Target position prediction method, computer device, machine-readable storage medium, and motor vehicle

Abstract

The present invention discloses a target position prediction method, computer equipment, machine-readable storage medium, and motor vehicle, relating to the field of autonomous driving technology. The method includes a modeling step and a position prediction step. The modeling step includes obtaining a trajectory heat map based on the distance between each pixel in a satellite map and a predicted target, converting the observed trajectory sequence of the predicted target at the start moment into a trajectory heat map sequence; performing semantic segmentation on a satellite map of an urban intersection scene to obtain a scene semantic map; constructing a distance heat map based on the influence of targets at different locations on the predicted target; selecting surrounding targets and generating a field of view mask based on their position and influence; selecting and reselecting targets in the distance heat map based on the field of view mask to generate an interaction heat map; and a position prediction step including a generator generating a destination heat map; and a discriminator determining the predicted location of the predicted target based on the destination heat map. The present invention significantly improves the accuracy of destination prediction.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a target position prediction method, computer equipment, machine-readable storage medium, and a motor vehicle. Background Art

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

[0003] The existing technology transforms the problem of predicting the trajectory of a behavioral target into a time series prediction problem that uses the past motion of the predicted target to infer the position of its future trajectory points. Due to the randomness and uncertainty of the motion of the predicted target, the predicted trajectory is changing from a single prediction of the future trajectory closest to the true value to a multimodal prediction of all possible future trajectories. However, this type of endpoint-oriented trajectory prediction method relies on the accuracy of the predicted endpoint. The deviation between the predicted endpoint and the true trajectory endpoint will have a huge impact on the prediction accuracy of the complete trajectory. The behavior prediction method based on the kinematic model in the existing technology contains too high a model complexity and cannot fit large amounts of trajectory data, so the prediction accuracy is low. Summary of the Invention

[0004] The present invention aims to solve one of the technical problems in the related art to a certain extent. To this end, the present invention provides a target position prediction method, which improves the accuracy of target position prediction.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A target position prediction method is used to predict the positions of targets around a vehicle when the vehicle is operating automatically at an urban intersection. The target position prediction method includes a modeling step and a position prediction step.

[0007] The modeling step includes the following sub-steps:

[0008] The trajectory heat map is obtained based on the distance between each pixel and the predicted target in the satellite map of the urban intersection scene, and the observed trajectory sequence of the predicted target at the starting time is converted into a trajectory heat map sequence;

[0009] Based on the obstacle categories in the urban intersection scene, semantic segmentation is performed on the satellite map of the urban intersection scene to classify the scene constraints and obtain a scene semantic map, wherein the scene semantic map corresponds to the trajectory heat map;

[0010] Construct a distance heat map based on the influence of targets at different locations around the predicted target within the historical observation time step on the predicted target;

[0011] Sort the targets in the observation field of view by distance, select several targets at different positions around the predicted target in sequence, and generate a field of view mask based on the positions of the selected targets and their influence on the predicted target;

[0012] Select and reselect targets in the range heat map based on the field of view mask, and generate an interaction heat map based on the selected targets;

[0013] The position prediction step includes the following sub-steps:

[0014] The generator generates a destination heatmap based on the predicted target trajectory heatmap, scene semantic map, and interaction heatmap;

[0015] The discriminator determines the predicted location of the predicted target based on the endpoint heat map.

[0016] Optionally, the observation trajectory sequence of the predicted target at the starting moment is:

[0017]

[0018] Among them, X tg is the observation trajectory sequence of the predicted target at the starting time, t0 is the starting time, tg is the predicted target, T obs is the historical time step, (x tg ,y tg ) are the coordinates of the observation trajectory;

[0019] The trajectory heat map sequence has T obs channels, the expression is:

[0020]

[0021] Where P is the trajectory heat map sequence;

[0022] The trajectory heat map is:

[0023]

[0024] Among them, P t is the expression of the trajectory heat map at time t, For each pixel (p x ,p y ) to the location of the predicted target tg The distance, i, j are natural numbers, maxd t The location of each pixel in the satellite map to the predicted target tg The maximum distance, w is the width of the trajectory heat map, and h is the height of the trajectory heat map.

