Structural RNN for Vehicle Lane Change Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in predicting lane changes by nearby vehicles when sensor data is unreliable, leading to difficulties in path planning and obstacle avoidance.
Innovation Solution
A structural recurrent neural network (S-RNN) based on a factor graph is used to predict lane changes by analyzing pose information of nearby vehicles, leveraging past sensor data to generate prediction indicators for future movements up to three seconds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If sensors are used to perceive surrounding vehicles and environment, then the vehicle can make decisions based on real-time data, but the system reliability deteriorates when sensors fail in real-world conditions
Solution Approach 1:
The system performs preliminary actions by training the S-RNN model offline with historical sensor data to learn patterns of vehicle behavior. This pre-trained model can then make predictions even when current sensor data is unavailable or unreliable, allowing the system to anticipate lane changes and make decisions without real-time sensor input.
Solution Approach 2:
The system creates a virtual copy of the physical sensor environment by using the S-RNN to generate predicted states of surrounding vehicles. This virtual model replicates the behavior of actual vehicles based on learned patterns, allowing the autonomous vehicle to interact with this virtual representation when physical sensors fail, thereby maintaining decision-making capability.
2Reliability
If the vehicle system uses knowledge of surrounding vehicles for path planning, then maneuver safety is improved, but the system cannot effectively plan when sensor data is unavailable
Solution Approach 1:
The S-RNN model incorporates feedback mechanisms by continuously updating its predictions based on the difference between predicted and actual vehicle states. The model uses historical sensor data to train and refine its predictions, creating a feedback loop that improves accuracy over time and allows the system to maintain safe path planning even when current sensor data is limited.
Solution Approach 2:
The system performs preliminary path planning by using the trained S-RNN model to predict future states of surrounding vehicles before actual sensor data becomes available or reliable. This allows the autonomous vehicle to pre-calculate safe maneuvers and plan paths in advance, maintaining maneuver safety even during periods of sensor unavailability.
3Measurement precision
If a structural recurrent neural network is used to predict lane changes, then prediction accuracy is improved, but the computational complexity and system structure become more complex
Solution Approach 1:
The S-RNN model segments the prediction task by processing different aspects of vehicle behavior separately - it divides the input data into distinct features (positions, velocities, lane indicators) and processes them through specialized neural network components. This segmentation allows the complex prediction task to be broken down into manageable parts, improving accuracy while making the system more tractable.
Solution Approach 2:
The factor graph serves as an intermediary structure that bridges the complex neural network computations and the practical prediction requirements. It provides a structured representation of the relationships between different vehicle attributes and prediction outcomes, simplifying the interface between the complex S-RNN model and the rest of the autonomous vehicle system.
Data Source
AI summary
System, methods, and other embodiments described herein relate to predicting lane changes for nearby vehicles of a host vehicle. In one embodiment, a method includes, in response to detecting that one or more of the nearby vehicles are present proximate to the host vehicle, collecting pose information about the nearby vehicles. The nearby vehicles are traveling proximate to the host vehicle and in a direction of the host vehicle. The method includes analyzing the pose information of the nearby vehicles using separate recurrent units of a structural recurrent neural network (S-RNN) to generate factors according to the lanes. The method includes generating prediction indicators for the nearby vehicles as a function of the factors for the lanes using the S-RNN. The method includes providing electronic outputs identifying the prediction indicators that specify a likelihood of the nearby vehicles changing between the lanes.