Vehicle Sensor Fusion Target Prediction During Measurement Gaps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current sensor fusion systems for vehicles face errors in predicting the position of a sensor fusion target, especially on curved roads, due to the lack of current measurement values and inadequate consideration of road curvature, leading to increased position errors.
Innovation Solution
A sensor fusion target prediction device that includes a learning unit using a recurrent neural network (RNN) to accumulate and learn sensor fusion target information over time, allowing for the calculation of prediction values based on learned parameters, and a target tracking unit that uses these predictions when current measurement values are not available.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If sensor fusion systems use traditional prediction methods without current measurement values, then the system can operate during sensor outages, but position prediction errors increase significantly
Solution Approach 1:
The system performs preliminary learning of target motion patterns during normal operation when measurement values are available. The learning unit stores learned parameters characterizing target behavior in advance, so that when sensor outages occur and current measurement values are unavailable, the prediction unit can rely on these pre-acquired parameters to maintain prediction accuracy without significant error increase.
Solution Approach 2:
The system continuously updates and refines the learned parameters based on incoming sensor fusion target information during normal operation. This feedback mechanism ensures that the prediction model adapts to actual target behavior patterns, improving prediction accuracy during outages. The learning unit receives feedback from measurement values and adjusts parameters accordingly, creating a closed-loop system that maintains reliability and precision.
2Device complexity
If sensor fusion systems do not consider road curvature in prediction, then the system complexity remains low, but position determination errors increase on curved roads
Solution Approach 1:
The learning unit dynamically adjusts prediction parameters based on road curvature characteristics. When the target is on a curved road, the system learns and adapts parameters that account for the curvature effect on target motion. This parameter adaptation allows the prediction unit to compensate for road curvature without requiring complex geometric calculations, maintaining position determination accuracy while avoiding excessive system complexity.
3Speed
If the system uses simple prediction without learning past transitions, then the calculation speed is fast, but prediction reliability decreases when no current measurement is available
Solution Approach 1:
The learning unit performs preliminary analysis of past target transitions and stores learned parameters in advance. During prediction, especially when no current measurement is available, the prediction unit quickly retrieves and applies these pre-learned parameters rather than performing complex real-time analysis. This preliminary action approach maintains fast calculation speed while significantly improving prediction reliability during sensor outages.
Data Source
AI summary
A sensor fusion target prediction device and method for vehicles that can estimate a prediction value in the state in which no current measurement value is present, and a vehicle including the device are disclosed. The sensor fusion target prediction device may include a learning unit for receiving sensor fusion target information and learning one or more parameters based on the received sensor fusion target information, a prediction unit for, upon receiving current sensor fusion target information, calculating a prediction value of the current sensor fusion target information based on the one or more parameters learned by the learning unit, and a target tracking unit for determining whether the sensor fusion target information is received and tracking a target using the prediction value calculated by the prediction unit based on not receiving the sensor fusion target information.