A vehicle trajectory prediction method based on drivable area feature extraction
By utilizing historical vehicle trajectories and high-precision maps to generate drivable area features, and combining historical motion state information to generate endpoint feature codes, the problem of trajectory prediction exceeding the drivable area is solved, achieving more accurate and reliable trajectory prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG LEAPMOTOR TECH CO LTD
- Filing Date
- 2022-09-20
- Publication Date
- 2026-07-07
AI Technical Summary
Existing trajectory prediction methods cannot guarantee that the predicted future trajectory will be within the drivable area, which may violate traffic regulations.
Driving areas are generated using historical trajectory information and high-precision map information of vehicles. Features of driving areas are extracted and combined with historical motion state information of the vehicle to be predicted to generate destination information. The destination information is feature-encoded and incorporated into the encoder, and then the final predicted trajectory information is generated through the decoder.
It improves the accuracy of trajectory prediction results, reduces the probability of trajectory prediction results exceeding the drivable area, and ensures the authenticity, reliability, and usability of trajectory prediction results.
Smart Images
Figure CN115703486B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of autonomous driving technology, and more specifically to a vehicle trajectory prediction method based on drivable area feature extraction. Background Technology
[0002] Autonomous driving technology is booming. To achieve fully autonomous driving, autonomous vehicles need to accurately predict the intentions and trajectories of surrounding vehicles and pedestrians, predicting their driving or walking trajectories over a future period, and controlling the autonomous vehicle to effectively avoid collisions. Simultaneously, surrounding vehicles must adhere to road regulations and travel within the drivable area. This requires trajectory prediction algorithms to ensure the predicted trajectory is accurate, reliable, and within the drivable area. Existing technologies primarily encode the vehicle's historical state information, employ attention mechanisms to extract features of the vehicle to be predicted, as well as its positional relationships with other vehicles, and then use a decoder to generate the future trajectory; or, based on historical information and interaction information with surrounding vehicles, generate the vehicle's intention information, and then use lane constraints and dynamics methods to predict the future trajectory. While these methods achieve the prediction of future trajectory information, they cannot guarantee that the predicted future trajectory will remain within the drivable area; that is, the predicted future trajectory may exceed the drivable area and violate traffic regulations. Therefore, there is an urgent need for a vehicle trajectory prediction method that can guarantee the predicted future trajectory complies with road regulations and stays within the drivable area. Summary of the Invention
[0003] This invention primarily addresses the problem that existing trajectory prediction methods cannot guarantee that the predicted trajectory lies within a drivable area. It provides a vehicle trajectory prediction method based on drivable area feature extraction. Utilizing historical vehicle trajectory information and high-precision map information, it generates the drivable area of the vehicle to be predicted, extracts features from this drivable area, and combines this with the vehicle's historical motion state information to generate endpoint information. This endpoint information is then feature-encoded and incorporated into an encoder, and finally, a decoder generates the final predicted trajectory information. This invention integrates drivable area feature information into the trajectory prediction model, improving the accuracy of the trajectory prediction results, reducing the probability of the predicted trajectory exceeding the drivable area, and ensuring the authenticity, reliability, and usability of the predicted trajectory results.
[0004] To achieve the above objectives, the present invention adopts the following technical solution:
[0005] A vehicle trajectory prediction method based on drivable area feature extraction includes the following steps:
[0006] Step S1: Extract the temporal features f of the vehicle to be predicted. agentTemporal characteristics f of the vehicles surrounding the vehicle to be predicted N , fusion f agent and f N Obtain the interaction features f social ;
[0007] Step S2: Generate the drivable area of the vehicle to be predicted, and extract the feature vector f of the drivable area. feasible , fusion f feasible and f agent Obtain the fusion feature f driver Further generate the predicted endpoint coordinates p last ;
[0008] Step S3: For the predicted endpoint coordinates p last Feature encoding is performed to obtain the representation vector f. last , fusion f last with f agent and f social Obtain the encoded feature f encoder , encode feature f encoder Decoding is performed to obtain the decoding feature f decoder The final predicted trajectory coordinate information C is then output.
[0009] This invention utilizes historical trajectory information and high-precision map information of vehicles to generate drivable areas for the vehicle to be predicted, extracts features from these drivable areas, and combines this with the historical motion state information of the vehicle to generate endpoint information. This endpoint information is then feature-encoded and incorporated into an encoder, and finally, a decoder generates the final predicted trajectory information. This invention integrates the feature information of drivable areas into the trajectory prediction model, improving the accuracy of the trajectory prediction results, reducing the probability of the predicted trajectory exceeding the drivable area, and ensuring the authenticity, reliability, and usability of the trajectory prediction results.
