U-turn intention and trajectory prediction method based on high-dimensional embedding and deep parallel computing

Abstract

The present invention relates to a U-turn intention and trajectory prediction method based on high-dimensional embedded deep parallel computing, including shooting a video of a U-turn near an intersection and extracting data to generate a corresponding time-series trajectory; analyzing the time-series trajectory of vehicles on the left-turn lane for U-turn intention, and forming a new input with straight-moving vehicles, U-turning vehicles, and straight-moving vehicles on the left-turn lane, and embedding the vehicle's behavior information into a high-dimensional space through an embedding layer; extracting the time and space information of the high-dimensional embedding and merging them into spatiotemporal features; inputting the spatiotemporal feature information into a deep separable convolutional network of the lower layer, decoding it through a decoder, and outputting the future trajectory information of all traffic participants. The above technical solution effectively solves the problem that the existing trajectory prediction method cannot make an accurate judgment on the vehicle's U-turn intention, and the vehicle trajectory prediction speed is slow and the accuracy is not high.

Description

Technical Field

[0001] The present invention relates to the technical fields of computers, traffic big data and automatic driving trajectory prediction, and in particular to a dangerous U-turn intention and trajectory prediction method based on high-dimensional embedded deep parallel computing. Background Art

[0002] Mixed-flow intersections pose a major challenge to urban traffic safety, especially the early U-turn behavior of vehicles near mixed-flow intersections, which involves frequent interactions with oncoming straight-moving vehicles. Left-turning U-turning vehicles and oncoming straight-moving vehicles cannot obtain complete traffic information, and therefore cannot accurately predict each other's interactive behavior, which may lead to the risk of serious traffic accidents.

[0003] At present, although there are many studies devoted to the modeling of vehicle behavior in mixed-flow intersection scenarios, there is insufficient research on the early U-turn behavior of vehicles near mixed-flow intersections and the various important features that affect the behavior of such vehicles, and no clear modular modeling scheme has been established. Deep learning is considered to be a promising vehicle trajectory prediction method because of its advantages of end-to-end training and the ability to capture long-term dependencies. However, in terms of modeling the U-turn behavior of vehicles near mixed-flow intersections, existing deep learning models are not comprehensive enough in feature extraction, or are more often used as "black boxes" without being explored in more detail and depth. Therefore, a new technical solution is urgently needed to comprehensively solve the problems existing in the existing technology. Summary of the invention

[0004] The purpose of the present invention is to provide a U-turn intention and trajectory prediction method based on high-dimensional embedded deep parallel computing, which can accurately and quickly predict the U-turn vehicle trajectory, so that the autonomous driving vehicle can quickly and accurately predict the vehicle's U-turn intention and future U-turn trajectory, thereby making better decisions and avoiding traffic accidents.

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

[0006] A method for predicting dangerous U-turn intention and trajectory based on high-dimensional embedding deep parallel computing, comprising the following steps:

[0007] R1: Use a drone to shoot video and extract data, preprocess the data, and generate corresponding time series data V1 and V2, where V1 represents the time series trajectory of the vehicle in the left turn lane, and V2 represents the time series trajectory of the vehicle in the straight lane.

[0008] R2: Read the time series data V1 and V2 in step R1, and analyze the U-turn intention of V1, and divide the vehicles in the U-turn lane into vehicles with obvious U-turn intention, vehicles with implicit U-turn intention and vehicles without U-turn intention (i.e. straight-moving vehicles in the left-turn lane). The straight-moving vehicles and U-turn vehicles in the left-turn lane and the straight-moving vehicles in the opposite lane constitute the new input. The vehicle behavior information is embedded into the high-dimensional space through the embedding layer and used as the input for the next step.

[0009] R3: Input the high-dimensional embedding of step R2 into the parallel structure based on temporal convolution and spatial self-attention, extract temporal and spatial information respectively through the temporal convolution network and spatial self-attention mechanism, and finally merge these two parts of information into spatiotemporal features. Since this structure is a parallel structure, and the temporal convolution network and spatial self-attention mechanism are both structures that can be calculated in parallel, the performance of the entire model can be greatly improved.

