Obstacle Trajectory Prediction Using Semantic Maps and LSTM
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current trajectory prediction methods for autonomous vehicles, relying on rule-based models, are limited in accuracy and efficiency, especially when insufficient scenario information is available, leading to suboptimal predictions.
Innovation Solution
The use of neural networks, specifically a combination of recurrent neural networks (RNNs) for dynamic feature extraction and convolutional neural networks (CNNs) for semantic map features, iteratively predicts the trajectory of obstacles by learning from past motion and environmental data, enhancing prediction accuracy and robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If rule-based methods are used for trajectory prediction, then the system is simple to implement, but prediction accuracy deteriorates when scenario information is insufficient
Solution Approach 1:
The patent replaces rule-based mechanical prediction systems with neural network-based learning systems. The neural networks learn complex driving patterns and trajectory predictions from data, substituting the rigid rule-based approach with a flexible, data-driven model that achieves higher accuracy while maintaining computational efficiency through optimized network architectures.
2Measurement precision
If neural networks are used for trajectory prediction, then prediction accuracy improves, but computational complexity increases
Solution Approach 1:
The patent segments the trajectory prediction task into multiple specialized neural networks, each handling specific aspects such as behavior prediction, trajectory generation, and semantic map processing. This segmentation allows each network to be optimized for its specific function, reducing overall computational complexity while maintaining high prediction accuracy through specialized processing.
Solution Approach 2:
The patent implements dynamic feature extraction using recurrent neural networks (LSTMs) that adaptively process temporal sequences of obstacle behavior. The system dynamically adjusts its processing based on the current state and historical data, allowing efficient computation that scales with the complexity of the driving scenario rather than using fixed, overly complex architectures.
3Reliability
If more scenario information is collected to improve rule-based prediction, then prediction reliability improves, but data processing time increases
Solution Approach 1:
The patent performs preliminary processing of scenario information by pre-training neural networks on extensive driving data and pre-computing semantic map features. During real-time operation, the pre-trained models quickly process incoming sensor data without requiring extensive computation, thus maintaining high reliability while minimizing data processing time through advance preparation.
Solution Approach 2:
The patent introduces semantic map features as an intermediary representation that bridges raw sensor data and trajectory predictions. The semantic map pre-processes environmental information into meaningful features (road boundaries, traffic signs, pedestrian zones) that the neural networks can efficiently utilize, reducing the computational burden of processing raw sensor data while improving prediction reliability.
Data Source
AI summary
A system and method for predicting the near-term trajectory of a moving obstacle sensed by an autonomous driving vehicle (ADV) is disclosed. The method applies neural networks such as a LSTM model to learn dynamic features of the moving obstacle's motion based on its past trajectory up to its current position and a CNN model to learn the semantic map features of the driving environment in a portion of an image map. From the learned dynamic features of the moving obstacle and the learned semantic map features of the environment, the method applies a neural network to iteratively predict the moving obstacle's positions for successive time points of a prediction interval. To predict the moving obstacle's position at the next time point from the currently predicted position, the methods may update the learned dynamic features of the moving obstacle based on its past trajectory up to the currently predicted position.