[0025] Optionally, the obstacle categories in urban intersection scenes include: roads, zebra crossings, sidewalks, buildings, green spaces, and obstacles; semantic segmentation is performed by the U-Net network structure, and the resulting scene semantic map Where C is the number of obstacle categories; when the scene semantic map and the trajectory heat map are the same pixel, the scene semantic map corresponds to the trajectory heat map.

[0026] Optionally, the side length of the distance heat map is S2, and the number of channels is T obs The expression is:

[0027]

[0028] Among them, N o is the distance heat map sequence, is the distance heat map at time t, and α is a custom hyperparameter used to control the influence of targets around the predicted target tg.

[0029] Optionally, the formula for generating the field of view mask is:

[0030]

[0031] Among them, M t is the field of view mask,∈{0,1}, r M is the influence of the selected target on the predicted target, X k The location of the selected target.

[0032] Optionally, the formula for generating an interaction heatmap based on the selected target is:

[0033]

[0034] Among them, N is the interaction heat map sequence, N t is the interaction heat map at time t, is the Hadamard product

[0035] Optionally, the generator is a fully convolutional network with skip connections, and the formula for generating the end point heat map by the generator is:

[0036]

[0037] Among them, O G is the end point heat map, H env is the input of the generator, W g are the network parameters of the generator.

[0038] Optionally, the discriminator determines the predicted position of the predicted target according to the following formula, including the following sub-steps:

[0039] Sample at least 10,000 locations from the interaction heat map based on the probability of the predicted location distribution, and extract the location information of the sampled locations;

[0040] Use K-means clustering algorithm to obtain K-1 cluster centers on the extracted location information;

[0041] According to the following formula, the cluster centers of K modes are output as the predicted positions:

[0042]

[0043] in, is the predicted position, and softargmax is the global coordinate normalization, which is used to calculate the mean of the position coordinates based on the probability distribution of the interaction heat map.

[0044] Optionally, the position prediction step further includes a training sub-step, in which a discriminator discriminates the authenticity of K×K regions according to the following formula:

[0045] O D =D(concat(P,O G );W D )

[0046] Among them, O D is the output of the true or false judgment, W D are the network parameters of the generator.

[0047] Optionally, in the training sub-step, the loss function is:

[0048]

[0049] Among them, L is the loss function, L GAN The end point heat map O G The loss function of the mid- and high-frequency characteristics, L BCE The end point heat map O G Loss function for mid- and low-frequency characteristics, Y gd To predict the future T fut The time-step trajectory sequence is a Gaussian distribution heat map with a mean and standard deviation of 4, and the future T of the target is predicted. fut The motion trajectory sequence of the time step is:

[0050]

[0051] The technical solution provided by the present invention models the predicted target and its surrounding environmental information as trajectory heat maps, interaction heat maps and scene semantic maps respectively, providing more effective input feature information for the model, making full use of social interaction, environmental perception and historical motion information, and realizing multimodal future trajectory prediction of the vehicle's surrounding targets. Based on the existing endpoint-oriented prediction model, the present invention provides a method for screening and calculating the influence of surrounding targets, which increases the model's learning of important target information, screens out surrounding targets with negligible influence, and constructs the surrounding target interaction information into an interaction heat map. At the same time, the present invention adopts a special generative adversarial network as the endpoint prediction model, so that the predicted endpoint is not limited to the constraints of the true value, has better multimodality, provides accurate data information for downstream trajectory prediction, and ensures the accuracy of downstream trajectory prediction.

[0052] Furthermore, the present invention further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements any one of the above-mentioned target position prediction methods when executing the computer program.

[0053] At the same time, a computer-readable storage medium stores a computer program thereon, and when the computer program is executed by a processor, it implements any of the target position prediction methods described above.

[0054] In addition, the present invention also provides a motor vehicle having an automatic driving function. When the motor vehicle is running the automatic driving function, the position of targets around the vehicle is predicted by using any of the target position prediction methods described above.

[0055] These features and advantages of the present invention will be further disclosed in the following detailed description and accompanying drawings. The preferred embodiments and means of the present invention will be fully illustrated in conjunction with the accompanying drawings, but are not intended to limit the technical solutions of the present invention. Furthermore, although multiple features, elements, and components may be present in each of the following text and accompanying drawings, they may be labeled with different symbols or numbers for convenience, but all represent components with the same or similar structure or function. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The present invention will be further described below in conjunction with the accompanying drawings:

[0057] Figure 1 A schematic diagram of constructing an interactive heat map in the present invention;

[0058] Figure 2 Schematic diagram of position prediction in the present invention. DETAILED DESCRIPTION

[0059] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described in the embodiments are intended to explain the present invention and are not to be construed as limiting the present invention.

