HD Map-Augmented Ground Truth for Autonomous DNN Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional approaches for training deep neural networks (DNNs) in autonomous vehicles require substantial manual effort and costly ground truth data generation, which may be inaccurate due to human error, especially when high-definition (HD) maps are unavailable or outdated.
Innovation Solution
Leverage HD map information to automatically generate training labels and annotations for DNNs by localizing sensor data and correlating it with HD map data, reducing manual effort and improving accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling and annotating is used to generate ground truth data, then the DNNs can be trained with accurate labels, but the process requires substantial manual effort and time (upwards of twenty minutes per training data instance)
Solution Approach 1:
The patent uses HD map data as a template or reference copy to automatically generate ground truth labels for training data. Instead of manually creating labels from scratch, the system copies and adapts existing accurate map information (lane markings, road edges, intersections) to generate corresponding labels for sensor data, dramatically reducing manual labeling time while maintaining accuracy
Solution Approach 2:
The patent introduces HD map data as an intermediary between the training data and the ground truth labels. The map information serves as a mediating resource that bridges the gap between raw sensor data and accurate annotations, enabling automatic label generation without direct manual intervention for each training instance
2Adaptability or versatility
If manual labeling is used to generate ground truth data, then the data can be created for diverse scenarios, but human error during labeling or annotating may reduce the accuracy of the ground truth data
Solution Approach 1:
The system copies verified accurate information from HD maps (which are professionally created and validated) to generate ground truth labels, thereby eliminating human error in the labeling process while maintaining comprehensive coverage of diverse driving scenarios through the richness of map data
3Adaptability or versatility
If DNNs are trained without HD map information, then the system can operate in areas without HD maps, but the DNNs may not perform as accurately as desired for safe autonomous operation
Solution Approach 1:
The patent performs preliminary action by using HD map information to pre-generate accurate ground truth labels for training data before the DNNs are deployed to areas without HD maps. This advance preparation using rich map data creates robust trained models that can then operate reliably in environments where the maps are unavailable, effectively transferring the accuracy benefit to broader operational coverage
Data Source
AI summary
In various examples, training sensor data generated by one or more sensors of autonomous machines may be localized to high definition (HD) map data to augment and/or generate ground truth data—e.g., automatically, in embodiments. The ground truth data may be associated with the training sensor data for training one or more deep neural networks (DNNs) to compute outputs corresponding to autonomous machine operations—such as object or feature detection, road feature detection and classification, wait condition identification and classification, etc. As a result, the HD map data may be leveraged during training such that the DNNs—in deployment—may aid autonomous machines in navigating environments safely without relying on HD map data to do so.