Vehicle AI Control Data Bundling for Balanced Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous driving technologies face challenges in accurately and efficiently dividing datasets for training and evaluation due to concentration of specific parameters in some data, leading to reduced accuracy and imbalanced distributions.
Innovation Solution
A vehicle control apparatus and method that classifies datasets into bundles based on predefined conditions, performs accuracy tests, and adjusts criteria to ensure balanced distributions between training and evaluation data, using sensors to detect external objects and adjust classification times to meet specified ratios and standard deviations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the dataset is simply divided at a predetermined ratio, then the division process is simple and quick, but specific parameters are concentrated in only some pieces of data and accuracy is reduced
Solution Approach 1:
The patent segments the dataset division process into multiple stages: initial bundling of sensor data, followed by iterative refinement through accuracy testing and rebundling. This multi-level segmentation allows the system to maintain both efficiency (through automated bundling) and accuracy (through controlled refinement), resolving the contradiction between simple division and accurate parameter distribution.
Solution Approach 2:
The patent implements a feedback mechanism where the accuracy test results are used to adjust subsequent bundling operations. The system tests whether divided datasets meet accuracy criteria, and if not, automatically rebundles the data and retests. This closed-loop feedback ensures that parameter concentration is corrected while maintaining an efficient automated process.
2Measurement precision
If adaptive criteria are used for dividing the dataset, then parameter distribution accuracy is improved, but the complexity of the division process increases
Solution Approach 1:
The patent enables the dataset division system to be self-correcting through automated accuracy testing and rebundling. The system automatically identifies parameter concentration issues, adjusts bundling criteria, and redivides data without manual intervention. This self-service approach handles the complexity internally while presenting a simplified interface to users.
Solution Approach 2:
The patent performs preliminary bundling of sensor data into structured groups before the actual division process. This preliminary organization of data by time, location, and sensor type creates a foundation that simplifies subsequent adaptive division operations, reducing the complexity of handling raw unstructured data during the accuracy-critical division phase.
3Loss of time
If the dataset is divided without considering parameter distribution, then the division process is fast, but the artificial intelligence model training accuracy is reduced
Solution Approach 1:
The patent applies partial refinement to the dataset division process by performing accuracy testing and selective rebundling only when necessary. The system initially divides data quickly, then applies refinement actions only to portions that fail accuracy criteria. This partial action approach maintains speed while ensuring reliability where it matters most.
Solution Approach 2:
The patent implements dynamic adjustment of bundling criteria based on real-time accuracy test results. The system adapts its division strategy during the process, modifying bundling parameters and rebundling data as needed to achieve target accuracy levels. This dynamic approach allows the system to balance speed and reliability by investing more time only where accuracy requirements demand it.
Data Source
AI summary
An apparatus for controlling driving of a vehicle is introduced. The apparatus may comprise a memory storing at least one instruction and a processor operatively coupled with the memory. The at least one instruction, when executed by the processor, is configured to cause the apparatus to obtain a dataset comprising a plurality of frames for driving control of the vehicle, classify the dataset into a plurality of bundles, and divide the plurality of bundles into training data and evaluation data. Based on the training data and the evaluation data satisfying a first condition, an accuracy test is performed. Based on the accuracy test satisfying a second condition, an artificial intelligence model is trained or its performance is evaluated. The apparatus may output a signal based on the trained artificial intelligence model or the evaluated performance of the artificial intelligence model and control, based on the signal, driving of the vehicle.