[0060] Reference in this specification to "one embodiment," "an example," or "an example" means that a particular feature, structure, or characteristic described in connection with the embodiment itself can be included in at least one embodiment of the present patent disclosure. The appearances of the phrase "in one embodiment" in various places in the specification are not necessarily all referring to the same embodiment.

[0061] Example:

[0062] This embodiment provides a target position prediction method for predicting the positions of objects around a vehicle during autonomous driving at an urban intersection. In short, the target position prediction method provided in this embodiment includes two major steps: a modeling step and a position prediction step. Each of these steps is described in detail below.

[0063] The modeling step includes the following sub-steps:

[0064] A trajectory heat map is obtained based on the distance between each pixel and the predicted target in the satellite map of the urban intersection scene, and the observed trajectory sequence of the predicted target at the starting time is converted into a trajectory heat map sequence. It should be noted here that in this embodiment, the difference between a heat map and a heat map sequence is that the superposition of heat maps of several time steps is a heat map sequence. In other words, a heat map is the heat map corresponding to a specific time step in the heat map sequence, while a heat map sequence is the collection of all heat maps within the predicted time step.

[0065] The observation trajectory sequence of the predicted target at the starting moment is:

[0066]

[0067] Among them, X tg is the observation trajectory sequence of the predicted target at the starting time, t0 is the starting time, tg is the predicted target, T obs is the historical time step, (x tg ,y tg ) is the coordinate of the observation trajectory. It should be pointed out here that in this embodiment, the superscripts in the upper right corner of all parameters are not powers or times, but refer to the dimensions of the data, and the symbol The meaning of the number field and the time step are both common technical meanings in this field.

[0068] The trajectory heatmap sequence has T obs channels, the expression is:

[0069]

[0070] Among them, P is the trajectory heat map sequence; and the trajectory heat map is:

[0071]

[0072] Among them, P t is the expression of the trajectory heat map at time t, For each pixel (p x ,p y ) to the location of the predicted target tg The distance, i, j are natural numbers, maxd t The location of each pixel in the satellite map to the predicted target tg The maximum distance, w is the width of the trajectory heat map, and h is the height of the trajectory heat map.

[0073] Based on the obstacle categories in urban intersection scenarios, semantic segmentation is performed on satellite maps of urban intersection scenarios to categorize scene constraints. Obstacle categories in urban intersection scenarios include six types: roads, zebra crossings, sidewalks, buildings, greenery, and obstacles. If the predicted target is a vehicle or non-motor vehicle, the future trajectory of the target is likely to be within the traversable area of ​​the road; if the predicted target is a pedestrian, when there are walkable areas such as sidewalks and zebra crossings, the range of pedestrian walking will also be fixed within a limited area. Pedestrian targets will be attracted by buildings and enter their interior, or they may walk parallel to the outer edge of the building; when there are greenery or obstacles on the pedestrian's movement path, they will choose to change the path to actively avoid them. Therefore, adding scene constraint information can effectively narrow the predicted range of the end point and improve accuracy. Semantic segmentation is performed by the U-Net network structure, which is an existing technology and will not be described here. The resulting scene semantic map Where C is the number of obstacle categories. Not all intersections have the six aforementioned obstacle types; therefore, different intersections may have 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 are pixel-by-pixel.

[0074] The social interactivity between the predicted target tg and its surrounding targets is an important factor affecting the downstream trajectory prediction, so the predicted target tg will always pay attention to the movements of the surrounding targets within the field of view during the movement, so as to plan its own route in time and avoid collisions. Therefore, this embodiment models social interactivity based on three elements: time, position, and field of view. The position and number of targets surrounding the predicted target tg will change over time. Based on this, a distance heat map is constructed based on the influence of targets at different positions around the predicted target on the predicted target within the historical observation time step. The side length of the distance heat map is S2, and the number of channels is Tobs The expression is:

[0075]

[0076] Among them, N o It is a distance heat map sequence, which provides the influence of targets at different positions around the predicted target tg on the predicted target tg during the historical observation time. is the distance heatmap at time t, and α is a custom hyperparameter, meaning it is not iterated in subsequent training steps. The setting of hyperparameters is flexibly selected by those skilled in the art based on practical circumstances and is not limited here. The hyperparameter α is used to control the influence of targets surrounding the predicted target tg.