[0010] R4: Normalize and perform residual connection operations on the spatiotemporal feature information obtained in step R3, and input it into the depthwise separable convolutional network of the lower layer to increase the fitting ability of the model and achieve higher accuracy. The depthwise separable convolution is divided into depthwise convolution and pointwise convolution, which are used for deep extraction of spatiotemporal features and feature fusion, respectively.

[0011] R5: The spatiotemporal features obtained after processing in step R4 are called the first type of output of the decoder; at the same time, the relative position coordinates of the predicted trajectory output at the previous moment are spliced ​​into a new output called the second type of output of the decoder through position encoding and input embedding. These two types of outputs are used as the input of the decoder, and the feature representation of the future trajectory coordinates is output after decoding; the feature representation output after decoding is input into the trajectory generator, and finally the future trajectory information of all traffic participants is output. where t obs+1 ,t pre represent the observation time and prediction time respectively.

[0012] Wherein, step R1 is specifically as follows:

[0013] R1-1. Use drones to shoot videos of U-turns near intersections during the morning and evening rush hours. The video field of view should at least include vehicles on the straight lane and the opposite left-turn lane;

[0014] R1-2. Extract the vehicle trajectory sequence on the left-turn lane and the opposite straight lane in the video. The proposed time series data should contain the following vehicle behavior information: x, y, l, w, θ, τ, where x and y represent the global coordinates of the vehicle, l and w represent the length and width of the vehicle respectively, θ represents the actual heading of the vehicle, and τ represents the category of the vehicle;

[0015] R1-3. Preprocess the extracted trajectory data and generate corresponding time series data V1 and V2, where V1 represents the time series trajectory of the vehicle in the left turn lane, and V2 represents the time series trajectory of the vehicle in the straight lane.

[0016] For the preprocessing operations of trajectory data, please refer to "Vehicle Trajectory Data Preprocessing Based on Hadoop" by Hu Siqin and Wu Shaofei.

[0017] Wherein, step R2 is specifically as follows:

[0018] R2-1, according to the actual heading θ and θ of the vehicle turn The relationship between θ and turn Minimum angle indicating that the vehicle has a clear intention to turn:

[0019] When θ ≥ θ turn When , the output a can be obtained, indicating a vehicle with an obvious intention to turn around;

[0020] When θ<θ turn When , the output b can be obtained, indicating that the vehicle does not have an obvious U-turn intention, so it is necessary to further analyze whether the vehicle has an implicit U-turn intention;

[0021] R2-2, input the above vehicles without explicit U-turn intention into the dilated causal convolutional network, and output 0 / 1 through the feedforward connection layer to further determine whether the vehicle has implicit U-turn intention;

[0022] R2-3, multiply the output 0 / 1 by the output b obtained in step R2-1, and combine them with the output a obtained in step R2-1 to obtain a new output c and output d, where output c represents a U-turn vehicle in the left-turn lane, and output d represents a straight-moving vehicle in the left-turn lane;

[0023] R2-4, combine the output c and output d of step R2-3 with V2 in step R1 into a new input, embed the vehicle's behavior information into a high-dimensional space through the embedding layer, and use it as the input of the next step.

[0024] Wherein, step R3 is specifically as follows:

[0025] R3-1. Input the high-dimensional embedding of step R3 into the spatial self-attention mechanism. The specific steps are as follows: For each vehicle i in the traffic scene of the tth frame, embed the upper layer output into Through three different learnable weights W q , W k , W v Get query separately Keywords Sum

[0026]

[0027] At the same time, the i-th U-turn vehicle in the t-th frame in this traffic scene is defined as the ego vehicle, and the message transmission between this vehicle and other vehicles j in the homogeneous and heterogeneous traffic flows is:

[0028]

[0029] Divide the resulting message passing by Scaling is performed and the softmax function is used to obtain the value v of the jth vehicle j The weight on the , and multiply the weight by v j Get the attention value between vehicle j and ego vehicle i, and sum all the messages sent from vehicle j to ego vehicle i to get the single-head attention of vehicle i:

[0030]

[0031] And query Keywords Sum After h different linear projections, the attention function is executed in parallel to obtain a dimension of d v The output of all the single-head attentions is concatenated and a linear projection operation is performed to obtain the spatial feature information of the ego car and other vehicles.

