Vehicle Trajectory Prediction With Attention-Based Interaction Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing trajectory prediction methods for vehicles in the surroundings of an ego vehicle suffer from high computational complexity and inefficiency due to the inclusion of all vehicles in the scene, leading to inadequate prediction results and increased training time.
Innovation Solution
A two-step method for trajectory prediction that uses an attention-based interaction algorithm to preselect relevant vehicles based on their interaction degrees, reducing the number of vehicles considered for prediction and thereby simplifying the prediction process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all vehicles in the scene are included in trajectory prediction, then comprehensive prediction coverage is achieved, but computational complexity and runtime increase significantly
Solution Approach 1:
The patent extracts and removes irrelevant vehicles from the complete scene graph, keeping only those vehicles that have meaningful interactions with the ego vehicle or other predicted vehicles. This is achieved through filtering mechanisms that evaluate vehicle relevance based on spatial relationships, motion patterns, and interaction potentials, thereby reducing computational complexity while preserving prediction coverage for important vehicles.
Solution Approach 2:
The patent segments the complete set of vehicles into relevant and irrelevant subsets. By dividing the scene into meaningful groups (e.g., vehicles in close proximity, vehicles with similar motion patterns, vehicles on intersecting paths), the system can apply trajectory prediction algorithms selectively to relevant segments, reducing overall computational burden while maintaining comprehensive coverage where needed.
2Reliability
If all vehicles are considered for trajectory prediction, then complete interaction analysis is performed, but training time increases excessively
Solution Approach 1:
The patent performs preliminary filtering of vehicles before the main trajectory prediction training process. By pre-identifying and selecting only those vehicles that will meaningfully contribute to the prediction outcomes, the system avoids wasting training computational resources on irrelevant vehicles, thereby significantly reducing training time while preserving complete interaction analysis for the relevant subset.
3Device complexity
If distance-based preselection of vehicles is used, then computational load is reduced, but relevant vehicles may be excluded based solely on distance criteria
Solution Approach 1:
The patent changes the selection parameters from purely distance-based criteria to multi-parameter criteria that include spatial relationships, velocity vectors, acceleration patterns, and predicted interaction potentials. This allows the system to identify relevant vehicles based on their actual impact on trajectory prediction needs, rather than simply their proximity, thereby improving selection accuracy while maintaining reduced computational load through selective filtering.
Data Source
AI summary
A method for trajectory prediction of vehicles in the surroundings of an ego vehicle involves determining degrees of interaction between the vehicles using an attention-based interaction algorithm trained by machine. Individual vehicles from the multitude of the vehicles located in the surroundings of the ego vehicle are identified as relevant to the trajectory prediction by the interaction algorithm and selected for this when their respective degree of interaction with at least one of the vehicles whose trajectory is to be predicted exceeds a predetermined threshold value. In a subsequent learning step, a trajectory prediction algorithm is trained with the vehicles selected as relevant to the trajectory prediction, and the trajectory prediction carried out by the trajectory prediction algorithm is carried out for the vehicles selected as relevant to the trajectory prediction.