Road User Movement Prediction with Latent State Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automated vehicles face challenges in predicting the movements of human-controlled and non-controlled road users, such as pedestrians and cyclists, due to uncertainty in their behavior, which can lead to potential collisions and unnecessary evasive maneuvers.
Innovation Solution
A method using a trained encoder network to map surroundings observations to reduced-dimensional representations, combined with prediction networks to forecast the future actions and states of traffic-relevant objects, enhancing prediction accuracy and robustness against noise, and utilizing kinematic models for further processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional prediction methods are used to forecast movement of road users, then the system is simpler to implement, but the prediction accuracy is insufficient leading to potential collisions
Solution Approach 1:
The patent introduces an encoder network as an intermediary component that transforms raw sensor observations into compressed latent representations. This encoder acts as a mediator between the observation system and prediction networks, enabling more accurate movement predictions by capturing essential patterns while filtering noise, thus resolving the contradiction between prediction accuracy and system complexity
Solution Approach 2:
The prediction system is segmented into multiple specialized networks: an encoder network for feature extraction, and multiple prediction networks for different prediction tasks. This segmentation allows each component to specialize in specific functions, improving overall prediction accuracy while maintaining manageable system complexity through modular architecture
2Reliability
If observations are processed with high dimensionality to preserve all information, then measurement precision is maintained, but noise and disturbances affect prediction reliability
Solution Approach 1:
The encoder network extracts only the most relevant features from high-dimensional observations by mapping them to a lower-dimensional latent space. This extraction process separates essential information from noise and disturbances, improving robustness against noise while preserving measurement precision through selective feature extraction
Solution Approach 2:
The patent changes the parameter representation from high-dimensional raw observations to lower-dimensional latent representations. This parameter transformation maintains the essential information needed for accurate prediction while reducing the impact of noise and disturbances through dimensionality reduction
3Reliability
If the system makes conservative predictions to avoid collisions, then safety is improved, but unnecessary evasive maneuvers increase
Solution Approach 1:
The system uses multiple prediction networks that provide probabilistic forecasts of future positions and actions. This feedback mechanism allows the control system to assess collision risk more accurately and distinguish between genuine threats requiring evasive maneuvers and normal variations in road user behavior, thus improving collision avoidance while reducing unnecessary maneuvers
Data Source
AI summary
A method is for predicting movement of at least one traffic-related object based on observations of the surroundings of the object. The method includes mapping an observation of the surroundings at a first time by a trained encoder network to a representation with reduced dimensionality. The method also includes, based on a first action performed by the object at the first time and the representation, using at least one trained prediction network to determine a first representation prediction of the representation to which a future observation is likely to be mapped by the trained encoder network at a first future time, and/or determine a first action prediction of a second action that the object is likely to perform at the first future time. The method also includes determining a first dynamic state prediction for a dynamic state of the object at the first future time.