Vehicle Perception Model Training With Weakly Annotated Worldviews
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for developing Automated Driving Systems (ADS) are hindered by the high cost and time-consuming process of creating hand-labelled datasets for training perception algorithms, which are essential for machine learning models, leading to significant expenses and inefficiencies in improving and verifying ADS features.
Innovation Solution
A method for weakly annotating data using a computing system that stores sensor data and perception data from vehicle-mounted sensors, forms a baseline worldview, and evaluates it against a perception model's output to identify matches and estimation errors, updating the model parameters using weakly supervised learning or transmitting data for off-board processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If hand-labelled datasets are used for training perception algorithms, then training data quality is improved, but annotation cost and time consumption increase significantly
Solution Approach 1:
The system performs preliminary actions by using the production perception system to generate baseline worldviews and pre-annotate data during normal operation. This preliminary annotation reduces the need for time-consuming manual labeling later, as the baseline worldview provides preprocessed training data that can be directly used or minimally refined for training perception algorithms.
Solution Approach 2:
The production perception system serves itself by generating baseline worldviews that automatically annotate sensor data. The system uses its own perception capabilities to create training data, eliminating the need for external manual annotation services. This self-service mechanism continuously produces training data during normal vehicle operation without requiring additional human resources.
2Measurement precision
If hand-labelled datasets are used for training perception algorithms, then training data quality is improved, but development cost increases significantly
Solution Approach 1:
The system eliminates external annotation services by using its own production perception system to generate baseline worldviews. This self-service approach converts the development cost structure from paying external annotators to using internal computational resources that are already deployed in production vehicles, significantly reducing development costs while maintaining data quality.
Solution Approach 2:
The production perception system serves multiple functions: it performs real-time perception for safe driving operations and simultaneously generates annotated training data for algorithm development. This multi-functionality eliminates the need for separate annotation processes and reduces overall development costs by leveraging existing system capabilities.
3Quantity of substance
If massive hand-labelled datasets are collected, then training data quantity is improved, but data processing time and storage requirements increase
Solution Approach 1:
The system performs preliminary processing of sensor data by generating baseline worldviews during normal operation. This preliminary action organizes and structures data in advance, creating ready-to-use training samples that reduce subsequent processing time. Data is pre-filtered and pre-annotated during collection, eliminating the need for extensive post-processing of raw datasets.
4Measurement precision
If domain expertise is required for annotation, then annotation quality is improved, but annotation speed decreases
Solution Approach 1:
The production perception system, which incorporates domain expertise in its algorithms, generates baseline worldviews automatically. This self-service mechanism applies expert-level annotation quality at machine speed, eliminating the trade-off between quality and speed. The system processes data at computational speeds while maintaining the quality standards embedded in the production perception algorithms.
Data Source
Figure 1
Figure 2(a)~2(d)
Figure 3
AI summary
The present invention relates to methods and systems that utilize the production vehicles to develop new perception features related to new sensor hardware as well as new algorithms for existing sensors by using federated learning. To achieve this the production vehicle's own worldview is post-processed and used as a reference, towards which the output of the software (SW) or hardware (HW) under development is compared. In case of a large discrepancy between the baseline worldview and perceived worldview by the module-under-test, the data is weakly annotated by the baseline worldview. Such weakly annotated data may subsequently be used to update the SW parameters of the "perception model" in the module-under-test in each individual vehicle, or to be transmitted to the "back-office" for off-board processing or more accurate annotations.