CN114559951A

CN114559951A - Obstacle trajectory prediction method, obstacle trajectory prediction device, storage medium, and electronic device

Info

Publication number: CN114559951A
Application number: CN202210261826.8A
Authority: CN
Inventors: 莫汇宇; 李鑫; 傅壮; 钟超; 钱德恒
Original assignee: Beijing Sankuai Online Technology Co Ltd
Current assignee: Beijing Sankuai Online Technology Co Ltd
Priority date: 2022-03-16
Filing date: 2022-03-16
Publication date: 2022-05-31

Abstract

The specification discloses a method and a device for predicting obstacle trajectories, a storage medium and an electronic device, which can acquire driving data of an obstacle within a first time period in history, and construct a state matrix by using state data of different times as row data and data of different data dimensions as column data. And then, respectively inputting the state matrix into a time coding layer and a space coding layer of the track prediction model to obtain the correlation characteristics in the time dimension and the correlation characteristics in the space dimension. And finally, determining the predicted driving track of the obstacle through a decoding layer of the track prediction model according to the correlation characteristics of the obstacle in the time dimension and the space dimension respectively. The historical driving data of the obstacle are coded from the time dimension and the space dimension respectively, the correlation between the characteristics of the obstacle on the time dimension and the space dimension is learned, and the accuracy of the track prediction is improved.

Description

Obstacle trajectory prediction method, obstacle trajectory prediction device, storage medium, and electronic device

Technical Field

The present application relates to the field of unmanned driving technologies, and in particular, to a method and an apparatus for predicting an obstacle trajectory, a storage medium, and an electronic device.

Background

Along with the development of the unmanned technology, the unmanned equipment is also widely applied to a plurality of fields, and more convenience is brought to the life of people.

In order to ensure safe driving of the unmanned aerial vehicle, when planning a driving track of the unmanned aerial vehicle, it is often necessary to predict a driving track of a surrounding obstacle so that the unmanned aerial vehicle can avoid each obstacle during driving.

At present, when a trajectory of an obstacle is predicted, a historical driving trajectory of the obstacle may be obtained first, and the historical driving trajectory may be encoded through a Long Short-Term Memory network (LSTM) to obtain a relationship between driving trajectories at different times. Then, the LSTM decoder decodes the obstacle to predict the driving track of the obstacle in a future time range.

However, only the correlation of the driving tracks in the time dimension is considered based on the relationship between the driving tracks at different moments, so that the track prediction result is not accurate enough, and the driving safety of the unmanned equipment is influenced.

Disclosure of Invention

The embodiment of the specification provides an obstacle trajectory prediction method, an obstacle trajectory prediction device, a storage medium and electronic equipment, which are used for partially solving the problems in the prior art.

The embodiment of the specification adopts the following technical scheme:

the present specification provides a method for predicting an obstacle trajectory, including:

acquiring running data of an obstacle within a first time period in history, wherein the running data comprises state data of a plurality of moments within the first time period;

taking state data at different moments as row data and taking data of different data dimensions of the state data as column data to construct a state matrix;

inputting the state matrix into a time coding layer of a pre-trained track prediction model, and performing column convolution on the state matrix to fuse the features of the same data dimension at different moments to obtain the correlation features of the barrier in the time dimension;

inputting the state matrix into a spatial coding layer of a pre-trained track prediction model, and performing row convolution on the state matrix to fuse the features of different data dimensions at the same moment to obtain the correlation features of the barrier in the spatial dimension;

and determining the predicted driving track of the obstacle through a decoding layer of the track prediction model according to the correlation characteristics of the obstacle in the time dimension and the space dimension respectively.

Optionally, before inputting the state matrix into a temporal coding layer and/or a spatial coding layer of a pre-trained trajectory prediction model, the method further includes:

for each moment in the first duration, determining a correlation matrix corresponding to the moment according to the state data of the obstacle corresponding to the moment and the state data of the obstacle corresponding to other moments, wherein the correlation matrix represents the correlation between the state data of the other moments and the state data of the moment;

determining an attention weight matrix according to the incidence matrix corresponding to each moment in the first duration;

and according to the attention weight matrix, performing attention weighting on the state matrix, and determining the weighted state matrix.

Optionally, the time-coding layer comprises a standard convolutional layer and a hole convolutional layer;

inputting the state matrix into a time coding layer of a pre-trained track prediction model, performing column convolution on the state matrix, and fusing features of the same data dimension at different moments to obtain associated features of the obstacle in the time dimension, wherein the method specifically comprises the following steps:

inputting the state matrix into a standard convolution layer of a time coding layer in a pre-trained track prediction model, and performing standard column convolution on the state matrix to fuse the features of the same data dimension at adjacent moments to obtain a first sub-correlation feature of the barrier in the time dimension;

inputting the state matrix into a cavity convolution layer of a time coding layer in a pre-trained track prediction model, performing cavity row convolution on the state matrix, fusing the characteristics of the same data dimension at interval moments to obtain a second sub-correlation characteristic of the barrier on the time dimension, wherein the interval moments are spaced for a preset time;

and performing feature fusion according to the first sub-correlation feature and the second sub-correlation feature to obtain the correlation feature of the obstacle in the time dimension.

Optionally, the spatial coding layer comprises a standard convolutional layer and a void convolutional layer;

inputting the state matrix into a spatial coding layer of a pre-trained track prediction model, performing row convolution on the state matrix, fusing features of different data dimensions at the same moment, and obtaining the correlation features of the obstacle in the spatial dimensions, specifically comprising:

inputting the state matrix into a standard convolution layer of a space coding layer in a pre-trained track prediction model, performing standard row convolution on the state matrix, and fusing the characteristics of adjacent data dimensions at the same moment to obtain a first sub-correlation characteristic of the barrier in the space dimension;

inputting the state matrix into a cavity convolution layer of a space coding layer in a pre-trained track prediction model, performing cavity row convolution on the state matrix, and fusing the characteristics of interval data dimensions at the same moment to obtain a second sub-correlation characteristic of the barrier in the space dimensions;

and performing feature fusion according to the first sub-associated feature and the second sub-associated feature to obtain the associated feature of the obstacle in the spatial dimension.

