robustness of the model's position prediction in different scenarios by using a generative adversarial learning approach, thus solving the problem of poor generalization ability of existing prediction models.

[0090] Then, the trajectory prediction step is performed. In this embodiment, the trajectory prediction step is completed by two modules: the destination driving module and the environmental force module. The destination driving module predicts the destination driving force F that the target experiences when it directly moves towards its destination E without being affected by the surrounding environment. goal The environmental force module takes into account surrounding environmental factors and estimates the environmental force F that affects the target's motion. env Based on F goal F env A dynamic model is established based on the combined effects of the target's motion. The velocity and position of the predicted target at each prediction step are calculated recursively to obtain the final predicted trajectory. This embodiment combines the SFM cross-matrix structure with a deep neural network, making the predicted trajectory more consistent with the kinematic model. Figure 3 As shown, the trajectory prediction step includes the following sub-steps:

[0091] When there is a definite endpoint When the target moves to The process can be controlled by the speed of movement Position X and final driving force F goal Construct as a pointer The trajectory is a polyline. In the future, over time T... fut Assuming the target is not disturbed by its surroundings during its movement, it will reach the endpoint. Expected velocity direction for:

[0092]

[0093] Expected speed for:

[0094]

[0095] in, As a scalar, For the next time step speed driving force Depend on and the speed of the current time step Therefore, the driving force constraint is constructed based on the direction and magnitude of the desired velocity using the following formula:

[0096]

[0097] in, The driving force constraint at time step t. W represents the velocity at the current time step. τThe network parameters for constructing driving force constraints; τ is the relaxation time, which is the time required for the predicted target to adjust its current velocity and reach the target again. The time interval is fitted using a neural network φ(·). In this embodiment, φ(·) is a neural network consisting of an LSTM layer and an MLP layer. The input to the LSTM layer is a concatenated vector of velocity and position. MLP layer extracts predicted destination The features are then concatenated with the hidden states output from the LSTM layer, and finally passed through an MLP layer to output the value τ. Different input variables will cause τ to change; the smaller τ is, the greater the acceleration.

[0098] In this embodiment, the endpoint determines the speed and direction of the predicted target's short-term motion. However, the predicted target is influenced by the surrounding environment as it moves towards the endpoint, altering its existing motion state to ensure its continued safe journey. Therefore, simply estimating the impact of the endpoint on the predicted target's motion is insufficient to construct a complete motion model. For this reason, this embodiment defines the influence of scene constraints on the motion trajectory as the environmental force F. env Its construction will be determined by the position, category, and distance of each structure within the scene:

[0099] Construct environmental force constraints based on the scene semantic graph:

[0100]

[0101] in, Let E be the environmental force constraint at time step t, k be a natural number, and E be the environmental force constraint at time step t. t To predict the position of stationary obstacles around the target, I env s is the segment of scene semantic graph S1 at time step t. env ×s env The image size is given by ψ(·), which represents the image feature extraction network. In this embodiment, a network structure consisting of multi-layer residual modules and linear layers is preferred. ξ represents the environmental force coefficients of the predicted stationary obstacles around the target, and W... ξ Network parameters for image feature extraction networks.

[0102] Then, the dynamic model under discrete conditions is established based on the driving force constraints and environmental force constraints using the following formula:

[0103]

[0104] Among them, F goal As a driving force constraint, F env The constraint is the environmental force, and Δt is the time interval.

[0105] Finally, the trajectory coordinates for the next time step are predicted based on the dynamic model under discrete states using the following formula:

[0106]

[0107] in, The trajectory coordinates for the next time step. Using the trajectory coordinates at the current time step, the trajectory of the predicted target is formed.

[0108] To address the problem of high coupling between model modules in existing prediction methods, which leads to optimization difficulties, the target trajectory prediction method provided in this embodiment further includes a multi-stage training step. This multi-stage training step employs a teacher forcing training strategy, including: a training sub-step, training of the neural network φ(·), and joint training of the neural network φ(·) and the image feature extraction network ψ(·). The first stage is the training sub-step in the position prediction step, where the optimal model is selected through testing. The second stage trains the neural network φ(·) in the endpoint driving module separately. The third stage trains the neural network φ(·) and the image feature extraction network ψ(·) together, achieving fine-tuning of the parameters of the neural network φ(·). In this embodiment, the model trained in each stage can be replaced with any higher-performance neural network model, allowing performance improvements to be made through module replacement, reducing the dependency between computational modules, and solving the problem of high coupling between model modules in existing prediction methods, which leads to optimization difficulties. In the second and third training phases, a teacher forcing training strategy was used to incorporate the ground truth of the predicted destination into the training, thereby learning all the effective information of the traffic scene, avoiding deviation of the predicted destination from the true location, and improving the accuracy of the final trajectory prediction.

