Synthetic Training Image Generation for Automated Vehicle Object Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Camera-based object recognition systems require a large number of training images to effectively recognize objects, especially when objects occur rarely in natural environments, making data collection time-consuming and inefficient.
Innovation Solution
A method using a shift-map algorithm to generate synthetic training images by replacing initial structural features in a base image with substitute features from a template image, maintaining the natural appearance and reducing data acquisition effort.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a large number of different training images are collected from natural environments, then the teaching result of the object recognition system is improved, but the time and effort required for data collection increases significantly
Solution Approach 1:
The patent creates synthetic training images by copying and transforming a single base image through geometric transformations (scaling, rotating, shearing) and adding synthetic objects to the background. This copying approach generates diverse training images without requiring extensive field data collection, thus improving teaching results while reducing data collection time and effort.
2Measurement precision
If training images for rare object classes are collected from natural environments, then the recognition accuracy for rare objects is improved, but the complexity of data gathering increases
Solution Approach 1:
The patent prepares a comprehensive set of synthetic objects and backgrounds in advance, along with transformation parameters. This preliminary preparation allows the system to generate training images for rare object classes on-demand without requiring complex field data gathering campaigns, thereby improving recognition accuracy for rare objects while reducing data gathering complexity.
3Reliability
If multiple different training images are obtained for each object class, then the classifier performance is improved, but the effort for provision of data increases
Solution Approach 1:
The patent creates a universal base image that can serve as the foundation for generating training images across multiple object classes. By applying different synthetic objects, backgrounds, and geometric transformations to this single base image, the system efficiently produces diverse training data for various classes (road signs, pedestrians, vehicles) without requiring separate data collection efforts for each class, thus improving classifier performance while maintaining high data provision efficiency.
Data Source
AI summary
In a method of generating a training image for teaching of a camera-based object recognition system suitable for use on an automated vehicle which shows an object to be recognized in a natural object environment, the training image is generated as a synthetic image by a combination of a base image taken by a camera and of a template image in that a structural feature is obtained from the base image and is replaced with a structural feature obtained from the template image by means of a shift-map algorithm.