Optionally, the trajectory prediction model further comprises a full coding layer;

determining the predicted driving track of the obstacle through a decoding layer of the track prediction model according to the relevant characteristics of the obstacle in the time dimension and the space dimension respectively, wherein the method specifically comprises the following steps:

inputting the state matrix into a full coding layer of the track prediction model, and determining comprehensive association characteristics of the barrier;

performing feature fusion according to the association features of the obstacles in the time dimension, the association features in the space dimension and the comprehensive association features to obtain fusion features;

and inputting the fusion features into a decoding layer of the track prediction model, and determining the predicted driving track of the obstacle.

Optionally, determining the predicted travel trajectory of the obstacle according to the correlation characteristics of the obstacle in the time dimension and the space dimension, respectively, through a decoding layer of the trajectory prediction model, specifically including:

according to the association characteristics of the obstacles in the time dimension and the space dimension respectively, performing characteristic fusion and determining fusion characteristics;

determining the encoding times of the state matrix, and judging whether the encoding times reach a preset value;

if yes, inputting the fusion features into a decoding layer of the track prediction model, and determining the predicted driving track of the obstacle;

if not, the fusion characteristics are used as new state matrixes again, the newly determined state matrixes are input into the time coding layer respectively, column convolution is carried out on the state matrixes to obtain the correlation characteristics of the obstacles in the time dimension, the fusion characteristics are input into the space coding layer, row convolution is carried out on the state matrixes to determine the correlation characteristics of the obstacles in the space dimension, characteristic fusion is carried out according to the correlation characteristics of the obstacles in the time dimension and the space dimension respectively, and the fusion characteristics are determined until the coding times reach a preset value.

Optionally, the training of the trajectory prediction model specifically includes:

acquiring running data of a plurality of obstacles in a first time length and a second time length in history, wherein the time corresponding to the first time length is earlier than the time corresponding to the second time length;

for each obstacle, taking state data of the obstacle at a plurality of different moments in the first time length as row data, taking data of different data dimensions of the state data as column data, constructing a state matrix as sample data of a training sample, and labeling the training sample according to driving data of the obstacle in the second time length;

inputting sample data of each training sample into a time coding layer of a trajectory prediction model to be trained, performing column convolution on the sample data to fuse the features of the same data dimension at different moments to obtain the associated features of the training sample on the time dimension, inputting the sample data of the training sample into a space coding layer of the trajectory prediction model, performing row convolution on the sample data to fuse the features of different data dimensions at the same moment to obtain the associated features of the training sample on the space dimension;

determining the predicted running track of the training sample in the second duration through a decoding layer of the track prediction model according to the associated characteristics of the training sample in the time dimension and the space dimension respectively;

and adjusting model parameters in the track prediction model by taking the difference between the predicted running track of each training sample and the label of each training sample as a target, wherein the track prediction model is used for predicting the running track of the obstacle.

Optionally, with a goal of minimizing a difference between the predicted travel trajectory of each training sample and the label of each training sample, adjusting the model parameters in the trajectory prediction model specifically includes:

for each training sample, determining a key position point corresponding to each key moment in the predicted driving track of the training sample;

connecting into a fitting track of the training sample according to each key position point of the training sample;

and adjusting model parameters in the track prediction model by taking the difference between the predicted driving track of each training sample and the label of each training sample and the difference between the fitting track of each training sample and the label of each training sample as targets.

The present specification provides an obstacle trajectory prediction device including:

the system comprises an acquisition module, a display module and a control module, wherein the acquisition module is configured to acquire driving data of an obstacle within a first time period historically, and the driving data comprises state data of a plurality of moments within the first time period;

the building module is configured to build a state matrix by taking state data at different moments as row data and taking data of different data dimensions of the state data as column data;

the time correlation module is configured to input the state matrix into a time coding layer of a pre-trained track prediction model, perform column convolution on the state matrix, and fuse features of the same data dimension at different moments to obtain correlation features of the barrier in the time dimension;

the spatial correlation module is configured to input the state matrix into a spatial coding layer of a pre-trained track prediction model, perform row convolution on the state matrix, enable features of different data dimensions at the same moment to be fused, and obtain correlation features of the obstacle in spatial dimensions;

and the prediction module is configured to determine the predicted driving track of the obstacle through a decoding layer of the track prediction model according to the associated characteristics of the obstacle in the time dimension and the space dimension respectively.

The present specification provides a computer-readable storage medium storing a computer program which, when executed by a processor, implements the above-described obstacle trajectory prediction method.

The present specification provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, the processor implements the above-mentioned obstacle trajectory prediction method.

The embodiment of the specification adopts at least one technical scheme which can achieve the following beneficial effects:

in this specification, the travel data of the obstacle within the first time period in history may be acquired, and the state data at different times may be used as the row data, and the data with different data dimensions may be used as the column data, to construct the state matrix. And then, respectively inputting the state matrix into a time coding layer and a space coding layer of the track prediction model to obtain the correlation characteristics in the time dimension and the correlation characteristics in the space dimension. And finally, determining the predicted driving track of the obstacle through a decoding layer of the track prediction model according to the correlation characteristics of the obstacle in the time dimension and the space dimension respectively. The historical driving data of the obstacle are coded from the time dimension and the space dimension respectively, the correlation between the characteristics of the obstacle on the time dimension and the space dimension is learned, and the accuracy of the track prediction is improved.

