Road User Trajectory Prediction With Class-Specific Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for predicting the trajectories of road users in autonomous driving and ADAS systems fail to effectively consider the specific characteristics of different types of road users, leading to reduced prediction accuracy due to the use of common features for all road users, which can negatively impact the reliability of trajectory predictions.
Innovation Solution
A computer-implemented method that determines specific characteristics for each road user, such as pedestrians and vehicles, and transforms them into input data for a prediction algorithm using an embedding algorithm, allowing for the consideration of unique features like body orientation, heading angles, and turning signals, which are then used to predict trajectories by allocating latent features to a dynamic grid map.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If common features are used for all road users, then the data representation is simplified and uniform, but the prediction accuracy deteriorates due to loss of specific characteristics
Solution Approach 1:
The patent segments the feature representation by creating separate embedding layers for different road user classes (pedestrians, cyclists, vehicles). Each class has its own dedicated embedding layer that processes class-specific characteristics, allowing the system to maintain uniform data structure while preserving specific features for each class through segmented processing paths.
Solution Approach 2:
The patent applies local quality by making the embedding layers class-specific rather than uniform. Each road user class has customized embedding parameters and characteristics tailored to its specific features (e.g., pedestrians have body orientation, vehicles have turning signals), while the overall data structure remains consistent. This allows different parts of the system to have different properties optimized for their specific class.
2Measurement precision
If class-specific characteristics are incorporated for each road user, then the prediction accuracy improves, but the device complexity increases due to multiple embedding layers
Solution Approach 1:
The patent achieves universality by designing a unified neural network architecture that handles multiple road user classes through a common framework. The system uses a single prediction algorithm that processes inputs from multiple class-specific embedding layers, allowing one system to serve multiple functions for different road user types without requiring separate prediction models for each class.
3Loss of information
If class-specific embedding layers are used for different road users, then the specific characteristics are preserved, but the processing time increases due to multiple transformation steps
Solution Approach 1:
The patent merges the processing of multiple road user classes into a unified neural network prediction step. After class-specific embedding layers transform their respective inputs, the results are combined and processed together by a single prediction algorithm, reducing the number of separate processing steps and minimizing overall processing time while preserving class-specific information.
Data Source
AI summary
A method is provided for predicting trajectories of a plurality of road users. For each road user, a set of characteristics detected by a perception system of a vehicle is determined, wherein the set of characteristics includes specific characteristics associated with a predefined class of road users. The set of characteristics is transformed to a set of input features for a prediction algorithm via a processing unit of the vehicle, wherein each set of input data comprises the same predefined number of data elements. At least one respective trajectory for each of the road users is determined by applying the prediction algorithm to the input data.