Vectorized Trajectory Prediction for Real-Time Autonomous Driving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing trajectory prediction systems for autonomous vehicles require computationally intensive and memory-heavy neural networks due to the use of rasterized image representations, which are not feasible for real-time autonomous driving applications.
Innovation Solution
Utilizing vectorized representations of the scene data, approximated by polylines, to generate trajectory predictions through an encoder and decoder neural network architecture, reducing the number of model parameters and floating-point operations (FLOPs) while maintaining performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If rasterized image representations are used for trajectory prediction, then the system can process visual scene data, but the computational complexity and memory requirements become excessively high
Solution Approach 1:
The patent changes the representation format parameter from rasterized images to vectorized polylines. This parameter change fundamentally alters the data structure from a grid-based pixel array to a coordinate-based geometric representation, reducing the dimensionality and computational complexity while preserving the essential spatial information needed for accurate trajectory prediction.
Solution Approach 2:
The patent extracts only the essential geometric features (polylines representing roads, curbs, and trajectories) from the complete scene data, discarding the redundant pixel information present in rasterized images. This extraction process retains the critical spatial relationships needed for trajectory prediction while eliminating the computational burden of processing entire image matrices.
2Reliability
If rasterized image representations are used, then comprehensive scene information is captured, but the processing time becomes too long for real-time applications
Solution Approach 1:
The patent changes the temporal parameter of data representation from static rasterized frames to dynamic vectorized polylines that inherently encode motion and spatial relationships. This parameter change enables the system to process only the essential geometric elements rather than entire image sequences, dramatically reducing processing time while maintaining reliable trajectory predictions through the preserved spatial coherence of polyline representations.
3Measurement precision
If high-resolution rasterized images are used for accurate predictions, then measurement precision improves, but memory consumption increases
Solution Approach 1:
The patent changes the data density parameter by transitioning from high-resolution rasterized images with millions of pixels to compact vectorized polyline representations with thousands of coordinates. This parameter change maintains the precision of spatial representation by using exact coordinate geometry rather than approximated pixel grids, while dramatically reducing memory consumption by storing only the essential geometric features.
Solution Approach 2:
The patent extracts and retains only the essential geometric boundaries and trajectories as polylines, removing all redundant pixel data from rasterized images. This extraction achieves accurate trajectory predictions by focusing computational resources on the critical spatial features (road edges, curbs, paths) while eliminating the memory burden of storing and processing complete high-resolution image data.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for agent trajectory prediction using vectorized inputs.