Drawings

The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the application and together with the description serve to explain the application and not to limit the application. In the drawings:

fig. 1 is a schematic flow chart of a method for predicting an obstacle trajectory according to an embodiment of the present disclosure;

FIG. 2 is a network architecture diagram of a trajectory prediction model provided in an embodiment of the present disclosure;

fig. 3 is a network structure diagram of a time coding layer provided in an embodiment of the present specification;

fig. 4 is a network structure diagram of a spatial coding layer according to an embodiment of the present disclosure;

FIG. 5 is a diagram of a network architecture of a trajectory prediction model provided in an embodiment of the present disclosure;

fig. 6 is a schematic structural diagram of an obstacle trajectory prediction apparatus provided in an embodiment of the present disclosure;

fig. 7 is a schematic diagram of an electronic device for implementing a method for predicting an obstacle trajectory according to an embodiment of the present specification.

Detailed Description

In order to make the objects, technical solutions and advantages of the present disclosure more apparent, the technical solutions of the present disclosure will be clearly and completely described below with reference to the specific embodiments of the present disclosure and the accompanying drawings. It should be apparent that the described embodiments are only some of the embodiments of the present application, and not all of the embodiments. All other embodiments obtained by a person skilled in the art without making any inventive step based on the embodiments in the description belong to the protection scope of the present application.

The present specification provides a method of predicting an obstacle trajectory. The technical solutions provided by the embodiments of the present application are described in detail below with reference to the accompanying drawings.

Fig. 1 is a schematic flow chart of a method for predicting an obstacle trajectory according to an embodiment of the present disclosure, which may specifically include the following steps:

s100: the driving data of the obstacle in the first time period in history is acquired.

S102: and taking the state data at different moments as row data and taking the data with different data dimensions of the state data as column data to construct a state matrix.

When planning the driving track of the unmanned device, it is often necessary to predict the driving track of surrounding obstacles so that the unmanned device can avoid each obstacle during driving. Therefore, the present specification provides an obstacle trajectory prediction method, which can extract effective correlation features from the historical behavior trajectory of the obstacle, and further predict a future travel trajectory.

The obstacle trajectory prediction method can be executed by a back-end server, and can also be executed by an unmanned device as an execution main body, which is not limited in the specification and can be set as required. For convenience of description, the server implementation is described as an example in the following. The obstacle is a dynamic obstacle capable of generating position change, and may be a vehicle, a pedestrian, or the like.

Specifically, the server may first acquire travel data of the obstacle for a first time period in history. The driving data of the obstacle includes state data of the obstacle at a plurality of moments in the first time period, and may further include a driving track, a size (length, width, and the like) and the like of the obstacle. The status data includes data for several different data dimensions, such as position coordinates, direction of travel, speed, acceleration, etc.

In order to make the track prediction effect more accurate, the first duration may be selected from a current recent time range, such as driving data of 5 minutes ago. To predict the travel trajectory for the next period of time.

The historical travel data of the obstacle acquired in this specification may be obtained by detecting a dynamic obstacle in the surrounding environment by a camera, a lidar, or other sensor device disposed in the vehicle during the historical travel of the vehicle. And determining the driving data of each dynamic obstacle in each frame based on the acquired environment image of each frame or the laser point cloud of each frame.

And then, the server can splice the collected multi-frame data into a two-dimensional image form, the state data of the barrier at different moments are used as row data, the data of different data dimensions of the state data are used as column data, and a state matrix is constructed.

For example, assuming that the obstacle is a pedestrian, the first time period includes T1 and T2 … Tn. Then from the pedestrian's travel data at time T1, it is determined that the pedestrian corresponds to the state data N1 at time T1, and the state data N1 includes k data dimensions such as the position coordinates, speed, acceleration, size, and travel direction of the pedestrian at time T1. And then, splicing the state data at n moments into an n × k state matrix by taking the state data as row data and taking the data with different dimensionalities of the state data as column data.

S104: and inputting the state matrix into a time coding layer of a pre-trained track prediction model, and performing column convolution on the state matrix to fuse the features of the same data dimension at different moments to obtain the correlation features of the barrier in the time dimension.

S106: and inputting the state matrix into a spatial coding layer of a pre-trained track prediction model, and performing row convolution on the state matrix to fuse the features of different data dimensions at the same moment to obtain the correlation features of the barrier in the spatial dimension.

In the prior art, the manner of encoding historical driving data through the LSTM only considers the correlation between features at different time points in the time dimension. But not every data dimension's features have the same effect on the trajectory prediction results, and there are differences in the correlation between every data dimension.

Thus, in the present specification, the correlation between data dimensions is referred to as the correlation in spatial dimension, i.e. the association between features of different data dimensions at the same time. When the obstacle trajectory is predicted in the present specification, the correlation of the historical driving data can be learned from two dimensions of time and space, respectively, so as to obtain an accurate trajectory prediction result.

Specifically, the constructed state matrix is input into a time coding layer of a pre-trained track prediction model, and the state matrix is subjected to column convolution, so that the features of the same data dimension at different moments are fused, and the correlation feature of the obstacle in the time dimension is obtained.

And inputting the state matrix into a spatial coding layer of a pre-trained track prediction model, performing row convolution on the state matrix, fusing the features of different data dimensions at the same moment, and obtaining the correlation features of the obstacle in the spatial dimension.

Fig. 2 is a network structure diagram of a trajectory prediction model provided in an embodiment of the present disclosure, where the trajectory prediction model includes a temporal coding layer, a spatial coding layer, and a decoding layer. The state matrix corresponding to the obstacle can be respectively input to the time coding layer to obtain the correlation characteristic of the obstacle in the time dimension, and input to the space coding layer to obtain the correlation characteristic of the obstacle in the space dimension.

In the trajectory prediction model shown in the embodiment of the present specification, the time coding layer and the space coding layer are in a parallel network structure, the state matrix corresponding to the obstacle may be simultaneously and respectively input to the time coding layer and the space coding layer, and the specification does not limit the order of inputting the state matrix.

