Driving Behavior Simulation With Sensor Error-Calibrated Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional computer models of driving behavior fail to reproduce realistic interactions between vehicles, leading to potential dangerous maneuvers and accidents, especially in complex traffic scenarios like merging lanes or on/off-ramps, as they are based on non-empirical parameters and do not account for measurement data from real vehicle interactions.

Innovation Solution

A system utilizing environmental sensors and position sensors to collect data on observed vehicle behavior, which is then used to create a machine learning-based driving behavior simulator that adjusts model parameters to reduce deviations between observed and simulated vehicle interactions, employing techniques like behavioral cloning, inverse reinforcement learning, and generative adversarial imitation learning to generate realistic driving trajectories.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If conventional computer models use non-empirical parameters, then model simplicity is maintained, but realism of driving behavior reproduction deteriorates

Engineering Contradiction:
Improvemodel simplicityVSAvoidrealism of driving behavior
Core Design Contradiction:
Ease of manufactureVSReliability

Solution Approach 1:

The patent copies real-world driving behavior data from observed vehicles into the simulation model. Positional data from multiple environmental sensors and position sensors is used to create training datasets that replicate actual driving patterns, allowing the model to learn and reproduce realistic interactions without requiring complex physical parameters

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent replaces traditional mechanical/physical parameter-based modeling with a data-driven machine learning approach. Instead of using empirical physical parameters to define driving behavior, the system uses neural networks trained on observed positional data to predict and reproduce realistic driving patterns

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Device complexity

If conventional computer models do not use measurement data from real vehicles, then data collection complexity is reduced, but accuracy of interaction reproduction deteriorates

Engineering Contradiction:
Improvedata collection complexityVSAvoidaccuracy of interaction reproduction
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent segments the data collection system into multiple independent sensor units distributed throughout the traffic area. Each environmental sensor and position sensor independently captures positional data, which are then aggregated to create comprehensive training datasets. This segmentation allows accurate data collection without requiring a single complex centralized system

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent uses universal sensor types (environmental sensors and position sensors) that can collect data from multiple different vehicles and traffic scenarios simultaneously. The same sensor infrastructure serves multiple purposes: tracking individual vehicles, capturing interaction patterns, and generating training data for various driving conditions

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Reliability

If machine learning simulator adjusts model parameters using training data, then realism of driving behavior improves, but computational complexity increases

Engineering Contradiction:
Improverealism of driving behaviorVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent performs preliminary data collection and processing by gathering positional data from multiple sensors before training the machine learning model. Training datasets are prepared in advance with labeled interaction patterns, allowing the neural network to learn from pre-processed data rather than requiring complex real-time computation during simulation

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The machine learning model serves itself by automatically adjusting its parameters through training on the collected data without requiring manual parameter tuning. The system uses the training datasets to self-train the neural networks, reducing the need for expert intervention and simplifying the overall system operation

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP3722907B1Learning a scenario-based distribution of human driving behavior for realistic simulation model and deriving an error model of stationary and mobile sensors
Publication Date: 2023.02.22 BAYERISCHE MOTOREN WERKE AG
  • EP3722907B1 patent drawingFigure 1
  • EP3722907B1 patent drawingFigure 2
  • EP3722907B1 patent drawingFigure 3

AI summary

An system 100 for determining a computer model of realistic driving behavior comprises one or more environmental sensors 110, which are configured to provide sensor data of observed vehicles 140-1 and 140-2 within a traffic area 150. One or more position sensors 160 installed on the observed vehicles 140-1 and 140-2 provide positional data of the observed vehicles 140-1 and 140-2. A processor 120 is configured to extract observed vehicle behavior data from the sensor data and to extract driving behavior training data 310 from observed vehicle behavior data of vehicles interacting with each other within the traffic area 150. The processor determines at least one error model by comparing the positional data with the sensor data of the one or more environmental sensors. The system 100 is further using a machine learning based driving behavior simulator 130, which is configured to adjust driving behavior model parameters 320. The driving behavior model parameters 320 can be determined by reducing a deviation between the driving behavior training data 310 modified by the at least one error model and simulated vehicle behavior data 340, which is generated by the machine learning based driving behavior simulator 130.