[0010] Preferably, in step S1, the temporal features f of the vehicle to be predicted are extracted. agent The specific process includes the following steps: Step A1: Obtain the historical motion state information of the vehicle to be predicted at T time points. The historical motion state information includes position coordinate information (x, y) and velocity information (v). x ,v y ), acceleration information (a x ,a y ) and heading angle information h;
[0011] Step A2: Use a one-dimensional convolutional neural network or a fully connected network to encode the T historical motion state information of the vehicle to be predicted, and obtain the high-dimensional features corresponding to the historical motion state information of the vehicle to be predicted.
[0012] Step A3: Input the high-dimensional features corresponding to the historical motion state information of the vehicle to be predicted into the serialization prediction model, and extract the temporal features f of the vehicle to be predicted. agent ;
[0013] Serialization prediction models include, but are not limited to, recurrent neural networks (RNNs), long short-term memory networks (LSTMs), gated recurrent units (GRUs), and other temporal networks.
[0014] Preferably, in step S1, the temporal features f of the surrounding vehicles of the vehicle to be predicted are extracted. N The specific process includes the following steps:
[0015] Step B1: Based on the current position coordinates of the vehicle to be predicted, obtain the historical motion state information of the nearest N surrounding vehicles at T time points. The historical motion state information includes position coordinate information (x, y) and velocity information (v). x ,v y ), acceleration information (a x ,a y ) and heading angle information h;
[0016] Step B2: Use a one-dimensional convolutional neural network or a fully connected network to encode the T historical motion state information of the surrounding vehicles to obtain the high-dimensional features corresponding to the historical motion state information of the surrounding vehicles.
[0017] Step B3: Input the high-dimensional features corresponding to the historical motion state information of surrounding vehicles into the serialization prediction model to extract the temporal features f of the surrounding N vehicles. N ;
[0018] Serialization prediction models include, but are not limited to, recurrent neural networks (RNNs), long short-term memory networks (LSTMs), gated recurrent units (GRUs), and other temporal networks.
[0019] Preferably, step S2 includes the following steps:
[0020] Step C1: Based on the current location coordinates and driving direction of the vehicle to be predicted, obtain a high-precision map of the current location;
[0021] Step C2: Using the current motion state of the vehicle to be predicted, select a map area within a specified range in the high-precision map, and construct the area into a single-channel map area with a width of W and a height of H. Set the pixels of the drivable area to 1 and the pixels of the non-drivable area to 0.
[0022] Step C3: Extract the feature vector f of the drivable region using a convolutional neural network (CNN) and pooling layers. feasible ;
[0023] Step C4: Extract the feature vector f of the drivable region. feasible and the temporal characteristics f of the vehicle to be predicted agent The fusion is performed to obtain the fusion feature f. driver ;
[0024] Step C5: The fusion feature f driver Input the fully connected layer to generate the predicted endpoint coordinates p last .
[0025] Surrounding vehicles must adhere to traffic regulations, meaning they must remain within the drivable area. This invention constructs the drivable area for surrounding vehicles, extracts its feature information, and incorporates it into the encoder for trajectory prediction. The construction and feature extraction of the drivable area improves trajectory prediction accuracy, ensuring the predicted trajectory remains within the drivable road surface and reducing the probability of the predicted trajectory veering off the road. Non-drivable areas include areas outside curbs, flower beds, and fences.
[0026] Preferably, step S3 includes the following steps:
[0027] Step D1: Use a fully connected layer to predict the endpoint coordinates p last Feature encoding is performed to obtain the representation vector f. last ;
[0028] Step D2: Convert the representation vector f last The temporal characteristics f of the vehicle to be predicted agent and interaction features f social By fusing the coded features, we obtain the encoded features f. encoder ;
[0029] Step D3: Encode feature f encoder The input is a decoder, which uses a serialized feature model for decoding and outputs decoded features f. decoder ;
[0030] Step D4: Decode feature f decoder The input is a fully connected layer, and the output is the final predicted trajectory coordinate information C, where the predicted trajectory coordinate information C contains K predicted trajectory coordinate points, C1, C2, C3…C… K .
[0031] Based on the historical motion state information of the drivable area and the vehicle to be predicted, the endpoint coordinates within the drivable area are predicted. Features are extracted from the predicted endpoint coordinates, fused into the encoder, and then fed into the subsequent decoder to obtain the final predicted trajectory, thus improving the accuracy of trajectory prediction. Serialization feature models include, but are not limited to, recurrent neural networks (RNNs), long short-term memory networks (LSTMs), gated recurrent units (GRUs), and other temporal networks.