Further, in the time coding layer of the trajectory prediction model, the time coding layer may be composed of a standard convolutional layer. Since one row of data in the state matrix represents state data at the same time, and one column of data represents data at different times in the same data dimension. Therefore, when the association relation among the features in the time dimension is extracted, different row data can be fused. And inputting the state matrix corresponding to the obstacle into a standard convolution layer of a time coding layer, and performing standard column convolution through a one-dimensional convolution kernel to fuse the characteristics of the same data dimension at adjacent moments to obtain a first sub-correlation characteristic of the training sample in the time dimension. Among them, 3 × 1, 5 × 1, etc. can be used as the one-dimensional convolution kernel.

Furthermore, in order to fuse feature information of other close moments, the time coding layer may further include a hole convolution layer, and when extracting the association relationship between features in the time dimension, the state matrix corresponding to the obstacle may be input into the hole convolution layer of the time coding layer, and hole column convolution is performed through one-dimensional convolution to fuse features of the same data dimension at interval moments, so as to obtain a second sub-association feature of the training sample in the time dimension. The hollow convolution kernel can be selected to be 3 × 1, 5 × 1, etc., and the interval between the interval moments is a preset time length, which is related to the expansion rate of the hollow convolution layer. The expansion rate of the void convolution layer can be set according to needs, but the specification does not limit this, and for example, the expansion rate can be set to 2 to blend features of a similar moment at a moment.

In addition, in order to fuse features between different moments of different pitches, in this specification, a plurality of different convolution kernels are used to perform feature fusion from different pitch dimensions such as a near pitch moment, an intermediate pitch moment, and a far pitch moment.

In one embodiment of the present description, the time coding layer may be configured with 3 convolutional layers for merging features between the near-pitch time, the intermediate-pitch time, and the far-pitch time, respectively. The near-distance time is a standard convolution layer, the intermediate-distance time and the far-distance time are cavity convolution layers with different convolution kernel sizes, and the convolution kernel size of the far-distance time is larger than that of the intermediate-distance time. And respectively inputting the state matrix corresponding to the obstacle into the convolution layers corresponding to the short-distance time to obtain the sub-correlation characteristics of the short-distance time, inputting the convolution layers corresponding to the middle-distance time to obtain the sub-correlation characteristics of the middle-distance time, and inputting the convolution layers corresponding to the long-distance time to obtain the sub-correlation characteristics of the long-distance time. And then, inputting the sub-associated features of the short-distance moment, the middle-distance moment and the long-distance moment into a feature fusion layer to obtain the associated features of the training sample in the time dimension.

Assume that the convolution layer corresponding to the near-pitch time is a 3 × 1 standard convolution layer, the convolution layer corresponding to the intermediate-pitch time is a 3 × 1 void convolution layer, the convolution layer corresponding to the far-pitch time is a 5 × 1 void convolution layer, and the feature fusion layer is a 1 × 1 standard convolution layer. As shown in fig. 3, the state matrix corresponding to the obstacle can be input into the 3 × 1 standard convolutional layer, the 3 × 1 hole convolutional layer, and the 5 × 1 hole convolutional layer, respectively, to obtain the sub-correlation features at the near-pitch time, the intermediate-pitch time, and the far-pitch time. And then, inputting each sub-correlation feature into a 1 × 1 standard convolutional layer to obtain a correlation feature in a fused time dimension.

Similarly, in the spatial coding layer of the trajectory prediction model, the spatial coding layer may be composed of standard convolutional layers. Since one row in the state matrix represents state data at the same time, and one column represents data at different times for the same data dimension. Therefore, when the association relation between the features in the spatial dimension is extracted, different columns of data can be fused. And inputting the state matrix corresponding to the obstacle into a standard convolution layer of a space coding layer, and performing standard row convolution through a one-dimensional convolution kernel to fuse the characteristics of adjacent data dimensions at the same moment to obtain a first sub-correlation characteristic of the obstacle in the space dimension. Wherein, 1 × 3, 1 × 5, etc. can be used as the one-dimensional convolution kernel.

Furthermore, in order to fuse the features of other data dimensions, the spatial coding layer may also include a hole convolution layer, and when extracting the correlation between the features in the spatial dimension, the state matrix corresponding to the obstacle may be input into the hole convolution layer of the spatial coding layer, and hole row convolution may be performed through a one-dimensional convolution kernel to fuse the features at the same time of the interval data dimension, so as to obtain a second sub-correlation feature of the obstacle in the spatial dimension. The hole convolution kernel can be selected from 1 × 3, 1 × 5, and the spacing distance of the spacing data dimension is related to the expansion rate of the hole convolution layer.

In addition, in order to associate the relationship between the features of different data dimensions, in this specification, a plurality of different convolution kernels may be used to perform feature fusion from a near-pitch data dimension, a middle-pitch data dimension, a far-pitch data dimension, and the like.

Assume that the convolution layer corresponding to the short-pitch data dimension is a 1 × 3 standard convolution layer, the convolution layer corresponding to the middle-pitch data dimension is a 1 × 3 void convolution layer, the convolution layer corresponding to the long-pitch data dimension is a 1 × 5 void convolution layer, and the feature fusion layer is a 1 × 1 standard convolution layer. As shown in fig. 4, the state matrix corresponding to the obstacle can be input into the 1 × 3 standard convolutional layer, the 1 × 3 void convolutional layer, and the 1 × 5 void convolutional layer, respectively, to obtain the sub-correlation features of the near-pitch data dimension, the intermediate-pitch data dimension, and the far-pitch data dimension. And then, inputting each sub-associated feature into a 1 × 1 standard convolutional layer to obtain the associated feature on the spatial dimension after fusion.

