Vehicle Trajectory Prediction Using Intent-Aware Relationship Graphs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current trajectory prediction methods for autonomous vehicles fail to adequately capture the interactive features between the target vehicle and its surroundings, particularly neglecting the impact of driving intentions on future vehicle trajectories.
Innovation Solution
The Cro-IntentFormer-based method integrates driving intentions into the prediction of surrounding vehicle trajectories by constructing physical and semantic relationship graphs, which consider both the distance between vehicles and the similarity of their behavioral intentions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If current trajectory prediction methods only consider distance between vehicles, then the model complexity is low, but the prediction accuracy and ability to capture interactive features deteriorates
Solution Approach 1:
The patent segments the trajectory prediction task into two distinct modules: an intent prediction module that predicts driving intentions (going straight, turning left, turning right) and a trajectory prediction module that generates future trajectories. This segmentation allows the system to separately model behavioral intentions and spatial movements, improving prediction accuracy by capturing the semantic meaning behind trajectory patterns without excessively increasing overall model complexity.
Solution Approach 2:
The patent adds a semantic dimension to traditional trajectory prediction by incorporating intent prediction. Instead of only considering spatial coordinates and distances, the model now operates in an enhanced feature space that includes behavioral intentions. This dimensional expansion allows the model to capture interactive features between vehicles based on their intended actions, significantly improving prediction accuracy while maintaining manageable complexity through modular architecture.
2Reliability
If driving intention prediction is separated from trajectory prediction, then the model structure is simple, but the integration of behavioral factors into trajectory prediction deteriorates
Solution Approach 1:
The patent merges intent prediction and trajectory prediction into a unified Cro-IntentFormer model architecture. The model shares common components including spatio-temporal feature extraction modules, attention mechanisms, and positional encoding. This merging ensures that intent information and trajectory information are processed together, allowing behavioral factors to directly influence trajectory predictions. The shared feature extraction and attention mechanisms enable seamless integration while maintaining model efficiency through parameter sharing.
Solution Approach 2:
The Cro-IntentFormer model is designed with multi-functionality, serving both intent prediction and trajectory prediction tasks within a single unified framework. The spatio-temporal feature extraction module and attention mechanisms are universally applied to both prediction tasks, allowing the model to leverage common patterns in both behavioral intentions and spatial movements. This universal design improves reliability by ensuring consistent processing of related features while avoiding the need for separate specialized models.
3Loss of information
If attention mechanisms are applied to individual time frames, then the computational load is low, but the connections between adjacent trajectory segments are overlooked
Solution Approach 1:
The patent segments the trajectory time series into multiple time frames and applies self-attention mechanisms within each frame while using cross-attention to capture connections between adjacent frames. This segmentation approach allows the model to process computational load in manageable chunks while still establishing temporal relationships through the cross-attention mechanism. The segmented architecture prevents the need to compute attention across the entire trajectory at once, managing computational energy efficiently while preserving segment connections.
Solution Approach 2:
The patent ensures continuity of useful action by implementing attention mechanisms that explicitly model temporal dependencies between adjacent trajectory segments. The cross-attention mechanism allows each time frame to attend to relevant features in adjacent frames, maintaining continuous information flow throughout the trajectory. This continuous attention mechanism prevents information loss at segment boundaries while distributing computational load across multiple processing stages, balancing information preservation with energy efficiency.
Data Source
AI summary
The provided is a method and system for trajectory prediction based on Cro-IntentFormer. The method starts by preprocessing vehicle trajectory data collected by sensors to produce raw data suitable for model input. Vehicles are treated as nodes, and the distance between vehicles serves as the basis for determining whether there is an edge between two vehicle nodes. A physical relationship graph is constructed and, along with the raw data, input into a spatio-temporal feature extraction module to obtain the spatio-temporal features of the trajectory. The spatio-temporal feature matrix is then input into an intent prediction module to determine the predicted intentions of the vehicles. Based on the intent information output by the intent prediction module, a semantic relationship graph is reconstructed and input, along with the raw data, into the spatio-temporal feature extraction module to derive the semantic features of the trajectory.