[0077] The surrounding targets that affect the future motion trajectory of the predicted target tg are often within the field of view. Therefore, it is necessary to screen all the surrounding targets in the intersection scene. In urban intersection scenes, the higher density of traffic participants increases the number of surrounding targets. Therefore, the maximum number of surrounding targets is selected to set the threshold m: Then, the targets in the observation field of view are sorted according to the distance, and up to m targets at different positions around the predicted target are selected in order as valid targets. According to the position of the selected target and the influence of the selected target on the predicted target, the field of view mask is generated by the following formula:

[0078]

[0079] Among them, M t is the field of view mask,∈{0,1}, r M is the influence of the selected target on the predicted target, X k The location of the selected target.

[0080] Then as Figure 1 As shown in the figure, based on the field of view mask, the target is selected again in the distance heat map. Based on the selected target, the target with negligible influence is filtered out, and the interaction information of the surrounding targets is used to generate an interaction heat map. The formula for generating the interaction heat map is:

[0081]

[0082] Among them, N is the interaction heat map sequence, N t is the interaction heat map at time t, is the Hadamard product

[0083] This completes the modeling process. This embodiment models the predicted target and its surrounding environmental information as a trajectory heatmap, an interaction heatmap, and a scene semantic map, respectively. This improves the accuracy of future trajectory predictions for targets around the vehicle itself, resolving the issue of incomplete utilization of surrounding environmental information in existing behavior prediction methods. This provides the model with more effective input feature information, fully leveraging social interactions, environmental perception, and historical motion information to achieve multimodal future trajectory predictions for targets around the vehicle itself.

[0084] Then the position prediction step is performed. This embodiment provides a special generative adversarial network as an endpoint prediction model, combining the U-Net and PatchGAN models to construct a generator G and a discriminator D, predicting the endpoints of multiple possible future target motion trajectories, providing additional target motion intention / trend information for the downstream trajectory prediction system, and assisting the prediction system in predicting more accurate multimodal future trajectories, so that the predicted endpoint is not limited to the constraints of the true value and has better multimodality, providing accurate data information for downstream trajectory prediction, and ensuring the accuracy of downstream trajectory prediction. Figure 2 As shown, the position prediction step in this embodiment includes the following sub-steps:

[0085] The generator generates a destination heatmap based on the predicted target trajectory heatmap, scene semantic map, and interaction heatmap. The generator is a fully convolutional network with skip connections. The formula for generating the destination heatmap is:

[0086]

[0087] Among them, O G H is the endpoint heat map, which represents the probability distribution of the future trajectory endpoint of the predicted target tg. env is the input of the generator, concat is the cascade operation, W g are the network parameters of the generator. In the above formula, G only represents the mapping of the generator input and the generator network parameters to the destination heat map. The specific mapping relationship and the generator network parameters are common technical meanings in this field and will not be repeated here.

[0088] Compared to 2D trajectory sequence data, multi-channel heatmap data has more intuitive underlying information sharing between input and output, and the shared information maps a more comprehensive range of features. Propagating this shared information within the model effectively preserves the spatial details of the intermediate feature maps. This embodiment uses a U-Net structure with skip connections as the generator to deliver these shared features.

[0089] Specifically in this embodiment, 2D convolution, activation function and pooling layer are the main components of the generator. The convolutional layer extracts features, and the pooling layer reduces the spatial scale until the bottleneck layer. The decoder then upsamples the bottleneck layer features to restore the prediction results. The whole process is:

[0090] O G =G(H env ;W G ).

[0091] Skip connections are added to the encoder-decoder structure to transfer features processed before and after the bottleneck layer. As shown in the following formula, the shallow features of the i'th layer decoder and the (n-i'+1)th layer encoder are connected through the concat cascade operation.

[0092]

[0093] Where n is the number of decoder layers, G enc , G dec is the output of the corresponding encoding layer and decoding layer, U represents the process of feature transfer, H skip represents the output of the skip connection module.

