Sensor Point Cloud Transformation for ML Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The high cost and time expenditure associated with creating diverse and annotated training data sets for complex machine learning models, particularly in scenarios with variable sensor configurations, limit the applicability and efficiency of perception systems in real-world applications.
Innovation Solution
A method is developed to generate input data for machine learning models by transforming point clouds from one sensor perspective to another, eliminating non-detectable points and adjusting depth or disparity images to create a target sensor point cloud that mimics the data from a target sensor, allowing training even when data from the target sensor is not available.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If large measurement campaigns are conducted to collect diverse training data, then the diversity and completeness of training data improve, but the time expenditure and costs increase significantly
Solution Approach 1:
The patent creates virtual copies of sensor data by transforming point clouds from one sensor's perspective to another sensor's perspective through coordinate transformations. This allows generating training data for target sensors without physical measurement campaigns, eliminating the need to physically replicate diverse scenarios while maintaining data diversity through mathematical transformations and occlusion simulations.
2Measurement precision
If manual annotation is performed on recorded training data, then the quality and accuracy of ground truth data improve, but the time expenditure and costs increase
Solution Approach 1:
The patent transforms existing annotated point cloud data into target sensor perspectives while preserving the ground truth annotations. The occlusion handling and coordinate transformations maintain the accuracy of original annotations without requiring manual re-annotation, thus copying the labeled data structure to the target sensor's coordinate system while eliminating time-consuming manual annotation processes.
3Ease of manufacture
If training data is collected from one sensor configuration, then the data collection process is simplified, but the applicability to other sensor configurations decreases
Solution Approach 1:
The patent implements a universal data transformation framework that can adapt point cloud data from any source sensor configuration to any target sensor configuration through coordinate transformations and occlusion handling. This multi-functional approach allows a single data collection campaign to serve multiple sensor configurations, making the data collection process universally applicable across different sensor setups without requiring separate campaigns for each configuration.
4Measurement precision
If complex machine learning models are deployed to handle complex perception tasks, then the perception accuracy improves, but the requirement for annotated training data increases
Solution Approach 1:
The patent generates additional virtual training samples by transforming existing point cloud data into multiple target sensor perspectives. This copying and transformation process effectively multiplies the available training data from a single source, providing sufficient diverse training data for complex machine learning models without requiring proportionally larger physical measurement campaigns.
Data Source
AI summary
A method of generating input data for a machine learning model includes determining, for a sensor, a point cloud with points detected by the sensor from surfaces in the environment of the sensor, generating a preliminary target sensor point cloud for a target sensor by transforming, for the sensor, points of the determined point cloud into points from the perspective of the target sensor according to the relative position of the target sensor to the sensor, generating a target sensor point cloud for the target sensor by using the preliminary target sensor point cloud, wherein points which, due to one or more surfaces for which points exist in the preliminary target sensor point cloud, are not detectable by the target sensor are eliminated in the target sensor point cloud, and using the target sensor point cloud as input for the machine learning model.