In addition, before inputting the state matrix into the temporal coding layer and the spatial coding layer, in order to improve the track prediction effect, an attention mechanism may be adopted to enhance the portions of the state matrix that need attention. Specifically, the state data of the obstacle corresponding to each time in the first time period may be determined from the state matrix. And then, aiming at each moment in the first duration, determining a correlation matrix corresponding to the moment according to the state data of the obstacle corresponding to the moment and the state data of the obstacle corresponding to other moments. And the incidence matrix represents the correlation between the state data at each other moment and the state data at the moment. And finally, determining an attention weight matrix according to the incidence matrix corresponding to each moment in the first duration, carrying out attention weighting on the state matrix according to the attention weight matrix, and determining the weighted state matrix.

Suppose that the first time period includes time T1 and time T2 … Tn, the state data of the obstacle corresponding to each time is Nn, and each state data includes position coordinates and speedAcceleration, size and driving direction, the state matrix spliced by the state data at each moment can be represented as

Each row represents the state data for the corresponding time Tn. And then, determining an attention weight matrix Q according to the state data of the obstacle corresponding to each moment in the state matrix. Wherein:

the incidence matrix corresponding to each moment is as follows: q. q.s_n＝[NnN1 NnN2 … NnNn]

Then, each row of data in the attention weight matrix Q is normalized through a normalization index function softmax, and the normalization result is multiplied by the original state matrix to obtain a weighted state matrix.

S108: and determining the predicted driving track of the obstacle through a decoding layer of the track prediction model according to the correlation characteristics of the obstacle in the time dimension and the space dimension respectively.

After the state matrix is encoded from the time dimension and the space dimension, the trajectory prediction can be performed according to the extracted correlation characteristics of the time dimension and the extracted correlation characteristics of the space dimension.

Specifically, the correlation characteristic of the obstacle in the time dimension and the correlation characteristic of the obstacle in the space dimension are subjected to characteristic fusion to obtain a fusion characteristic. Then, as shown in fig. 2, the fused feature is input into a decoding layer of the trajectory prediction model, and the predicted travel trajectory of the obstacle within the second duration is determined. Wherein, the decoding layer is composed of a plurality of full Connected layers (FC). The predicted travel path includes a location point at which the obstacle is located at each time within the second period of time.

Furthermore, in order to fully fuse local features and reduce data calculation amount, when feature fusion is performed, a cavity convolution layer can be firstly adopted to perform local feature fusion on the correlation features of the time dimension and the space dimension, and meanwhile, the effect of reducing the dimension is achieved. And then, the feature after dimension reduction is promoted through the standard convolution layer to obtain a fusion feature. The hole convolution layer can adopt a 3 x 3 hole convolution kernel, and the standard convolution layer can adopt a 1 x 1 standard convolution kernel.

Based on the obstacle trajectory prediction method shown in fig. 1, driving data of the obstacle within a first time period in history can be acquired, state data at different times are used as row data, data with different data dimensions are used as column data, and a state matrix is constructed. And then, respectively inputting the state matrix into a time coding layer and a space coding layer of the track prediction model to obtain the associated characteristics in the time dimension and the associated characteristics in the space dimension. And finally, determining the predicted driving track of the obstacle through a decoding layer of the track prediction model according to the correlation characteristics of the obstacle in the time dimension and the space dimension respectively. The historical driving data of the obstacle are coded from the time dimension and the space dimension respectively, the correlation between the characteristics of the obstacle on the time dimension and the space dimension is learned, and the accuracy of the track prediction is improved.

In one embodiment of the present specification, in order to fully retain the feature information of the historical driving data, the trajectory prediction model is further provided with a full coding layer, the state matrix of the obstacle can be input into the full coding layer, and the features between different data dimensions at different times are correlated to obtain the comprehensive correlation features of the obstacle. Wherein the full code layer may be a standard convolutional layer, and the convolutional kernel may be set to 1 × 1. And then, performing feature fusion according to the associated features of the training sample in the time dimension, the associated features in the space dimension and the comprehensive associated features to obtain fusion features. And finally, inputting the fusion characteristics into a decoding layer of the track prediction model, and determining the predicted driving track of the obstacle within the second time length.

Fig. 5 is a network structure diagram of a trajectory prediction model provided in an embodiment of this specification, where the trajectory prediction model is composed of an attention network, an encoding layer, and a decoding layer, and after a state matrix of the obstacle is input into the trajectory prediction model, the obstacle may be first weighted by the attention network to obtain a weighted state matrix, and then the weighted state matrix is input into a time encoding layer, a full encoding layer, and a space encoding layer, respectively, to obtain a correlation feature, a comprehensive correlation feature, and a correlation feature in a space dimension of the obstacle. Then, feature fusion is performed through the feature fusion layer to obtain fusion features. And finally, inputting the fusion characteristics into a decoding layer to obtain a predicted driving track within a second duration.

In one embodiment of the present specification, in order to improve the fusion effect of the features, multiple iterative encoding may be performed. And performing feature fusion according to the correlation features of the barrier in the time dimension and the space dimension respectively, determining the fusion features, determining the encoding times of the state matrix, and judging whether the encoding times reach a preset value. The encoding times refer to the times of obtaining the fusion characteristics of the state matrix through the encoding layer. And if the encoding times reach a preset value, inputting the fusion characteristics into a decoding layer, and determining the predicted driving track of the obstacle in the second time length. And if the encoding frequency does not reach the preset value, re-using the fusion characteristic as a state matrix corresponding to the obstacle, coding the re-determined state matrix again by the method in the step S104 and the step S106, performing characteristic fusion on the output association characteristics of the obstacle in the time dimension and the space dimension, and determining the fusion characteristic until the encoding frequency reaches the preset value.