[0032]

[0033] Among them, W o represents the learnable weights;

[0034] R3-2. Input the high-dimensional embedding of step R2 into the temporal convolutional network. The temporal convolutional network benefits from its unique convolutional structure, causal dilated convolution, which enables it to expand the receptive field exponentially without leaking future information or losing feature information, and shows good performance in trajectory prediction of dense traffic. For traditional convolutional networks, the size of the receptive field can only be expanded by increasing the size of the convolution kernel or the number of convolutional layers, but increasing the size of the convolution kernel will increase the amount of calculation and memory consumption, and increasing the number of convolutional layers will lead to too many model parameters and prone to overfitting; in contrast, the dilated convolutional network introduces a dilation factor inside the convolution kernel, which expands the receptive field without increasing the size of the convolution kernel and the number of layers, which can effectively improve the performance of the model.

[0035] The specific steps are as follows: For the trajectory sequence of the i-th vehicle in time period T Given a convolution kernel f = (f 1,f 2 ,…,f K ), trajectory sequence X i Elements on The dilated convolution operation can be defined as F:

[0036]

[0037] Where *d represents the dilated convolution operator, d represents the dilation factor, K represents the number of convolution kernels, and x t-(K-k)d Indicates the location of past historical information. When d = 1, it indicates a standard convolution operation, that is, no dilated convolution is introduced. When d>1&d∈N * When , it means that the dilated convolution operation is introduced, which is equivalent to adding d-1 holes between every two convolution kernel elements. Generally speaking, d = 2 i , where i represents the i-th layer of the network, and the size of d grows exponentially with the depth of the network. By using a larger expansion factor, we can ensure that the overall network structure can cover a longer period of historical information, thereby expanding the network's receptive field more efficiently. At the same time, the following formula is used to calculate the receptive field rf of the dilated convolution:

[0038]

[0039] Where k(i) represents the size of the convolution kernel of the i-th layer, and d(i) represents the dilation factor of the i-th layer. In order to solve the problem of gradient disappearance caused by the increase of network depth, a general residual module is used to increase the stability of TCN. The residual module can be further expressed as:

[0040] output = X i +F i (X i )

[0041] Where output is the output of the residual module, F i Represents operations such as dilated causal convolution, non-linear activation, and normalization.

[0042] Wherein, step R4 is specifically as follows:

[0043] R4-1, first is the deep convolution operation, which only performs convolution operation on each channel of the spatiotemporal feature information input in step R3. More specifically, for the input C in The feature map of channels, the deep convolution layer uses C in A 3x 3 convolution kernel is used to convolve each channel to obtain C in The convolutional feature map of channels.

[0044] R4-2. After step R4-1, use the ReLU activation function to perform nonlinear transformation on the feature map, which can enhance the nonlinear expression ability of the model and make the model more flexible and adaptable.

[0045] R4-3, the point-by-point convolution layer convolves the convolution feature map obtained by the depth convolution, uses a 1x 1 convolution kernel to convolve each channel, and convolves the C output of the depth convolution layer in The feature maps of the channels are linearly transformed and feature fused, and the output is C out The main function of the point-by-point convolution layer is to fuse the feature information of different channels, and it also reduces the dimension of the feature map and reduces the amount of calculation.

[0046] R4-4. After step R4-3, the ReLU activation function is also used to perform nonlinear transformation on the feature map.

