Lane-Aware Trajectory Prediction for Diverse Driving Paths
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing trajectory prediction methods for autonomous vehicles fail to fully utilize map data, relying on it as a soft constraint, which limits their effectiveness in predicting diverse and accurate vehicle trajectories.
Innovation Solution
A computer-implemented method for driving path prediction that concatenates past trajectory features and lane centerline features, uses a convolutional encoder to derive scene representations, and employs a hypercolumn trajectory encoder and decoder to generate primary and auxiliary trajectory predictions, along with an Inverse Optimal Control-based ranking module to refine these predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If map data is used as a soft constraint in trajectory prediction, then the prediction method is simple to implement, but the prediction accuracy and diversity are limited
Solution Approach 1:
The patent segments the trajectory prediction problem into multiple independent trajectory hypotheses that are evaluated separately. Instead of treating map data as a single soft constraint, the system divides the prediction space into multiple plausible trajectories, each evaluated against the map data and other constraints independently, then combines them to form the final prediction distribution.
Solution Approach 2:
The patent introduces a new dimensional approach by incorporating lane-level semantic information and map data as hard constraints in the prediction framework. This adds a semantic dimension to the prediction space, transforming the problem from purely spatial-temporal prediction to include semantic consistency with road infrastructure.
2Device complexity
If map data is used as a soft constraint, then the system complexity remains low, but the ability to predict diverse trajectories is insufficient
Solution Approach 1:
The patent implements a dynamic prediction framework where multiple trajectory hypotheses are generated and evaluated. The system dynamically adjusts the prediction distribution based on the evaluation of each hypothesis against map constraints, vehicle dynamics, and observed agent behavior, allowing diverse trajectories to emerge naturally from the competitive evaluation process.
Solution Approach 2:
The patent introduces an intermediary evaluation module that assesses each trajectory hypothesis against multiple criteria including map consistency, vehicle dynamics, and behavioral patterns. This intermediary layer mediates between the simple generation of hypotheses and the complex requirements for diverse, accurate predictions, enabling versatility without proportionally increasing system complexity.
3Use of energy by moving object
If traditional trajectory prediction methods are used, then computational resources are saved, but prediction accuracy and reliability are compromised
Solution Approach 1:
The patent performs preliminary filtering and evaluation of trajectory hypotheses before final selection. By pre-evaluating hypotheses against map constraints and vehicle dynamics early in the process, the system eliminates implausible trajectories before more computationally intensive processing, maintaining reliability while managing computational resources efficiently.
Solution Approach 2:
The patent replaces traditional mechanical optimization approaches with a learning-based evaluation framework. Instead of using complex mechanical models to directly compute optimal trajectories, the system uses trained neural networks to evaluate hypotheses, substituting iterative mechanical optimization with more efficient learned predictions that maintain reliability.
Data Source
AI summary
A method for driving path prediction is provided. The method concatenates past trajectory features and lane centerline features in a channel dimension at an agent's respective location in a top view map to obtain concatenated features thereat. The method obtains convolutional features derived from the top view map, the concatenated features, and a single representation of the training scene the vehicle and agent interactions. The method extracts hypercolumn descriptor vectors which include the convolutional features from the agent's respective location in the top view map. The method obtains primary and auxiliary trajectory predictions from the hypercolumn descriptor vectors. The method generates a respective score for each of the primary and auxiliary trajectory predictions. The method trains a vehicle trajectory prediction neural network using a reconstruction loss, a regularization loss objective, and an IOC loss objective responsive to the respective score for each of the primary and auxiliary trajectory predictions.