Attention Mechanism for Road User Trajectory Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for predicting the future trajectories of road users in autonomous driving systems face challenges in modeling interactions between road users at far distances without increasing computational effort or losing finer details, due to the limited receptive field of traditional convolutional neural networks (CNNs).
Innovation Solution
A computer-implemented method using a machine learning algorithm with an attention mechanism, including set attention blocks and multi-head attention, encodes trajectory characteristics and static environment data to predict road user trajectories without rasterization, allowing for interactions to be modeled across various distances and scales, reducing computational requirements and maintaining detail.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Area of stationary object
If the receptive field of CNN is increased by stacking multiple CNN-blocks or increasing kernel size, then the spatial range for learning interactions is improved, but computational cost increases and finer details are lost
Solution Approach 1:
The patent introduces an attention mechanism as an intermediary component that selectively focuses on relevant road users and interactions. Instead of processing all possible interactions within an expanded receptive field, the attention mechanism computes attention scores to identify and prioritize important interactions, thereby achieving comprehensive interaction modeling without the computational burden of processing the entire expanded field.
Solution Approach 2:
The attention mechanism applies local quality by assigning different weights to different interactions based on their importance. Rather than uniformly processing all interactions within the receptive field, the system focuses computational resources on locally important interactions (those with high attention scores) while reducing or eliminating processing of less important interactions, thus maintaining detail while reducing computational cost.
2Length of stationary object
If the receptive field of CNN is increased, then interactions at far distances can be learned, but finer details in interactions are lost
Solution Approach 1:
The attention mechanism serves as an intermediary that bridges the gap between global spatial coverage and local detail preservation. It enables the system to consider interactions across the entire spatial range while maintaining detail by selectively focusing attention on specific important interactions rather than averaging or diluting information across the entire field.
Solution Approach 2:
By assigning different attention weights to different spatial locations and interactions, the system preserves local details in important regions while still considering the global spatial context. This allows fine details in critical interactions to be maintained even as the spatial range expands to include far-distance interactions.
3Reliability
If traditional CNN approaches are used to model road user interactions, then the model can capture spatial correlations, but it cannot effectively model complex interactions at various distances and time horizons
Solution Approach 1:
The patent applies dynamics by making the interaction modeling adaptive rather than static. The attention mechanism dynamically adjusts which interactions to focus on based on the current traffic scene, allowing the model to adapt to varying distances and time horizons. This enables reliable interaction modeling across diverse scenarios without requiring separate models for different distance ranges.
Solution Approach 2:
The attention mechanism provides universality by creating a single unified framework that can handle interactions at all distances and time horizons. Rather than requiring separate CNN blocks or mechanisms for near-field and far-field interactions, the attention-based approach universally models all interactions within a single framework, adapting to the specific needs of each situation.
Data Source
AI summary
A method is provided for predicting respective trajectories of a plurality of road users. Trajectory characteristics of the road users are determined with respect to a host vehicle via a perception system, wherein the trajectory characteristics are provided as a joint vector describing respective dynamics of each of the road users for a predefined number of time steps. The joint vector of the trajectory characteristics is encoded via an algorithm which included an attention algorithm for modelling interactions of the road users. The encoded trajectory characteristics and encoded static environment data obtained for the host vehicle are fused in order to provide fused encoded features. The fused encoded features are decoded in order to predict the respective trajectory of each of the road users for a predetermined number of future time steps.