[0047] The U-turn intention and trajectory prediction method based on high-dimensional embedded deep parallel computing provided in the above technical solution has the following beneficial effects compared with the prior art:

[0048] 1. The present invention utilizes the encoder-decoder architecture and adopts parallel computing in the encoder. It extracts the temporal correlation between vehicles through a temporal convolutional network and calculates the relative position relationship and feature similarity between vehicles through a spatial self-attention mechanism, thereby effectively extracting feature information between data in spatiotemporal prediction tasks. Compared with traditional methods, the present invention makes full use of the advantages of parallel computing and solves the problems that the existing trajectory prediction methods cannot accurately judge the vehicle's intention to turn around, as well as the slow speed and low accuracy of vehicle trajectory prediction. The present invention is suitable for an autonomous driving vehicle on a straight lane to accurately and quickly predict the vehicle's intention to turn around in the opposite left-turn lane and the future trajectory of the turning vehicle.

[0049] Specifically, the trajectory of the U-turning vehicle is predicted by using the parallel calculation of temporal convolution and spatial attention. Since the temporal convolution network and the spatial self-attention mechanism are structures that can be calculated in parallel, and the two structures process temporal or spatial information respectively, unnecessary data dependence and communication are avoided, thereby further improving the calculation speed and efficiency. At the same time, the deep separable convolution is used to deeply extract spatiotemporal features and feature fusion, and the residual structure is used to ensure the accuracy of the prediction.

[0050] 2. The present invention adopts a method based on dilated causal convolution and supplemented by vehicle heading condition constraints to make a fast and accurate prediction of the vehicle's U-turn intention in the left turn lane; for the U-turn intention prediction task in dense traffic, not only multi-scale contextual information is required, but also a sufficiently large receptive field, and the method of the present invention can just meet these two conditions; by utilizing asynchronous and parallel convolution operations, it has obvious advantages in training speed and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A U-turn scene diagram near an intersection in an embodiment of the present invention;

[0052] Figure 2 It is the overall model structure diagram of the present invention;

[0053] Figure 3 It is a dangerous U-turn intention prediction module in the overall model diagram of the present invention;

[0054] Figure 4 It is a flow chart of a method for predicting U-turn intention and trajectory based on high-dimensional embedded deep parallel computing of the present invention;

[0055] Figure 5 Extracting spatial information between vehicles by the spatial self-attention mechanism in the embodiment of the present invention;

[0056] Figure 6 A comparison diagram of the actual vehicle trajectory and the predicted vehicle trajectory provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is specifically described below in conjunction with embodiments. It should be understood that the following text is only used to describe one or several specific embodiments of the present invention, and does not strictly limit the scope of protection of the specific claims of the present invention.

[0058] This embodiment takes the U-turn scene near the intersection of Haitang Bay in Hefei as an example, and combines the following Figures 1 to 6 , the present invention is further analyzed and explained.

[0059] The U-turn intention and trajectory prediction method based on high-dimensional embedding deep parallel computing includes the following steps:

[0060] Step 1: Use drones to shoot videos and extract data, and pre-process the data:

[0061] 1.1 Use drones to photograph the U-turn near the intersection during the morning and evening rush hours (reference Figure 1 ) video, the video field of view should at least include the vehicles in the left turn lane and the opposite straight lane;

[0062] 1.2 Extract the vehicle trajectory sequence on the left-turn lane and the opposite straight lane in the video. The proposed time series data should contain the following vehicle behavior information: x, y, l, w, θ, τ, where x and y represent the global coordinates of the vehicle, l and w represent the length and width of the vehicle respectively, θ represents the actual heading of the vehicle, and τ represents the category of the vehicle;

[0063] 1.3 Preprocess the extracted trajectory data and generate corresponding time series data V1 and V2, where V1 represents the time series trajectory of the vehicle in the left turn lane, and V2 represents the time series trajectory of the vehicle in the straight lane.