[0032] Preferably, the method further includes: using the mean squared error loss function (MSE) to supervise and iterate the predicted trajectory C and the true trajectory G, where G represents the true future trajectory coordinate information, containing K true trajectory coordinate points, G1, G2, G3…G K The accuracy of trajectory prediction can be improved by utilizing the mean squared error loss function (MSE).
[0033] As a preferred option, it also includes:
[0034] Step E1: Add the consistency loss function L con For the predicted endpoint coordinates p last and the final predicted trajectory endpoint coordinates C K By comparing the results, the predicted endpoint coordinates p are obtained. last The final predicted trajectory endpoint coordinates C K The mean square error;
[0035] Step E2: Combine the loss function L using weights α and β. traj and loss function L con Thus, the total loss function L is obtained;
[0036] Step E3: Iterate and optimize using the loss function L.
[0037] To make the predicted endpoint coordinates p last and the final predicted trajectory endpoint coordinates C K To maintain consistency, add a consistency loss function L. con For the directly predicted endpoint coordinates p last and the final predicted trajectory endpoint coordinates C K Compare the two results and calculate their mean squared errors. Loss function L traj and L con The total loss function L is obtained by combining weights α and β. The trajectory prediction model is then iterated and optimized using the loss function L to improve the accuracy of trajectory prediction. This invention supervises both the predicted endpoint coordinates and the final predicted endpoint coordinates of the complete trajectory to ensure consistency in the prediction results.
[0038] Preferably, the loss function L traj It is expressed as follows:
[0039]
[0040] Preferably, the loss function L con It is expressed as follows:
[0041] L con =||C K -p last ||2.
[0042] Preferably, the loss function L is expressed as follows:
[0043] L=αL traj +βL con .
[0044] Therefore, the advantages of the present invention are:
[0045] (1) Integrating the feature information of the drivable area into the trajectory prediction model improves the accuracy of the trajectory prediction results, reduces the probability of the trajectory prediction results exceeding the drivable area, and ensures the authenticity, reliability and usability of the trajectory prediction results.
[0046] (2) The loss function is used to supervise and iterate the predicted endpoint coordinates and the final predicted endpoint coordinates of the complete trajectory to ensure the consistency of the prediction results and thus improve the accuracy of trajectory prediction. Attached Figure Description
[0047] Figure 1 This is a flowchart of a vehicle trajectory prediction method based on drivable area feature extraction in an embodiment of the present invention.
[0048] Figure 2 This is a schematic block diagram of a vehicle trajectory prediction method based on drivable area feature extraction in an embodiment of the present invention. Detailed Implementation
[0049] The present invention will now be further described with reference to the accompanying drawings and specific embodiments.
[0050] A vehicle trajectory prediction method based on drivable area feature extraction, such as Figure 1 As shown, it includes the following steps:
[0051] Step S1: Extract the temporal features f of the vehicle to be predicted. agent Temporal characteristics f of the vehicles surrounding the vehicle to be predicted N , fusion f agent and f N Obtain the interaction features f social ;
[0052] In step S1, extract the temporal features f of the vehicle to be predicted. agent The specific process includes the following steps:
[0053] Step A1: Obtain the historical motion state information of the vehicle to be predicted at T time points. The historical motion state information includes position coordinates (x, y) and velocity information (v). x ,v y ), acceleration information (a x ,a y ) and heading angle information h;
[0054] Step A2: Use a one-dimensional convolutional neural network or a fully connected network to encode the T historical motion state information of the vehicle to be predicted, and obtain the high-dimensional features corresponding to the historical motion state information of the vehicle to be predicted.
[0055] Step A3: Input the high-dimensional features corresponding to the historical motion state information of the vehicle to be predicted into the serialization prediction model, and extract the temporal features f of the vehicle to be predicted. agent ;
[0056] Step S1 extracts the temporal features f of the surrounding vehicles of the vehicle to be predicted. N The specific process includes the following steps:
[0057] Step B1: Based on the current position coordinates of the vehicle to be predicted, obtain the historical motion state information of the nearest N surrounding vehicles at T time points. The historical motion state information includes position coordinate information (x, y) and velocity information (v). x ,v y ), acceleration information (a x ,a y ) and heading angle information h;
[0058] Step B2: Use a one-dimensional convolutional neural network or a fully connected network to encode the T historical motion state information of the surrounding vehicles to obtain the high-dimensional features corresponding to the historical motion state information of the surrounding vehicles.