In addition, compared with the method for predicting the track by adopting the LSTM time sequence model in the prior art, serial calculation needs to be carried out on the position point of each moment, and the overall timeliness of the model is seriously influenced. While in the unmanned driving process, long-time calculation delay may cause serious safety problems. The trajectory prediction model in the application adopts the efficient convolution network, the calculation efficiency is higher, the trajectory prediction time is shorter, and the timeliness of the trajectory prediction is greatly improved.

When the trajectory prediction model is trained, the following training modes can be specifically adopted:

a0: and acquiring the driving data of a plurality of obstacles in a first time period and a second time period in history.

A2: and aiming at each obstacle, taking the state data of the obstacle at a plurality of different moments in the first time length as row data, taking the data of different data dimensions of the state data as column data, constructing a state matrix as sample data of a training sample, and labeling the training sample according to the driving data of the obstacle in the second time length.

In training the trajectory prediction model, a large amount of travel data of the obstacle historically over a first time period and travel data over a second time period may be obtained. Wherein, the time corresponding to the first time length is earlier than the time corresponding to the second time length. The first time period and the second time period can be set as required, such as 5min, 30s, and so on.

And then, for each obstacle, taking the state data of the obstacle at a plurality of different moments in a first time period as row data, taking the data of different data dimensions of the state data as column data, and constructing a state matrix as sample data of a training sample.

Because the time period corresponding to the second time period is a future time period relative to the time period of the first time period, the training sample can be labeled according to the driving data of the obstacle in the second time period. When the sample marking is performed, the actual driving track of the obstacle in the second time period and the position point corresponding to each moment on the actual driving track can be marked according to the driving data of the obstacle in the second time period.

A4: for each training sample, inputting sample data of the training sample into a time coding layer of a track prediction model to be trained, performing column convolution on the sample data to fuse the features of the same data dimension at different moments to obtain the associated features of the training sample on the time dimension, inputting the sample data of the training sample into a space coding layer of the track prediction model, performing row convolution on the sample data to fuse the features of different data dimensions at the same moment to obtain the associated features of the training sample on the space dimension.

A6: and determining the predicted driving track of the training sample in the second time length through a decoding layer of the track prediction model according to the associated characteristics of the training sample in the time dimension and the space dimension respectively.

The trajectory prediction model to be trained may be a network structure as shown in fig. 2 or a network structure as shown in fig. 5, and feature extraction performed by the temporal coding layer and the spatial coding layer has been described in detail in the above steps S104 to S106, which is not described in detail herein.

A8: and adjusting model parameters in the track prediction model by taking the minimum difference between the predicted driving track of each training sample and the label of each training sample as a target, wherein the track prediction model is used for predicting the driving track of the obstacle.

In the model training process, model parameters can be adjusted based on the difference between the predicted driving track output by the track prediction model to be trained and the sample label.

Therefore, the Average Displacement Error (ADE) between the predicted travel track of each training sample and the actual travel track labeled by each training sample can be determined, and the model parameters in the track prediction model are adjusted to predict the travel track of the obstacle through the track prediction model with the aim of minimizing the ADE between the predicted travel track and the actual travel track, so that the travel path of the unmanned equipment can be reasonably planned in time.

Further, average displacement error calculation is carried out on each position point on the predicted running track, and influence weights of each position point on the errors are consistent. In the actual driving process, the starting point, the end point and the position point at the intermediate time often have a larger influence on the trajectory deviation.

Thus, for each training sample, the key position points corresponding to the key times in the predicted travel locus of the training sample can be determined. And then, connecting into a fitting track of the training sample according to each key position point of the training sample. And finally, determining a loss function according to the ADE between the predicted driving track of each training sample and the actual driving track marked by each training sample and the ADE between the actual driving track marked by each training sample and the fitting track of each training sample, and adjusting the model parameters in the track prediction model by taking the minimization of the loss function as a target. The key time can be set as required, for example, the key time is set as a starting point time, an end point time and a middle time of the second duration, and the key position point is a position point where the obstacle is located at the key time. For example, the second duration is set to 5s, and the key time may be selected to be 1s, 3s, and 5 s.

It should be noted that the difference between the fitted actual driving trajectory and the fitted trajectory, not between the key position points, is because the single position point is not strongly constrained to the trajectory prediction, and the fitted trajectory includes more position point information, thereby bringing about stronger constraint. And moreover, smoothness of the finally output predicted driving track can be guaranteed by using fitting estimation for constraint, and the problem of excessive track abnormal points is effectively avoided.

Based on the method for predicting the trajectory of the obstacle shown in fig. 1, an embodiment of the present specification further provides a schematic structural diagram of an apparatus for predicting the trajectory of the obstacle, as shown in fig. 6.

Fig. 6 is a schematic structural diagram of an obstacle trajectory prediction apparatus provided in an embodiment of the present disclosure, including:

the system comprises an acquisition module 200, a storage module and a control module, wherein the acquisition module is configured to acquire driving data of an obstacle within a first time period in history, and the driving data comprises state data of a plurality of moments within the first time period;

the building module 202 is configured to build a state matrix by using state data at different moments as row data and using data of different data dimensions of the state data as column data;

the time correlation module 204 is configured to input the state matrix into a time coding layer of a pre-trained trajectory prediction model, perform column convolution on the state matrix, and fuse features of the same data dimension at different moments to obtain correlation features of the obstacle in the time dimension;

the spatial correlation module 206 is configured to input the state matrix into a spatial coding layer of a pre-trained trajectory prediction model, perform row convolution on the state matrix, and fuse features of different data dimensions at the same time to obtain correlation features of the obstacle in spatial dimensions;

a prediction module 208 configured to determine a predicted travel trajectory of the obstacle through a decoding layer of the trajectory prediction model according to the associated features of the obstacle in the time dimension and the space dimension, respectively.