[0109] Meanwhile, this embodiment also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the target trajectory prediction method described above. The steps of the target trajectory prediction method here can be the steps in the memory analysis methods of the various embodiments described above.

[0110] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. Accordingly, the computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can implement the methods of any of the above embodiments. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0111] Furthermore, this embodiment also provides a motor vehicle with an autonomous driving function. When the motor vehicle provided in this embodiment is operating the autonomous driving function, in an urban intersection scenario, the position of the target around the vehicle is predicted by the aforementioned target trajectory prediction method.

[0112] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Those skilled in the art should understand that the present invention includes, but is not limited to, the contents described in the accompanying drawings and the specific embodiments above. Any modifications that do not depart from the functional and structural principles of the present invention will be included within the scope of the claims.

Claims

1. A target trajectory prediction method for predicting the trajectory of targets around a vehicle during autonomous driving in an urban intersection scenario, characterized in that: The target trajectory prediction method includes, The modeling steps are as follows: obtain trajectory heatmaps based on satellite maps of the predicted target and urban intersection scene; perform semantic segmentation on the satellite map of the urban intersection scene according to the obstacle categories in the urban intersection scene to obtain scene semantic map; construct distance heatmaps based on the influence of other targets at different locations around the predicted target on the predicted target; filter other targets in the distance heatmap; and generate interaction heatmaps based on the filtered other targets. The location prediction step generates an endpoint heatmap based on the trajectory heatmap, scene semantic map, and interaction heatmap of the predicted target, and determines the predicted location of the predicted target from the endpoint heatmap. Trajectory prediction steps include: Calculate the direction and magnitude of the expected velocity of the predicted target moving to the predicted position, and construct driving force constraints based on the direction and magnitude of the expected velocity; Construct environmental force constraints based on the scene semantic graph; A dynamic model under discrete conditions is established based on driving force constraints and environmental force constraints. The trajectory coordinates of the next time step are predicted based on the dynamic model under discrete conditions, thus forming the trajectory of the predicted target.

2. The target trajectory prediction method according to claim 1, characterized in that, The direction and magnitude of the expected velocity of the target moving to the predicted position are calculated using the following formula: in, Let be a vector, representing the direction of the desired velocity. Let be a scalar representing the magnitude of the desired velocity, and t be the current time step. For the future, This refers to the observation trajectory at the current time step, where the observation trajectory is a two-dimensional coordinate system. The predicted position is obtained by the position prediction step, and the predicted position is a two-dimensional coordinate.

3. The target trajectory prediction method according to claim 2, characterized in that, The driving force constraint is constructed based on the direction and magnitude of the desired velocity using the following formula: in, for t Time-step driving force constraints, It is a vector, representing the desired velocity. It is a vector representing the velocity at the current time step. To construct the network parameters for driving force constraints; The relaxation time is the time required for the predicted target to adjust its current speed and re-reach the target. The time interval; It is a neural network.

4. The target trajectory prediction method according to claim 3, characterized in that, The environmental force constraints are constructed based on the semantic graph of the scene using the following formula: in, for t Environmental force constraints of time step, k For natural numbers, To predict the positions of stationary obstacles around the target, for t Time step is extracted from the scene semantic map Image of size For image feature extraction networks, To predict the environmental force coefficients of stationary obstacles around the target, Network parameters for image feature extraction networks, C This represents the number of obstacle categories.

5. The target trajectory prediction method according to claim 4, characterized in that, The following formula is used to establish a dynamic model under discrete conditions based on driving force constraints and environmental force constraints: in, As a driving force constraint, Due to environmental constraints, For time intervals.

6. The target trajectory prediction method according to claim 5, characterized in that, The trajectory coordinates for the next time step are predicted using the following formula based on the dynamic model under discrete states: in, The trajectory coordinates for the next time step. The coordinates of the trajectory at the current time step.

7. The target trajectory prediction method according to claim 4, characterized in that, The target trajectory prediction method further includes a multi-stage training step, which employs... teacherforcing The training strategy includes: training the location prediction step, and training the neural network. Training and the neural network Image feature extraction network Joint training.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the target trajectory prediction method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the target trajectory prediction method according to any one of claims 1 to 7.

10. A motor vehicle, characterized in that, The motor vehicle has an autonomous driving function. When the motor vehicle is operating the autonomous driving function, in an urban intersection scenario, the trajectory prediction method described in any one of claims 1 to 7 is used to predict the trajectory of targets around the vehicle.

Images