Autonomous Vehicle Test Data Labeling with Removable Hardware Pods
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current autonomous vehicle technologies face challenges in efficiently generating and processing large amounts of training and testing data for machine learning models, particularly in complex environments, due to the time-consuming and error-prone nature of manual data labeling and the difficulty in accurately labeling objects at varying distances and occlusions.
Innovation Solution
The use of removable hardware pods mounted on additional vehicles to collect and synchronize data with autonomous vehicles, enabling automatic labeling and generation of training and testing instances through a computing system that processes and synchronizes data from multiple sources, including sensors and CAN bus data, to improve the accuracy and efficiency of machine learning model training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual data labeling is used for autonomous vehicle training data, then labeling accuracy can be maintained through human judgment, but the process becomes time-consuming and error-prone
Solution Approach 1:
The patent uses additional vehicles equipped with sensors to capture data that is then mapped and copied to the autonomous vehicle's perspective. This creates synthetic training data instances that replicate real-world scenarios without requiring manual labeling of each individual case, thereby reducing time while maintaining accuracy through the use of ground truth data from the additional vehicles.
Solution Approach 2:
The system enables autonomous vehicles and additional vehicles to automatically collect, synchronize, and generate training data without human intervention. The vehicles self-service by capturing sensor data, determining spatial relationships, and creating labeled training instances autonomously, eliminating the time-consuming manual labeling process while maintaining accuracy through automated ground truth generation.
2Adaptability or versatility
If data is collected from multiple additional vehicles to increase data diversity, then training data robustness improves, but system complexity and processing requirements increase
Solution Approach 1:
The patent divides the data collection system into modular components: autonomous vehicles, additional vehicles with sensors, and a synchronization system. Each component performs a specific function (data collection, spatial relationship determination, training instance generation), allowing the system to scale by adding more vehicles while maintaining manageable complexity through clear functional segmentation.
Solution Approach 2:
The additional vehicles serve multiple functions: they collect sensor data, provide ground truth information, and enable generation of training instances from various perspectives and scenarios. This multi-functionality increases data diversity without proportionally increasing system complexity, as the same vehicles perform multiple roles in the data collection ecosystem.
3Productivity
If automatic labeling is implemented using additional vehicle data, then data generation speed increases, but labeling accuracy may be compromised without manual verification
Solution Approach 1:
The patent uses additional vehicles as intermediaries that capture ground truth data about the environment and other vehicles. This intermediary data serves as a reliable reference that automatically labels the autonomous vehicle's sensor data, providing both high-speed automatic labeling and high accuracy through the use of authentic ground truth information from vehicles positioned in the actual environment.
Data Source
AI summary
Sensor data collected from an autonomous vehicle can be labeled using sensor data collected from an additional vehicle. Labeled sensor data can generate targeted testing instances for a trained machine learning model, where the trained machine learning model is used in generating control signals for an autonomous vehicle. In many implementations, targeted training instances can generate an accuracy value for the trained neural network model. Additionally or alternatively, the sensor suite on the additional vehicle can include a removable hardware pod which can be mounted on a variety of vehicles.