Optionally, the building module 202 is further configured to, for each time within the first duration, determine a correlation matrix corresponding to the time according to the state data of the obstacle corresponding to the time and the state data of the obstacle corresponding to each other time, where the correlation matrix represents a correlation between the state data of each other time and the state data of the time, determine an attention weight matrix according to the correlation matrix corresponding to each time within the first duration, perform attention weighting on the state matrix according to the attention weight matrix, and determine a weighted state matrix.

Optionally, the time coding layer includes a standard convolutional layer and a void convolutional layer, the time correlation module 204 is specifically configured to input the state matrix into the standard convolutional layer of the time coding layer in the pre-trained trajectory prediction model, perform standard column convolution on the state matrix, fuse features of the same data dimension at adjacent times to obtain a first sub-correlation feature of the obstacle in the time dimension, input the state matrix into the void convolutional layer of the time coding layer in the pre-trained trajectory prediction model, perform void column convolution on the state matrix, fuse features of the same data dimension at interval times to obtain a second sub-correlation feature of the obstacle in the time dimension, where an interval between the interval times is a preset duration, and according to the first sub-correlation feature and the second sub-correlation feature, and carrying out feature fusion to obtain the associated features of the obstacle in the time dimension.

Optionally, the spatial coding layer includes a standard convolutional layer and a void convolutional layer, and the spatial correlation module 206 is specifically configured to input the state matrix into the standard convolutional layer of the spatial coding layer in the pre-trained trajectory prediction model, performing standard row convolution on the state matrix to enable the features of adjacent data dimensions at the same moment to be fused to obtain a first sub-correlation feature of the barrier on the space dimension, inputting the state matrix into a cavity convolution layer of a space coding layer in a pre-trained track prediction model, performing cavity row convolution on the state matrix to fuse the features of interval data dimensionality at the same moment to obtain a second sub-correlation feature of the barrier on the space dimensionality, and performing feature fusion according to the first sub-associated feature and the second sub-associated feature to obtain the associated feature of the obstacle in the spatial dimension.

Optionally, the trajectory prediction model further includes a full coding layer, and the prediction module 208 is specifically configured to input the state matrix into the full coding layer of the trajectory prediction model, determine a comprehensive correlation characteristic of the obstacle, perform feature fusion according to the correlation characteristic of the obstacle in the time dimension, the correlation characteristic in the space dimension, and the comprehensive correlation characteristic, obtain a fusion characteristic, input the fusion characteristic into a decoding layer of the trajectory prediction model, and determine the predicted travel trajectory of the obstacle.

Optionally, the prediction module 208 is specifically configured to perform feature fusion according to the correlation features of the obstacle in the time dimension and the space dimension, determine a fusion feature, determine the number of times of encoding of the state matrix, and determine whether the number of times of encoding reaches a preset value, if so, input the fusion feature into a decoding layer of the trajectory prediction model, determine a predicted travel trajectory of the obstacle, if not, take the fusion feature as a new state matrix again, input the newly determined state matrices into the time encoding layer, perform column convolution on the state matrices, obtain the correlation feature of the obstacle in the time dimension, input the space encoding layer, perform row convolution on the state matrices, determine the correlation feature of the obstacle in the space dimension, and according to the correlation features of the obstacle in the time dimension and the space dimension, and performing feature fusion, and determining fusion features until the coding times reach a preset value.

Optionally, the obstacle trajectory prediction apparatus further includes a model training module 210, where the model training module 210 is specifically configured to obtain travel data of a plurality of obstacles in a first time length and a second time length in history, where a time corresponding to the first time length is earlier than a time corresponding to the second time length, construct, for each obstacle, a state matrix as row data using state data of the obstacle at a plurality of different times in the first time length, and using data of different data dimensions of the state data as column data, construct a state matrix as sample data of a training sample, label the training sample according to the travel data of the obstacle in the second time length, for each training sample, input the sample data of the training sample into a time coding layer of a trajectory prediction model to be trained, and perform column convolution on the sample data, the method comprises the steps of fusing features of the same data dimension at different moments to obtain associated features of training samples in a time dimension, inputting sample data of the training samples into a space coding layer of a track prediction model, performing row convolution on the sample data to fuse the features of the different data dimensions at the same moment to obtain the associated features of the training samples in the space dimension, determining predicted travelling tracks of the training samples in a second duration through a decoding layer of the track prediction model according to the associated features of the training samples in the time dimension and the space dimension respectively, and adjusting model parameters in the track prediction model aiming at minimizing the difference between the predicted travelling tracks of the training samples and labels of the training samples, wherein the track prediction model is used for predicting the travelling tracks of obstacles.

Optionally, the model training module 210 is specifically configured to, for each training sample, determine key position points corresponding to each key time in the predicted travel track of the training sample, and connect a fitting track of the training sample according to each key position point of the training sample, so as to minimize a difference between the predicted travel track of each training sample and a label of each training sample, and a difference between the fitting track of each training sample and the label of each training sample as targets, and adjust model parameters in the track prediction model.

Embodiments of the present specification also provide a computer-readable storage medium, which stores a computer program, and the computer program can be used to execute the obstacle trajectory prediction method provided in fig. 1.

According to the method for predicting the obstacle trajectory shown in fig. 1, the embodiment of the present specification further provides a schematic structural diagram of the electronic device shown in fig. 7. As shown in fig. 7, at the hardware level, the electronic device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory, but may also include hardware required for other services. The processor reads a corresponding computer program from the non-volatile memory into the memory and then runs the computer program to implement the obstacle trajectory prediction method shown in fig. 1.

Of course, besides the software implementation, the present specification does not exclude other implementations, such as logic devices or a combination of software and hardware, and the like, that is, the execution subject of the following processing flow is not limited to each logic unit, and may be hardware or logic devices.