[0064] Step 2: read the time series data V1 and V2 in step 1, and analyze the U-turn intention of V1, and divide the vehicles on the U-turn lane into vehicles with obvious U-turn intention, vehicles with implicit U-turn intention, and vehicles without U-turn intention (i.e., straight-moving vehicles in the left-turn lane), specifically:

[0065] 2.1 According to the actual heading of the vehicle θ and θ turn (the minimum angle at which the vehicle has an obvious U-turn intention) to determine whether the vehicle has an obvious U-turn intention;

[0066] When θ ≥ θ turn When , the output a can be obtained, indicating a vehicle with an obvious intention to turn around;

[0067] When θ<θ turn When , the output b can be obtained, indicating that the vehicle does not have an obvious U-turn intention, so it is necessary to further analyze whether the vehicle has an implicit U-turn intention;

[0068] 2.2 The above vehicles without explicit U-turn intention are input into the dilated causal convolutional network, and 0 / 1 is output through the feedforward connection layer to further determine whether the vehicle has implicit U-turn intention;

[0069] 2.3 Multiply the output 0 / 1 with the output b obtained in step 2.1, and combine it with the output a obtained in step 2.1 to obtain a new output c and output d, where output c represents a U-turn vehicle in the left turn lane, and output d represents a straight-moving vehicle in the left turn lane.

[0070] 2.4 Combine the output c and output d of step 2.3 with V2 in step 1 into a new input, embed the vehicle's behavior information into a high-dimensional space through the embedding layer, and use it as the input of the next step.

[0071] Step 3: Input the high-dimensional embedding of step 2 into the parallel structure based on temporal convolution and spatial self-attention, extract temporal and spatial information respectively through the temporal convolution network and spatial self-attention mechanism, and finally merge these two parts of information into spatiotemporal features. Since this structure is a parallel structure, and the temporal convolution network and spatial self-attention mechanism are both networks that can be calculated in parallel, the performance of the entire model can be greatly improved.

[0072] 3.1 Input the high-dimensional embedding in step 2 into the spatial self-attention mechanism. The specific steps are as follows: For each vehicle i in the traffic scene of the tth frame, embed the upper layer output into Through three different learnable weights W q , W k , W v Get query separately Keywords Sum

[0073]

[0074] At the same time, the i-th U-turn vehicle in the t-th frame in this traffic scene is defined as the ego vehicle, and the message transmission between this vehicle and other vehicles j in the homogeneous and heterogeneous traffic flows is:

[0075]

[0076] Divide the resulting message passing by Scaling is performed and the softmax function is used to obtain the value v of the jth vehicle j The weight on the , and multiply the weight by v j Get the attention value between vehicle j and ego vehicle i, and sum the message transmissions sent by all vehicles j to ego vehicle i to get the single-head attention of vehicle i, as follows: Figure 5 As shown:

[0077]

[0078] And query Keywords Sum After h different linear projections, the attention function is executed in parallel to obtain a dimension of d v The output of all the single-head attentions is concatenated and a linear projection operation is performed to obtain the spatial feature information of the ego car and other vehicles.

[0079]

[0080] Where W o represents the learnable weights;

[0081] 3.2 Input the high-dimensional embedding in step 2 into the temporal convolutional network. The temporal convolutional network benefits from its unique convolutional structure, causal dilated convolution, which enables it to expand the receptive field exponentially without leaking future information or losing feature information, and shows good performance in trajectory prediction of dense traffic. For traditional convolutional networks, the size of the receptive field can only be expanded by increasing the size of the convolution kernel or the number of convolutional layers, but increasing the size of the convolution kernel will increase the amount of calculation and memory consumption, and increasing the number of convolutional layers will lead to too many model parameters and prone to overfitting; in contrast, the dilated convolutional network introduces a dilation factor inside the convolution kernel, which expands the receptive field without increasing the size of the convolution kernel and the number of layers, which can effectively improve the performance of the model.

[0082] The specific steps are as follows: For the trajectory sequence of the i-th vehicle in time period T Given a convolution kernel f = (f 1 ,f 2 ,…,f K ), trajectory sequence X i Elements on The dilated convolution operation can be defined as F:

[0083]

[0084] Where *d represents the dilated convolution operator, d represents the dilation factor, K represents the number of convolution kernels, and x t-(K-k)d Indicates the positioning of past historical information. When d = 1, it indicates a standard convolution operation, that is, no dilated convolution is introduced.