[0094] The discriminator determines the predicted location of the target based on the endpoint heat map, as follows:

[0095] Since the end point heat map O G It is impossible to directly provide the specific target future destination coordinates, so this embodiment is based on the destination heat map. G The probability distribution provided uses the TTST method to sample K points as the predicted position of the final output.

[0096] Sample at least 10,000 locations from the interaction heat map based on the probability of the predicted location distribution, and extract the location information of the sampled locations;

[0097] Use K-means clustering algorithm to obtain K-1 cluster centers on the extracted location information;

[0098] Combined with the global coordinate normalization method, the cluster centers of K modes are output as the predicted positions according to the following formula:

[0099]

[0100] in, is the predicted position, softargmax is the global coordinate normalization, and Kmeans is the K-means clustering algorithm, which is used to calculate the mean of the position coordinates based on the probability distribution of the interaction heat map.

[0101] At the same time, in order to further improve the accuracy of position prediction, in this embodiment, the position prediction step also includes a training sub-step.

[0102] In this embodiment, the discriminator and the generator are the objects of mutual game, and through adversarial training, they improve their ability to distinguish true from false, and then synchronously constrain the generation effect of the generator. At the same time, during the training process, the network parameters of the generator and the discriminator are synchronously updated to realize game training and improve the reliability of the generator. During the implementation of this embodiment, only the generator network that has completed training is run. In the training sub-step, as shown in the following formula, in this embodiment, the discriminator uses a full convolution structure to extract the features of the input image with the help of convolution operations, and attempts to classify the authenticity of each K×K patch. The discriminator determines the authenticity of K×K areas according to the following formula:

[0103] O D =D(concat(P,O G );W D )

[0104] Among them, O D is the output of the true or false judgment, W D is the network parameter of the generator. In the implementation process, this embodiment uses pixel loss to compare the pixel-level difference between the generated result and the true value data to control the end point heat map. G Complementary to this, the discriminator is used to calculate the end point heat map O G The high-frequency characteristics of the image can be used to control the global differences of the image.

[0105] The loss function of game training is:

[0106]

[0107] Among them, L is the loss function, L GAN The end point heat map O G The loss function of the mid- and high-frequency characteristics, L BCE The end point heat map O G The loss function of the low- and medium-frequency features jointly guides the predicted endpoint heat map O G To the true probability graph Y gd Closer, so that the predicted endpoint distribution is close to the true endpoint distribution. gd To predict the future T fut The time-step trajectory sequence is a Gaussian distribution heat map with a mean and standard deviation of 4, and the future T of the target is predicted. fut The motion trajectory sequence of the time step is:

[0108]

[0109] This embodiment improves the utilization rate of input environment features through a generative adversarial learning method, improves the robustness of the model's position prediction in different scenarios, and solves the problem of poor generalization ability of existing prediction models.

[0110] At the same time, this embodiment also provides a computer device including a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor performs the steps of the target location prediction method described above. The steps of the target location prediction method here can be the steps of the memory analysis method in each of the above embodiments.

[0111] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program. Accordingly, the computer program can be stored in a non-volatile computer-readable storage medium, and when the computer program is executed, it can implement the method of any of the above-mentioned embodiments. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may 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 many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double 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), etc.

[0112] In addition, this embodiment also provides a motor vehicle with an automatic driving function. When the motor vehicle provided by this embodiment runs the automatic driving function, the position of targets around the vehicle is predicted by the aforementioned target position prediction method.

[0113] The above are only 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 drawings and the above specific embodiments. Any modifications that do not deviate from the functional and structural principles of the present invention are intended to be included within the scope of the claims.

Claims

1. A target position prediction method for predicting the positions of targets around a vehicle when the vehicle is operating autonomously at an urban intersection. The target position prediction method comprises a modeling step and a position prediction step, and is characterized by: The modeling step includes the following sub-steps: The trajectory heat map is obtained based on the distance between each pixel and the predicted target in the satellite map of the urban intersection scene, and the observed trajectory sequence of the predicted target at the starting time is converted into a trajectory heat map sequence; Based on the obstacle categories in the urban intersection scene, semantic segmentation is performed on the satellite map of the urban intersection scene to classify the scene constraints and obtain a scene semantic map, wherein the scene semantic map corresponds to the trajectory heat map; Construct a distance heat map based on the influence of targets at different locations around the predicted target within the historical observation time step on the predicted target; Sort the targets in the observation field of view by distance, select several targets at different positions around the predicted target in sequence, and generate a field of view mask based on the positions of the selected targets and their influence on the predicted target; Select and reselect targets in the range heat map based on the field of view mask, and generate an interaction heat map based on the selected targets; The position prediction step includes the following sub-steps: The generator generates a destination heatmap based on the predicted target trajectory heatmap, scene semantic map, and interaction heatmap; The discriminator determines the predicted location of the predicted target based on the endpoint heat map.