In the 90 s of the 20 th century, improvements in a technology could clearly distinguish between improvements in hardware (e.g., improvements in circuit structures such as diodes, transistors, switches, etc.) and improvements in software (improvements in process flow). However, as technology advances, many of today's process flow improvements have been seen as direct improvements in hardware circuit architecture. Designers almost always obtain the corresponding hardware circuit structure by programming an improved method flow into the hardware circuit. Thus, it cannot be said that an improvement in the process flow cannot be realized by hardware physical modules. For example, a Programmable Logic Device (PLD), such as a Field Programmable Gate Array (FPGA), is an integrated circuit whose Logic functions are determined by programming the Device by a user. A digital system is "integrated" on a PLD by the designer's own programming without requiring the chip manufacturer to design and create a dedicated integrated circuit chip. Furthermore, nowadays, instead of manually generating an Integrated Circuit chip, such Programming is often implemented by "logic compiler" software, which is similar to the software compiler used in program development and writing, but the original code before compiling is written in a specific Programming Language, which is called Hardware Description Language (HDL), and the HDL is not only one kind but many kinds, such as abel (advanced boot Expression Language), ahdl (alternate Language Description Language), communication, CUPL (computer universal Programming Language), HDCal (Java Hardware Description Language), langa, Lola, mylar, HDL, las, harddl (software Description Language), vhh-Language, and vhigh-Language, which are currently used most commonly. It will also be apparent to those skilled in the art that hardware circuitry that implements the logical method flows can be readily obtained by merely slightly programming the method flows into an integrated circuit using the hardware description languages described above.

The controller may be implemented in any suitable manner, for example, the controller may take the form of, for example, a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro) processor, logic gates, switches, an Application Specific Integrated Circuit (ASIC), a programmable logic controller, and an embedded microcontroller, examples of which include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320, the memory controller may also be implemented as part of the control logic for the memory. Those skilled in the art will also appreciate that, in addition to implementing the controller as pure computer readable program code, the same functionality can be implemented by logically programming method steps such that the controller is in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, embedded microcontrollers and the like. Such a controller may thus be considered a hardware component, and the means included therein for performing the various functions may also be considered as a structure within the hardware component. Or even means for performing the functions may be regarded as being both a software module for performing the method and a structure within a hardware component.

The systems, devices, modules or units illustrated in the above embodiments may be implemented by a computer chip or an entity, or by a product with certain functions. One typical implementation device is a computer. In particular, the computer may be, for example, a personal computer, a laptop computer, a cellular telephone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

For convenience of description, the above devices are described as being divided into various units by function, and are described separately. Of course, the functions of the various elements may be implemented in the same one or more software and/or hardware implementations of the present description.

As will be appreciated by one skilled in the art, embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

In a typical configuration, a computing device includes one or more processors (CPUs), input/output interfaces, network interfaces, and memory.

The memory may include forms of volatile memory in a computer readable medium, Random Access Memory (RAM) and/or non-volatile memory, such as Read Only Memory (ROM) or flash memory (flash RAM). Memory is an example of a computer-readable medium.

Computer-readable media, including both non-transitory and non-transitory, removable and non-removable media, may implement information storage by any method or technology. The information may be computer readable instructions, data structures, modules of a program, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), Read Only Memory (ROM), Electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), Digital Versatile Discs (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device. As defined herein, a computer readable medium does not include a transitory computer readable medium such as a modulated data signal and a carrier wave.

It should also be noted that the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.

As will be appreciated by one skilled in the art, embodiments of the present description may be provided as a method, system, or computer program product. Accordingly, the description may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the description may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

This description may be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media including memory storage devices.

The embodiments in the present specification are described in a progressive manner, and the same and similar parts among the embodiments are referred to each other, and each embodiment focuses on the differences from the other embodiments. In particular, for the system embodiment, since it is substantially similar to the method embodiment, the description is simple, and for the relevant points, reference may be made to the partial description of the method embodiment.

The above description is only an example of the present specification, and is not intended to limit the present specification. Various modifications and alterations to this description will become apparent to those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present specification should be included in the scope of the claims of the present specification.

Claims

1. An obstacle trajectory prediction method, comprising:

2. The method of claim 1, wherein prior to inputting the state matrix into a temporal coding layer and/or a spatial coding layer of a pre-trained trajectory prediction model, the method further comprises:

3. The method of claim 1, wherein the time-coding layers comprise a standard convolutional layer and a hole convolutional layer;

and performing feature fusion according to the first sub-associated feature and the second sub-associated feature to obtain the associated feature of the obstacle in the time dimension.

4. The method of claim 1, wherein the spatial coding layer comprises a standard convolutional layer and a hole convolutional layer;

5. The method of claim 1, wherein the trajectory prediction model further comprises a full coding layer;

determining the predicted driving track of the obstacle according to the correlation characteristics of the obstacle in the time dimension and the space dimension respectively through a decoding layer of the track prediction model, and specifically comprises the following steps:

6. The method according to claim 1, wherein determining the predicted travel trajectory of the obstacle through a decoding layer of the trajectory prediction model according to the associated features of the obstacle in the time dimension and the space dimension, respectively, specifically comprises:

7. The method of claim 1, wherein training the trajectory prediction model specifically comprises:

8. The method of claim 7, wherein adjusting model parameters in the trajectory prediction model with the goal of minimizing a difference between the predicted travel trajectory of each training sample and the label of each training sample comprises:

9. An obstacle trajectory prediction device characterized by comprising:

the spatial correlation module is configured to input the state matrix into a spatial coding layer of a pre-trained track prediction model, and perform row convolution on the state matrix to fuse features of different data dimensions at the same moment to obtain correlation features of the barrier in spatial dimensions;

10. A computer-readable storage medium, characterized in that the storage medium stores a computer program which, when executed by a processor, implements the method of any of claims 1 to 8.

11. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method of any of claims 1 to 8 when executing the program.