[0085] When d>1&d∈N * When , it means that the dilated convolution operation is introduced, which is equivalent to adding d-1 holes between every two convolution kernel elements. Generally speaking, d = 2 i , where i represents the i-th layer of the network, and the size of d grows exponentially with the depth of the network. By using a larger expansion factor, we can ensure that the overall network structure can cover a longer period of historical information, thereby expanding the network's receptive field more efficiently. At the same time, the following formula is used to calculate the receptive field rf of the dilated convolution:

[0086]

[0087] Where k(i) represents the size of the convolution kernel of the i-th layer, and d(i) represents the dilation factor of the i-th layer. In order to solve the problem of gradient disappearance caused by the increase of network depth, a general residual module is used to increase the stability of TCN. The residual module can be further expressed as:

[0088] output = X i +F i (Xi )

[0089] Where output is the output of the residual module, F i Represents operations such as dilated causal convolution, nonlinear activation, and normalization, as shown in Figure 4 shown.

[0090] Step 4: Normalize and residually connect the spatiotemporal feature information obtained in step 3, and input it into the depthwise separable convolutional network of the lower layer to increase the fitting ability of the model and achieve higher accuracy. The depthwise separable convolution is divided into depthwise convolution and pointwise convolution, which are used for deep extraction of spatiotemporal features and feature fusion, respectively.

[0091] 4.1 The first is the deep convolution operation, which only performs convolution operations on each channel of the spatiotemporal feature information input in step 3. More specifically, for the input C in The feature map of channels, the deep convolution layer uses C in A 3x3 convolution kernel is used to convolve each channel to obtain C in The convolutional feature map of channels.

[0092] 4.2 After step 4.1, using the ReLU activation function to perform nonlinear transformation on the feature map can enhance the nonlinear expression ability of the model, making the model more flexible and adaptable.

[0093] 4.3 The point-by-point convolution layer convolves the convolution feature map obtained by the depth convolution, uses a 1x1 convolution kernel to convolve each channel, and convolves the C output of the depth convolution layer. in The feature maps of the channels are linearly transformed and feature fused, and the output is C out The main function of the point-by-point convolution layer is to fuse the feature information of different channels, and it also reduces the dimension of the feature map and reduces the amount of calculation.

[0094] 4.4 After step 4.3, the ReLU activation function is also used to perform nonlinear transformation on the feature map.

[0095] Step 5: Decode through the decoder and output the feature representation of the future trajectory coordinates, then input the feature representation into the trajectory generator, and finally output the future trajectory information of all traffic participants.

[0096] 5.1 The spatiotemporal features obtained after processing in step 4 are called the first type of output of the decoder; at the same time, the relative position coordinates of the predicted trajectory output at the previous moment are spliced ​​into a new output called the second type of output of the decoder after position encoding and input embedding. These two types of outputs are used as the input of the decoder, and after decoding, the feature representation of the future trajectory coordinates is output.

[0097] 5.2 Input the feature representation output after decoding in step 5 into the trajectory generator, and finally output the future trajectory information of all traffic participants where t obs+1 ,t pre Represent the observation time and prediction time respectively. Figure 6 : is a schematic diagram comparing the actual vehicle trajectory and the predicted vehicle trajectory of this embodiment. It can be seen that the actual vehicle trajectory and the predicted vehicle trajectory of this embodiment are substantially the same, which verifies the effectiveness of the method of the present invention.

[0098] In the U-turn intention and trajectory prediction method provided in this embodiment, a method based on dilated causal convolution and supplemented by vehicle heading condition restriction is adopted to make a fast and accurate prediction of the U-turn intention of the vehicle in the left turn lane; at the same time, a method of temporal convolution and spatial attention parallel calculation is adopted to predict the trajectory of the U-turn vehicle. Since the temporal convolution network and the spatial self-attention mechanism are structures that can be calculated in parallel, and the two structures process time or space information respectively, unnecessary data dependence and communication are avoided, thereby further improving the calculation speed and efficiency.

