Vision Training Using Future Frames for Vehicle Ground Truth
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional approaches to training machine learning algorithms for vision-only systems in vehicles are inefficient due to the need for manual ground truth label determination, which is impractical for large-scale implementation and often results in incomplete or ambiguous data interpretation, requiring additional sensors for confirmation.
Innovation Solution
A network service that processes captured vision system data to automatically generate ground truth labels by utilizing subsequent data sets to update and refine initial label determinations, allowing for the generation of updated machine learning algorithms without additional sensors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual ground truth label determination is used for training vision systems, then label accuracy can be ensured, but the training process becomes inefficient and impractical for large-scale implementation
Solution Approach 1:
The system uses its own vision system outputs to generate ground truth labels. The vision system processes images to detect objects and their attributes, and these detections are automatically used as ground truth labels for training, eliminating the need for external manual annotation while maintaining scalability
Solution Approach 2:
The system performs preliminary processing of vision data to generate training labels before actual model training. By pre-processing the vision system outputs and creating labeled datasets in advance, the system prepares training data efficiently at scale without requiring manual intervention during the training phase
2Measurement precision
If additional sensors are added to confirm vision system data, then data interpretation accuracy improves, but system complexity and cost increase
Solution Approach 1:
The patent replaces physical sensor systems (雷达, LIDAR) with a computational approach using vision system data processing and machine learning algorithms. The system uses software-based object detection and attribute extraction to achieve accurate data interpretation without additional hardware sensors
Solution Approach 2:
The system creates virtual representations of the environment through image processing and generates synthetic training data that mimics real-world scenarios. This allows the vision system to learn from copied representations without needing physical sensors to verify each detection
3Loss of information
If manual ground truth label determination is used, then data completeness can be verified, but the process becomes impractical for large-scale implementation
Solution Approach 1:
The vision system automatically generates complete training datasets by processing large volumes of images through its detection algorithms. The system self-generates ground truth labels for multiple attributes (object type, position, velocity, acceleration) simultaneously, ensuring data completeness while maintaining scalability to large datasets
Solution Approach 2:
The system performs preliminary extraction of all relevant attributes (object detection, classification, positioning, motion parameters) from vision data before training. This pre-processing ensures that complete training data is generated in advance at scale, with all necessary information extracted systematically from the image sequences
Data Source
AI summary
Aspects of the present application correspond to utilization of a combined set of inputs to generate or train machine learned algorithms for utilization in vehicles with vision system-only based processing. A network service can receive a first set of inputs (e.g., a first data set) from a target vehicle including captured vision system data at a first point in time. The network service can receive a second set of inputs (e.g., a second data set) from the target vehicle including captured vision system data at a second point in time. The second point in time is subsequent to the first point in time. Based on the second set of ground truth labels and values, the network service can then determine or derive labels and associated values for the first set of ground truth labels and values.