2. The target position prediction method according to claim 1, characterized in that: The observation trajectory sequence of the predicted target at the starting moment is: Among them, X tg is the observation trajectory sequence of the target predicted at the starting time, t0 is the starting time, t g is the prediction target, T obs is the historical time step, (x tg ,y tg ) are the coordinates of the observation trajectory; The trajectory heat map sequence has T obs channels, the expression is: Where P is the trajectory heat map sequence; The trajectory heat map is: Among them, P t is the expression of the trajectory heat map at time t, For each pixel (p x ,p y ) to the location of the predicted target tg The distance, i, j are natural numbers, max d t The location of each pixel in the satellite map to the predicted target tg The maximum distance, w is the width of the trajectory heat map, and h is the height of the trajectory heat map.

3. The target position prediction method according to claim 2, characterized in that: Obstacle categories in urban intersection scenes include: roads, zebra crossings, sidewalks, buildings, green spaces, and obstacles; semantic segmentation is performed by the U-Net network structure, and the resulting scene semantic map Where C is the number of obstacle categories; when the scene semantic map and the trajectory heat map are the same pixel, the scene semantic map corresponds to the trajectory heat map.

4. The target position prediction method according to claim 3, characterized in that: The side length of the distance heat map is S2, and the number of channels is T obs The expression is: Among them, N o is the distance heat map sequence, is the distance heat map at time t, and α is a custom hyperparameter used to control the influence of targets around the predicted target tg.

5. The target position prediction method according to claim 4, characterized in that: The formula for generating the field of view mask is: Among them, M t is the field of view mask,∈{0,1}, r M is the influence of the selected target on the predicted target, X k The location of the selected target.

6. The target position prediction method according to claim 5, characterized in that: The formula for generating an interaction heat map based on the selected target is: Among them, N is the interaction heat map sequence, N t is the interaction heat map at time t, is the Hadamard product.

7. The target position prediction method according to claim 6, characterized in that: The generator is a fully convolutional network with skip connections. The formula for generating the end point heat map is: Among them, O G is the end point heat map, H env is the input of the generator, W g are the network parameters of the generator.

8. The target position prediction method according to claim 7, wherein: The discriminator determines the predicted position of the predicted target according to the following formula, which includes the following sub-steps: Sample at least 10,000 locations from the interaction heat map based on the probability of the predicted location distribution, and extract the location information of the sampled locations; Use K-means clustering algorithm to obtain K-1 cluster centers on the extracted location information; According to the following formula, the cluster centers of K modes are output as the predicted positions: in, is the predicted position, and softargmax is the global coordinate normalization, which is used to calculate the mean of the position coordinates based on the probability distribution of the interaction heat map.

9. The target position prediction method according to claim 8, characterized in that: The position prediction step further includes a training sub-step, in which a discriminator discriminates the authenticity of K×K regions according to the following formula: O D =D(concat(P,O G );W D ) Among them, O D is the output of true or false judgment, W D are the network parameters of the generator.

10. The target position prediction method according to claim 9, characterized in that: In the training sub-step, the loss function is: Among them, L is the loss function, L GAN The end point heat map O G The loss function of the mid- and high-frequency characteristics, L BCE The end point heat map O G Loss function for mid- and low-frequency characteristics, Y gd To predict the future T fut The time-step trajectory sequence is a Gaussian distribution heat map with a mean and standard deviation of 4, and the future T of the target is predicted. fut The motion trajectory sequence of the time step is:

11. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the target position prediction method according to any one of claims 1 to 10 is implemented.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the target position prediction method according to any one of claims 1 to 10 is implemented.

13. A motor vehicle, characterized in that: The motor vehicle has an automatic driving function. When the automatic driving function is running, the motor vehicle predicts the positions of targets around the vehicle using the target position prediction method described in any one of claims 1 to 10.

Images