Traffic Trajectory Prediction With Class-Specific Interaction Graphs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing trajectory prediction methods for intelligent vehicles in complex heterogeneous environments, such as those with pedestrians, bicycles, and vehicles, fail to accurately account for the dynamic characteristics, social attributes, and information processing differences among various traffic participants, leading to poor prediction accuracy.
Innovation Solution
A method involving the creation of position, velocity, acceleration, and class graphs, using expert experience correction coefficients and multi-mode decoding with graph convolutional and temporal neural networks to capture topological and time-dependent relationships, allowing for distinct processing of environmental information by different traffic participants.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing trajectory prediction methods based on LSTM or graph convolutional neural networks are used, then social interaction influence on future trajectory can be studied, but prediction accuracy deteriorates in complex heterogeneous environments composed of pedestrians, bicycles, and vehicles
Solution Approach 1:
The patent segments traffic participants into different classes (pedestrians, bicycles, vehicles) and creates separate prediction models for each class. Each class has its own set of parameters including attention weights, motion capability coefficients, and social constraint factors. This segmentation allows the system to account for the unique characteristics of each participant type, resolving the contradiction between general applicability and class-specific accuracy.
Solution Approach 2:
The patent applies local quality by assigning different parameter values and model configurations to different traffic participant classes. Each class receives customized attention weights, motion capability coefficients, and social constraint factors based on its specific characteristics. For example, pedestrians receive different parameters than vehicles, allowing the system to optimize prediction accuracy for each local class while maintaining overall system functionality.
2Device complexity
If a unified trajectory prediction method is used for all traffic participants, then device complexity is reduced, but the method cannot account for differences in dynamic characteristics, social attributes, and information processing between different traffic participants
Solution Approach 1:
The patent implements universality by designing a modular prediction framework that can handle multiple traffic participant classes using a consistent architectural structure. The same neural network architecture and processing pipeline are used across all classes, but with class-specific parameter configurations. This allows the system to maintain relatively low complexity while achieving high reliability for different traffic classes through parameter customization rather than structural complexity.
Data Source
AI summary
Disclosed is a method for predicting a trajectory of a traffic participant in a complex heterogeneous environment, including the following steps: obtaining traffic participant information in a complex heterogeneous environment; arranging and numbering traffic participant classes based on the class information, to obtain serial numbers of the traffic participant classes; establishing a position graph, a velocity graph, an acceleration graph, and a class graph, into each of which expert experience is introduced; and capturing topological structure relationships and time dependence relationships to obtain a position hidden state, a velocity hidden state, an acceleration hidden state, and a class hidden state; classifying the position hidden state, the velocity hidden state, the acceleration hidden state, and the class hidden state to obtain a hidden state set of traffic participants; and decoding hidden states of the traffic participants separately using a corresponding decoder to obtain future trajectory predictions of the traffic participants.