3D Road Environment Data Generation for Neural Network Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Collecting training data for neural networks in intelligent traffic control systems is hindered by privacy regulations, requiring significant manpower and resources, and existing datasets for night and rainy conditions are insufficient and inaccurate.
Innovation Solution
A data generation apparatus and method that uses 3D graphic road environment data to create virtual captured images and ground truth data, allowing for the generation of training data for neural networks without violating privacy regulations, at a lower cost and with high precision, by simulating various conditions such as time and weather.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If actual photographs are collected from local governments for training data, then real-world traffic data can be obtained, but privacy protection regulations prevent collection and enormous manpower and costs are required
Solution Approach 1:
The patent creates virtual copies of real-world traffic scenes using 3D rendering technology. Instead of collecting actual photographs from cameras, the system generates synthetic images that replicate real traffic environments, including vehicles, pedestrians, road conditions, and environmental factors like weather and time of day. This copying approach eliminates privacy concerns while providing sufficient training data quality.
Solution Approach 2:
The patent introduces a 3D rendering engine as an intermediary between the need for real traffic data and the constraints of privacy regulations. This intermediary system processes simplified 3D scene data into photorealistic images that serve as training data, avoiding direct collection from sensitive sources while maintaining data reliability.
2Measurement precision
If manual annotation of training data is performed to ensure high-precision ground truth, then accurate training data can be generated, but enormous manpower and costs are required
Solution Approach 1:
The patent copies ground truth information directly from the 3D scene data used to generate the rendered images. Since the 3D scenes contain precise object positions, sizes, and attributes by definition, the ground truth can be automatically extracted without manual annotation, maintaining high precision while dramatically improving productivity.
Solution Approach 2:
The system makes the 3D scene data self-annotating. The same data structure that defines the virtual scene geometry and object properties automatically provides the ground truth labels needed for training, eliminating the need for separate manual annotation processes.
3Ease of manufacture
If training data is collected from existing datasets, then data generation cost is reduced, but datasets for night and rainy environments are insufficient and GT information is inaccurate
Solution Approach 1:
The patent implements dynamic environmental control in the 3D rendering system. Weather conditions (rain, snow, fog), time of day, lighting conditions, and road surface states can be dynamically adjusted and combined in various configurations, allowing generation of training data for any environmental condition without being limited by existing dataset availability.
Solution Approach 2:
The system changes environmental parameters (weather, time, lighting) in the 3D rendering process to generate diverse training scenarios. By modifying these parameters, the same base 3D scene can produce multiple variations suitable for different environmental conditions, greatly expanding adaptability while keeping generation costs low.
Data Source
AI summary
There is provided an apparatus for generating a training data. The apparatus comprises a memory storing instructions; and a processor executing the instructions, wherein the instructions, when executed by the processor, cause the processor to: prepare 3D graphic road environment data required for rendering of 3D graphic road environment including a road and at least one object moving on the road, and set a photographing environment of a camera capturing the road and the at least one object moving on the road within the rendered 3D graphic road environment, generate a virtual captured image obtained by capturing the road and the at least one object moving on the road in the 3D graphic road environment based on information on the photographing environment of the camera, and extract training ground truth data from the virtual captured image to generate the training data including the virtual captured image and the training ground truth data.