Autonomous Driving Scenario Detection Using Real Sensor Frames
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional simulation scenarios for testing autonomous vehicle algorithms fail to accurately represent real-world driving conditions, particularly in regards to human driving behaviors and passenger safety perceptions, leading to inadequate testing of safety and efficacy.
Innovation Solution
A system and method that utilize real-world sensor data frames from autonomous vehicles to train machine learning models to identify unsafe driving conditions, which are then used to refine and update algorithms, ensuring safer decision-making by simulating realistic scenarios that account for human behaviors and passenger comfort.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional simulation scenarios are used for testing autonomous vehicle algorithms, then testing can be performed with standardized procedures, but the scenarios fail to accurately represent real-world driving conditions and human driving behaviors
Solution Approach 1:
The patent creates simulation scenarios by copying and reproducing actual sensor data from real-world driving conditions. Instead of using artificially generated simulation data, the system captures authentic sensor frames from vehicles operating in diverse real-world environments and uses these actual recordings as simulation scenarios, ensuring faithful representation of human driving behaviors and environmental conditions
Solution Approach 2:
The system performs preliminary data collection and labeling of sensor frames from real-world driving conditions before they are needed for testing. By pre-capturing and pre-labeling diverse driving scenarios including edge cases and unsafe conditions in advance, the system prepares a ready-to-use library of authentic simulation scenarios that can be immediately applied for algorithm testing
2Measurement precision
If machine learning models are trained to identify unsafe driving conditions using real sensor data, then detection accuracy improves, but the complexity of data processing and model training increases
Solution Approach 1:
The patent segments the complex task of training machine learning models into distinct, manageable components: (1) collecting and organizing raw sensor data frames, (2) labeling frames as safe or unsafe through automated rules and manual review, (3) creating balanced training datasets with appropriate class distributions, (4) training individual model components, and (5) evaluating performance. This segmentation reduces overall system complexity by making each subtask independently manageable
Solution Approach 2:
The system introduces an intermediary data labeling and preprocessing layer between raw sensor data and machine learning model training. This intermediary component automatically labels sensor frames based on predefined safety criteria and disengagement events, creating structured training data without requiring direct complex interaction between raw data and model training processes
3Productivity
If simulation scenarios are selected based on safety scores from machine learning models, then testing focuses on critical unsafe conditions, but this requires additional computational resources for model inference
Solution Approach 1:
The system applies partial action by using the machine learning model to evaluate only a strategically selected subset of simulation scenarios rather than all possible scenarios. By identifying and prioritizing scenarios with higher uncertainty or potential safety concerns for detailed model evaluation, the system achieves effective safety testing with reduced computational energy consumption compared to exhaustive evaluation of every scenario
Data Source
AI summary
Systems, methods, and non-transitory computer-readable media configured to obtain one or more series of successive sensor data frames during a navigation of a vehicle. Disengagement data is obtained. The disengagement data indicates whether a vehicle is in autonomous mode. A training dataset with which train a machine learning model is determined based on the one or more series of successive sensor data frames and the disengagement data. The training dataset includes a subset of the one or more series of successive sensor data frames and a subset of the disengagement data, the machine learning model being trained to identify unsafe driving conditions.