Detour Path Data Augmentation for Autonomous Driving Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The complexity and unpredictability of road and traffic conditions pose challenges for training autonomous vehicles, as existing data collection and deep learning techniques are costly, inefficient, and result in insufficient or low-quality training data, limiting the accuracy of autonomous vehicle models.
Innovation Solution
Implementing data augmentation methods to enhance training data by adding or modifying road features, traffic safety objects, and driver images, which are scalable and cost-effective, allowing for high-fidelity coverage of diverse driving scenarios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If large amounts of real road and traffic data are collected and transmitted from vehicles to remote servers, then the training data quantity increases, but communication bandwidth consumption increases and costs become prohibitively expensive
Solution Approach 1:
The patent creates synthetic copies of road and traffic scenes by extracting background images and overlaying object images (vehicles, pedestrians, cyclists) to generate artificial training data. This copying approach produces diverse training scenarios without requiring actual data collection and transmission, thereby increasing training data quantity while avoiding communication bandwidth consumption
Solution Approach 2:
The patent extracts background images by removing traffic objects from original images, creating clean background scenes that can be reused multiple times with different object overlays. This extraction process enables efficient generation of diverse training data from limited source material, improving training data quantity without proportional increases in data transmission requirements
2Measurement precision
If deep learning techniques are trained using real collected training data, then model accuracy improves, but data collection costs and time consumption increase significantly
Solution Approach 1:
The patent performs preliminary image processing by extracting background images and organizing them into structured formats before they are needed for training. This preliminary preparation creates a ready-to-use library of background scenes that can be quickly combined with object images to generate training data, eliminating time-consuming real-world data collection while maintaining model training efficiency
Solution Approach 2:
The patent generates synthetic training data by copying and overlaying object images onto background images, creating diverse training scenarios without requiring additional real-world data collection. This copying process produces sufficient training data volume and variety to train deep learning models effectively, reducing both data collection costs and time consumption
3Ease of manufacture
If existing training data is used to train autonomous vehicle models, then training costs are reduced, but the diversity and quality of training scenarios become insufficient
Solution Approach 1:
The patent creates dynamic and diverse training scenarios by randomly selecting background images and object images, varying their positions, sizes, and orientations through overlay operations. This dynamic generation process produces infinitely varied training scenarios from static source images, greatly enhancing training scenario diversity without incurring additional data collection or transmission costs
Solution Approach 2:
The patent segments training data into separate background components and object components, allowing independent manipulation and recombination. This segmentation enables the creation of diverse training scenarios by mixing and matching different backgrounds with different objects, maintaining low training costs while significantly improving scenario diversity and coverage of edge cases
Data Source
AI summary
This application is directed to augmenting training images used for generating vehicle driving models. A computer system obtains a first image of a road, identifies within the first image a drivable area of the road, obtains an image of a traffic safety object, and determines a detour path on the drivable area. The computer system determines positions of a plurality of traffic safety objects to be placed adjacent to the detour path, and generates a second image from the first image by adaptively overlaying a respective copy of the image of the traffic safety object at each of the positions of the plurality of traffic safety objects on the drivable area within the first image. The second image is added to a corpus of training images to be used by a machine learning system to generate a model for facilitating driving a vehicle (e.g., at least partial autonomously).