Vehicle Trajectory Prediction Using Vectorized Interaction Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for predicting vehicle trajectories in autonomous driving struggle to fully utilize interaction between the surrounding environment and the vehicle, due to the limited receptive field of convolutional neural networks, which affects the accuracy of multimodal trajectory predictions.
Innovation Solution
A method that obtains vectorized features from perceived information of autonomous vehicles, including self-vehicle, surrounding vehicles, and road network information, and performs feature interaction and fusion to generate accurate trajectory predictions, using an encoder-decoder network model with attention submodules to integrate rich vehicle and environment information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If a convolutional neural network is used for trajectory prediction, then the processing of image features is simplified, but the receptive field is limited and cannot fully utilize interaction between surrounding environment and self-vehicle
Solution Approach 1:
The patent segments the feature processing into two distinct parts: (1) using a convolutional neural network to process image features from the surrounding environment, and (2) using a graph neural network to process vectorized features representing vehicles and their interactions. This segmentation allows each network type to operate on its most suitable data format, with the CNN handling spatial image data and the GNN handling relational vehicle interaction data, thereby resolving the contradiction between processing simplicity and information completeness.
Solution Approach 2:
The patent merges the outputs of two different neural network architectures (CNN and GNN) into a unified trajectory prediction result. The CNN processes image features to extract environmental context, while the GNN processes vectorized vehicle features to model interactions; their combined results provide both processing efficiency and comprehensive environmental interaction information, resolving the contradiction between operational ease and information loss.
2Device complexity
If only image features from convolutional neural network are used, then the computational model remains simple, but the interaction between surrounding environment and vehicles cannot be fully considered
Solution Approach 1:
The computational model is segmented into two parallel processing streams: a CNN stream for image feature extraction and a GNN stream for vehicle interaction modeling. Each stream maintains its own complexity level appropriate for its task, while their integration achieves high prediction reliability without requiring one stream to承担 the full complexity burden alone.
Solution Approach 2:
The patent introduces vectorized features as an intermediary representation that bridges the image features from CNN and the interaction modeling requirements of GNN. These vectorized features encode vehicle states, positions, and relationships in a format suitable for graph neural network processing, enabling reliable trajectory prediction through multi-modal feature fusion without excessive computational complexity.
3Power
If the receptive field is limited, then the computational load is reduced, but the ability to utilize surrounding environment interaction is insufficient
Solution Approach 1:
The patent transitions from a single-dimension image processing approach to a multi-dimensional framework by introducing vectorized features that represent vehicles in spatial, temporal, and relational dimensions. The GNN operates on this multi-dimensional vectorized representation to capture environmental interactions that cannot be expressed in traditional image space, thereby reducing computational load on the CNN while preserving rich interaction information through the additional vectorized feature dimension.
Data Source
Figure 1~2
Figure 3~5
Figure 6
AI summary
The disclosure relates to the technical field of autonomous driving, and specifically provides a method for predicting a vehicle trajectory, a control device, a readable storage medium, and a vehicle, with a view to solve the problem of effectively predicting a multimodal trajectory of a vehicle. According to the disclosure, perceived information of an autonomous vehicle is converted into vectorized features of target vehicles, where the target vehicles include the autonomous vehicle and a plurality of first surrounding vehicles; and trajectory prediction results of the target vehicles are obtained based on the vectorized features. The vectorized features in the disclosure are obtained based on the perceived information of the autonomous vehicle, which enables the vectorized features to include rich vehicle trajectory information and environment information of the target vehicle; and interaction and fusion are performed on the vectorized features of the target vehicle, so that a plurality of vehicle trajectories of each target vehicle can be effectively predicted.