[0099] At the same time, the present invention uses deep separable convolution to deeply extract spatiotemporal features and feature fusion, and adopts a residual structure, which ensures the accuracy of the prediction. In summary, the method of the present invention effectively solves the problems that the existing trajectory prediction method cannot make accurate judgments on the vehicle's intention to turn around, and the vehicle trajectory prediction speed is slow and the accuracy is not high; it not only has a significant improvement in prediction accuracy compared to the existing model, but also has a higher calculation speed, so it is more suitable for deployment in the real world.

[0100] The embodiments of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. For ordinary technicians in this technical field, after knowing the contents recorded in the present invention, they can make several equivalent changes and substitutions without departing from the principle of the present invention. These equivalent changes and substitutions should also be regarded as belonging to the protection scope of the present invention.

Claims

1. A U-turn intention and trajectory prediction method based on high-dimensional embedded deep parallel computing, characterized in that: The following steps are involved: R1: Shoot the video of the U-turn near the intersection and extract the data, pre-process the data, and generate the corresponding V1 and V2; Where V1 represents the time-series trajectory of the vehicle in the left-turn lane, and V2 represents the time-series trajectory of the vehicle in the straight lane; R2: Read V1 and V2 in step R1, and analyze V1 for U-turn intention, and divide the vehicles on the U-turn lane into vehicles with obvious U-turn intention, vehicles with implicit U-turn intention, and vehicles without U-turn intention, wherein the vehicles without U-turn intention are the straight-moving vehicles on the left-turn lane; The straight-moving vehicles, U-turning vehicles in the left-turn lane, and straight-moving vehicles in the opposite lane are used as new inputs. The vehicle behavior information is embedded into a high-dimensional space through the embedding layer and used as the input for the next step. R3: Input the high-dimensional embedding of step R2 into the parallel structure based on temporal convolution and spatial self-attention, extract temporal and spatial information respectively through the temporal convolution network and spatial self-attention mechanism, and finally merge these two parts of information into spatiotemporal features; R4: Normalize and perform residual connection operations on the spatiotemporal feature information obtained in step R3, and input it into the depth-separable convolutional network of the lower layer; R5: The spatiotemporal features obtained after processing in step R4 are called the first type of output of the decoder. At the same time, the relative position coordinates of the predicted trajectory output at the previous moment are spliced ​​into a new output called the second type of output of the decoder through position encoding and input embedding. These two types of outputs are used as the input of the decoder, and the feature representation of the future trajectory coordinates is output after decoding. This feature representation is then input into the trajectory generator, which ultimately outputs the future trajectory information of all traffic participants.

2. The method for predicting U-turn intention and trajectory based on high-dimensional embedded deep parallel computing according to claim 1 is characterized in that: Step R1 is specifically as follows: R1-1: Shoot a video of a U-turn near an intersection during the morning and evening rush hours. The video field of view includes at least the vehicles in the left-turn lane and the opposite straight lane. R1-2: Extract the vehicle trajectory sequence on the left-turn lane and the opposite straight lane in the video. The proposed time series data contains the following vehicle behavior information: x, y, l, w, θ, τ; where x and y represent the global coordinates of the vehicle, l and w represent the length and width of the vehicle respectively, θ represents the actual heading of the vehicle, and τ represents the category of the vehicle; R1-3: Preprocess the extracted trajectory sequence data and generate corresponding time series trajectories V1 and V2.

3. The method for predicting U-turn intention and trajectory based on high-dimensional embedded deep parallel computing according to claim 1, characterized in that: Step R2 is specifically: R2-1: According to the actual heading of the vehicle θ and θ turn The relationship between θ and turn The minimum angle indicating that the vehicle has an obvious intention to turn around; When θ ≥ θ turn When , the output a is obtained, indicating a vehicle with an obvious intention to turn around; When θ<θ turn When , the output b is obtained, indicating that the vehicle has no obvious intention to turn around; R2-2: Input the above vehicles without obvious U-turn intention into the dilated causal convolutional network, and output 0 / 1 through the feedforward connection layer to further determine whether the vehicle has an implicit U-turn intention; R2-3: Multiply the output 0 / 1 by the output b obtained in step R2-1, and combine it with the output a obtained in step R2-1 to obtain a new output c and output d, where output c represents a U-turn vehicle in the left-turn lane, and output d represents a straight-moving vehicle in the left-turn lane; R2-4: Combine the output c and output d of step R2-3 with V2 in step R1 into a new input, embed the vehicle's behavior information into a high-dimensional space through the embedding layer, and use it as the input of step R3.

