Autoregressive Trajectory Prediction With Unified Transformer Patches
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current learning-based agent simulators for autonomous driving systems face high computational complexity and cost due to heterogeneous encoder-decoder architectures that require manual separation of historical and future data, limiting simulation efficiency and scalability.
Innovation Solution
An autoregressive architecture is adopted, treating all time points as present, allowing for more efficient data utilization by predicting movement information in patches without distinguishing between history and future, and utilizing multi-head self- and cross-attention algorithms to integrate information across time and spatial dimensions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If heterogeneous encoder-decoder architecture is used to predict future states, then trajectory prediction capability is improved, but computational complexity and model implementation complexity increase
Solution Approach 1:
The patent merges the encoder and decoder into a single unified transformer model that processes all time points simultaneously. The model uses self-attention mechanisms to handle both historical and future predictions within one architecture, eliminating the need for separate encoder-decoder components and manual data separation, thus reducing implementation complexity while maintaining prediction capability
Solution Approach 2:
The transformer model serves multiple functions: it processes historical data, predicts future states, and handles both position and velocity predictions within a single universal architecture. The model can adapt to different prediction tasks through its attention mechanisms without requiring separate specialized components
2Measurement precision
If manual separation of historical and future data is performed, then trajectory prediction accuracy is improved, but computational burden increases
Solution Approach 1:
The model performs preliminary processing of all time point data simultaneously through parallel attention computations, avoiding the need for sequential processing of historical and future data. The self-attention mechanism computes relationships across all time points in one pass, reducing redundant computations
Solution Approach 2:
The patent changes the processing dimension from sequential (historical then future) to parallel (all time points simultaneously). The transformer architecture processes the entire time series in parallel using attention mechanisms, eliminating the computational overhead of sequential processing and manual data separation
3Measurement precision
If encoder-decoder architecture with different structures is used, then future state prediction is improved, but simulation efficiency decreases
Solution Approach 1:
The patent combines the functionality of separate encoder and decoder structures into a single transformer model that processes all time points uniformly. This merger eliminates the computational overhead of switching between different structural modes and enables more efficient simulation through unified parallel processing
Solution Approach 2:
The model changes the fundamental parameter of architectural structure from heterogeneous (different encoder/decoder structures) to homogeneous (single transformer structure). This parameter change enables more efficient computation while maintaining the ability to predict future states through attention-based relationships
Data Source
AI summary
A trajectory prediction method and an electronic device are provided. The trajectory prediction method includes: during a training phase, obtaining movement information of a plurality of agents at a plurality of time points, and dividing the time points into a plurality of training patches; selecting one of the training patches as a current training patch, obtaining at least one previous training patch preceding the current training patch, and predicting the current training patch according to the previous training patch to update parameters of a machine learning model; changing the current training patch to another training patch, and repeatedly performing the training phase to update the parameters of the machine learning model; and during an inference phase, inputting inference patches into the machine learning model to predict a first patch; and inputting the inference patches and the first patch into the machine learning model to predict a second patch.