[0059] Step B3: Input the high-dimensional features corresponding to the historical motion state information of surrounding vehicles into the serialization prediction model to extract the temporal features f of the surrounding N vehicles. N ;
[0060] Step S2: Generate the drivable area of the vehicle to be predicted, and extract the feature vector f of the drivable area. feasible , fusion f feasible and f agent Obtain the fusion feature f driver Further generate the predicted endpoint coordinates p last ;
[0061] The specific process of step S2 includes the following steps:
[0062] Step C1: Based on the current location coordinates and driving direction of the vehicle to be predicted, obtain a high-precision map of the current location;
[0063] Step C2: Using the current motion state of the vehicle to be predicted, select a map area within a specified range in the high-precision map, and construct the area into a single-channel map area with a width of W and a height of H. Set the pixels of the drivable area to 1 and the pixels of the non-drivable area to 0.
[0064] Step C3: Extract the feature vector f of the drivable region using a convolutional neural network (CNN) and pooling layers. feasible ;
[0065] Step C4: Extract the feature vector f of the drivable region. feasible and the temporal characteristics f of the vehicle to be predicted agent The fusion is performed to obtain the fusion feature f. driver ;
[0066] Step C5: Fuse features f driver Input the fully connected layer to generate the predicted endpoint coordinates p last ;
[0067] Step S3: For the predicted endpoint coordinates p last Feature encoding is performed to obtain the representation vector f. last , fusion f last with f agent and f social Obtain the encoded feature f encoder , encode feature f encoder Decoding is performed to obtain the decoding feature f decoder Furthermore, the final predicted trajectory coordinate information C is output;
[0068] The specific process of step S3 includes the following steps:
[0069] Step D1: Use a fully connected layer to predict the endpoint coordinates p last Feature encoding is performed to obtain the representation vector f. last ;
[0070] Step D2: Convert the representation vector f last The temporal characteristics f of the vehicle to be predicted agent and interaction features f social By fusing the coded features, we obtain the encoded features f. encoder ;
[0071] Step D3: Encode feature f encoder The input is a decoder, which uses a serialized feature model for decoding and outputs decoded features f. decoder ;
[0072] Step D4: Decode feature f decoder The input is a fully connected layer, and the output is the final predicted trajectory coordinate information C, where the predicted trajectory coordinate information C contains K predicted trajectory coordinate points, C1, C2, C3…C… K .
[0073] The mean squared error (MSE) loss function is used to supervise and iterate the predicted trajectory C and the true trajectory G, where G represents the true future trajectory coordinate information, containing K true trajectory coordinate points, G1, G2, G3…G K To ensure the predicted endpoint coordinates p last and the final predicted trajectory endpoint coordinates C K To maintain consistency, add a consistency loss function L. con For the directly predicted endpoint coordinates p last and the final predicted trajectory endpoint coordinates C K Compare the two results and calculate their mean squared errors. Loss function L traj and L con The total loss function L is obtained by combining the weights α and β. The trajectory prediction model is then iterated and optimized using the loss function L to improve the accuracy of trajectory prediction.
[0074] Loss function L traj It is expressed as follows:
[0075]
[0076] Loss function L con It is expressed as follows:
[0077] L con =||C K -p last ||2;
[0078] The loss function L is expressed as follows:
[0079] L=αL traj +βL con .
[0080] like Figure 2 As shown, this embodiment utilizes the vehicle's historical trajectory information and high-precision map information to generate the drivable area of the vehicle to be predicted, extracts features from the drivable area, and combines this with the vehicle's historical motion state information to generate endpoint information. The endpoint information is then feature-encoded and incorporated into the encoder, and finally, the decoder generates the final predicted trajectory information. This embodiment incorporates the feature information of the drivable area into the trajectory prediction model, reducing the probability that the predicted trajectory will exceed the drivable area. Based on the drivable area and the vehicle's historical motion state information, the endpoint coordinates within the drivable area are predicted. Features of the predicted endpoint coordinates are extracted, fused into the encoder, and then fed into the subsequent decoder to obtain the final predicted trajectory, thus improving the accuracy of trajectory prediction. Serialization prediction models / serialization feature models include, but are not limited to, recurrent neural networks (RNNs), long short-term memory networks (LSTMs), gated recurrent units (GRUs), and other temporal networks.