4. The method for predicting U-turn intention and trajectory based on high-dimensional embedded deep parallel computing according to claim 1, characterized in that: Step R3 is specifically: R3-1: Input the high-dimensional embedding in step R2 into the spatial self-attention mechanism. The specific steps are as follows: For each vehicle i in the traffic scene of the tth frame, embed the upper layer output into Through three different learnable weights W q , W k , W v Get query separately Keywords Sum in: At the same time, the i-th U-turn vehicle in the t-th frame of the traffic scene is defined as the ego vehicle, and the message transmission between this vehicle and other vehicles j in the homogeneous and heterogeneous traffic flows is: in: Divide the resulting message passing by Scaling is performed and the softmax function is used to obtain the value v of the jth vehicle j The weight on the , and multiply the weight by v j Get the attention value between vehicle j and ego vehicle i, and sum all the messages sent from vehicle j to ego vehicle i to get the single-head attention of vehicle i: And query Keywords Sum After h different linear projections, the attention function is executed in parallel to obtain a dimension of d v The output of all the single-head attentions is concatenated and a linear projection operation is performed to obtain the spatial feature information of the ego car and other vehicles: Where W o represents the learnable weights; R3-2: Input the high-dimensional embedding of step R2 into the temporal convolutional network. The specific steps are as follows: For the trajectory sequence of the i-th vehicle in time period T Given a convolution kernel f = (f1, f2, ..., f K ), trajectory sequence X i Elements on The dilated convolution operation can be defined as F: Where *d represents the dilated convolution operator, d represents the dilation factor, K represents the number of convolution kernels, and x t-(K-k)d Indicates the location of past historical information; When d = 1, it indicates a standard convolution operation, that is, no dilated convolution is introduced; When d>1&d∈N * When , it means that the dilated convolution operation is introduced, which is equivalent to adding d-1 holes between every two convolution kernel elements; Where d = 2 i , i represents the i-th layer of the network, and the size of d grows exponentially with the depth of the network. At the same time, the following formula is used to calculate the receptive field rf of the dilated convolution: Where k(i) represents the size of the convolution kernel of the i-th layer, and d(i) represents the expansion factor of the i-th layer. In order to solve the problem of gradient disappearance caused by the increase of network depth, a general residual module is used to increase the stability of TCN. The residual module is expressed as: output=X i +F i (X i ) Where output is the output of the residual module, F i Represents dilated causal convolution, nonlinear activation, and normalization operations.

5. The method for predicting U-turn intention and trajectory based on high-dimensional embedded deep parallel computing according to claim 1, characterized in that: The depth-separable convolution in step R4 is divided into depth convolution and point-by-point convolution, which are used for deep extraction of spatiotemporal features and feature fusion respectively; and step R4 specifically The following steps are involved: R4-1: The first is a deep convolution operation, which only performs a convolution operation on each channel of the spatiotemporal feature information output in step R3; for the input C in The feature map of channels, the deep convolution layer uses C in A 3×3 convolution kernel is used to convolve each channel to obtain C in Convolutional feature map of channels; R4-2: After step R4-1, use the ReLU activation function to perform nonlinear transformation on the feature map; R4-3: The point-by-point convolution layer convolves the convolution feature map obtained by the depth convolution, uses a 1×1 convolution kernel to convolve each channel, and convolves the C output of the depth convolution layer. in The feature maps of the channels are linearly transformed and feature fused, and the output is C out Feature map of each channel; R4-4: After step R4-3, the ReLU activation function is also used to perform nonlinear transformation on the feature map.

Images