Occupancy Grid and Semantic Map Discrepancy Detection for Autonomous Vehicle Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in accurately identifying objects with unknown semantic classes due to the need for large and comprehensive training data sets, which are difficult to generate, leading to potential operational hindrances when encountering unfamiliar objects.
Innovation Solution
The method involves comparing occupancy grids and semantic segmentation maps to identify discrepancies, extracting sensor data from these discrepancies, and using this data to generate training sets that can improve the accuracy of machine learning models by retraining them to recognize previously unknown objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large and comprehensive training data sets are used to improve object classification accuracy, then the ability to recognize unknown semantic classes improves, but the difficulty and resource requirements for generating such data sets increase significantly
Solution Approach 1:
The system uses its own operational data (sensor data, occupancy grids, segmentation maps) to automatically generate training examples without requiring external data sources. The conflict detection mechanism identifies discrepancies that serve as self-generated training data, enabling the system to improve its own classification capabilities through self-service data generation
Solution Approach 2:
The patent introduces an intermediary process that compares occupancy grids with segmentation maps to identify conflicts. This intermediary conflict detection mechanism serves as a bridge between existing system operations and training data generation, automatically extracting valuable training examples from operational discrepancies without requiring manual data collection
2Speed
If traditional semantic segmentation models are used, then processing speed is maintained, but the ability to accurately identify objects with unknown semantic classes deteriorates
Solution Approach 1:
The system implements a feedback mechanism where detected conflicts between occupancy grids and segmentation maps are used to generate training data, which is then used to retrain and improve the segmentation model. This closed-loop feedback enables continuous improvement of unknown object identification accuracy while maintaining processing efficiency through iterative model enhancement
Solution Approach 2:
The patent performs preliminary conflict detection between occupancy grids and segmentation maps to identify potential training examples before full model retraining is performed. This preliminary action allows the system to prepare training data in advance and selectively update models based on detected discrepancies, maintaining processing speed while improving accuracy
Data Source
AI summary
Provided are methods, systems, and computer program products for programmatic detection of objects with an unknown or uncertain semantic class, and generation of a training data set for such objects that can facilitate further development of machine learning models trained to extract semantic information. Embodiments of the present disclosure can detect discrepancies between sensor data corresponding to multiple representations of an environment, and utilize the identified discrepancies to programmatically select a portion of the sensor data for use as a training data set, e.g., for input to a machine learning model.