Sensor-View Label Transformation for Autonomous Driving Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The lack of appropriate training data for neural networks in autonomous and assisted driving systems limits their effectiveness, as existing labels often have different views than the environmental sensors they are intended to support, making them unsuitable for training.
Innovation Solution
A method that transforms top view or polar representations of traffic environments into perspective representations aligned with the view of environmental sensors, using a transformation matrix that includes a sensor calibration matrix to create suitable training data for neural networks, allowing the use of diverse sensor data for improved training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If top view or polar representations of traffic environments are used as training data, then the quantity of training data increases, but the suitability for training neural networks with environmental sensors decreases due to mismatched views
Solution Approach 1:
The patent introduces an intermediary transformation process that converts top view or polar representations into perspective representations. This intermediary step bridges the gap between the available training data format and the required sensor view format, allowing data from one coordinate system to be adapted for use with sensors in another coordinate system.
Solution Approach 2:
The patent applies parameter changes by transforming the coordinate system parameters of the training data. The transformation matrix modifies the spatial parameters (x, y, z coordinates) and viewing angle parameters to convert data from top view or polar coordinates to perspective coordinates, making the training data suitable for environmental sensors.
2Reliability
If training data is transformed to match sensor perspectives, then the suitability of training data improves, but the complexity of data processing increases
Solution Approach 1:
The patent applies preliminary action by performing the coordinate transformation during the data preparation and labeling phase, before the actual neural network training begins. The transformation matrix is pre-calculated and applied to create properly formatted training data sets, so that the complex transformation work is completed in advance rather than during training execution.
3Adaptability or versatility
If diverse sensor data from different sensor types is used for training, then the versatility of training data increases, but the difficulty of processing and aligning different coordinate systems increases
Solution Approach 1:
The patent creates a universal transformation framework that can handle multiple sensor types (cameras, lidars, radars) and multiple coordinate systems (top view, polar, perspective). The same transformation matrix approach can be applied universally to convert data from any sensor type's coordinate system to the target perspective coordinate system, making the system multi-functional and adaptable to diverse sensors.
Data Source
AI summary
An autonomous/assisted driving system (ADS) obtains a top view representation of labels of a traffic environment in Cartesian coordinates from sample data. The system obtains a transformation matrix for transforming Cartesian coordinates in an observation coordinate system of a perspective of an environmental sensor of the ADS. The transformation matrix is applied to the top-view representation of the labels to obtain a perspective representation of the traffic environment in the observation coordinate system.