[0081] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A vehicle trajectory prediction method based on drivable area feature extraction, characterized in that, Includes the following steps: Step S1: Extract the temporal features of the vehicle to be predicted. Temporal characteristics of vehicles surrounding the vehicle to be predicted fusion and Obtain interaction features; Obtain the historical motion state information of the vehicle to be predicted at T time points; A one-dimensional convolutional neural network or a fully connected network is used to encode the T historical motion state information of the vehicle to be predicted, obtaining high-dimensional features corresponding to the historical motion state information of the vehicle to be predicted; these features are then input into a serialization prediction model to extract the temporal features of the vehicle to be predicted. ; Step S2: Generate the drivable area of the vehicle to be predicted and extract the feature vector of the drivable area. fusion and Obtain fusion features Further generate predicted endpoint coordinates ; Step S3: Predict the endpoint coordinates Feature encoding is performed to obtain the representation vector. fusion and as well as Obtaining encoded features The encoded features are decoded to obtain the decoded features. The final predicted trajectory coordinate information C is then output.
2. The vehicle trajectory prediction method based on drivable area feature extraction according to claim 1, characterized in that, In step S1, the temporal features of the vehicle to be predicted are extracted. The specific process includes: Historical motion status information includes position coordinate information. Speed information Acceleration information and heading angle information .
3. The vehicle trajectory prediction method based on drivable area feature extraction according to claim 1, characterized in that, In step S1, the temporal features of the surrounding vehicles of the vehicle to be predicted are extracted. The specific process includes the following steps: Step B1: Based on the current position coordinates of the vehicle to be predicted, obtain the historical motion state information of the nearest N surrounding vehicles at T time points. The historical motion state information includes position coordinate information. Speed information Acceleration information and heading angle information ; Step B2: Use a one-dimensional convolutional neural network or a fully connected network to encode the T historical motion state information of the surrounding vehicles to obtain the high-dimensional features corresponding to the historical motion state information of the surrounding vehicles. Step B3: Input the high-dimensional features corresponding to the historical motion state information of surrounding vehicles into the serialization prediction model to extract the temporal features of N surrounding vehicles. .
4. The vehicle trajectory prediction method based on drivable area feature extraction according to claim 1, characterized in that, The specific process of step S2 includes the following steps: Step C1: Based on the current location coordinates and driving direction of the vehicle to be predicted, obtain a high-precision map of the current location; Step C2: Using the current motion state of the vehicle to be predicted, select a map area within a specified range in the high-precision map, and construct the area into a single-channel map area with a width of W and a height of H. Set the pixels of the drivable area to 1 and the pixels of the non-drivable area to 0. Step C3: Extract the feature vector of the drivable region using a convolutional neural network (CNN) and pooling layers. ; Step C4: Extract the feature vector of the drivable area Temporal characteristics of the vehicle to be predicted By performing fusion, fusion characteristics are obtained. ; Step C5: The fusion features Input the fully connected layer to generate the predicted endpoint coordinates. .
5. A vehicle trajectory prediction method based on drivable area feature extraction according to claim 1 or 4, characterized in that, The specific process of step S3 includes the following steps: Step D1: Use a fully connected layer to predict the endpoint coordinates Feature encoding is performed to obtain a representation vector; Step D2: Convert the representation vector Temporal characteristics of the vehicle to be predicted and interactive features By fusing the coded features, the encoded features are obtained. ; Step D3: Encode features The input is a decoder, which uses a serialized feature model for decoding, and the output is decoded features. ; Step D4: Decode the features The input is a fully connected layer, and the output is the final predicted trajectory coordinate information C, which contains K predicted trajectory coordinate points. .
6. The vehicle trajectory prediction method based on drivable area feature extraction according to claim 1, characterized in that, Also includes: The mean squared error (MSE) loss function is used to supervise and iterate the predicted trajectory C and the true trajectory G, where G represents the true future trajectory coordinate information, containing K true trajectory coordinate points. .
7. The vehicle trajectory prediction method based on drivable area feature extraction according to claim 1, characterized in that, Also includes: Step E1: Add a consistency loss function For the predicted endpoint coordinates and the final predicted trajectory endpoint coordinates By comparing the results, the predicted endpoint coordinates can be obtained. The final coordinates of the predicted trajectory The mean square error; Step E2: Through weights and Combination loss function and loss function The total loss function is obtained. ; Step E3: Iterate and optimize using the loss function L.
8. The vehicle trajectory prediction method based on drivable area feature extraction according to claim 7, characterized in that, The loss function It is expressed as follows: 。 9. A vehicle trajectory prediction method based on drivable area feature extraction according to claim 7, characterized in that, The loss function It is expressed as follows: 。 10. A vehicle trajectory prediction method based on drivable area feature extraction according to claim 7, characterized in that, The loss function It is expressed as follows: 。