Ultrasonic Data Augmentation for Bias-Resistant Autonomous Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional techniques face challenges in training machine learning models to effectively generalize various scenarios and objects using ultrasonic sensor data, leading to biases based on sensor type, road type, and sensor pose, which affects the accuracy of environmental representations for autonomous and semi-autonomous machines.
Innovation Solution
The system augments ultrasonic sensor data to include information about different driving environments, machine poses, and additional sensor details, using a new architecture to generate more accurate input data for machine learning models, reducing biases and improving map generation accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional ultrasonic sensor data is used for training machine learning models, then the training process is simple, but the model develops biases based on sensor type, road type, and sensor pose, reducing generalization ability
Solution Approach 1:
The system performs data augmentation in advance during the training phase, generating synthetic ultrasonic data with varied sensor poses, road types, and environmental conditions. This preliminary preparation of diverse training data enables the model to learn robust features that generalize across different scenarios without requiring complex real-world data collection for every possible condition.
Solution Approach 2:
The system creates synthetic copies of ultrasonic sensor data by simulating various sensor configurations, road surfaces, and environmental scenarios. These copied and modified data samples are used to train the model, allowing it to learn from diverse scenarios without requiring physical sensors in every possible configuration or environment.
2Measurement precision
If machine learning models are trained on limited ultrasonic data, then training is faster and easier, but the models produce inaccurate environmental representations for autonomous navigation
Solution Approach 1:
Comprehensive data augmentation is performed in advance during the training phase, generating diverse synthetic ultrasonic data covering various sensor poses, road types, and environmental conditions. This preliminary preparation of extensive training data enables the model to learn robust features that generalize across different scenarios, achieving high accuracy without requiring lengthy training on limited real-world data.
Solution Approach 2:
The system varies multiple parameters in the synthetic data generation process, including sensor position, orientation, road surface characteristics, and environmental conditions. By systematically changing these parameters during data augmentation, the model learns to handle variations in real-world conditions, improving measurement precision for environmental representations.
3Adaptability or versatility
If ultrasonic sensors are fixed in specific positions on machines, then the hardware configuration is simple, but the system cannot accurately represent environments from different machine poses
Solution Approach 1:
The system creates synthetic copies of ultrasonic sensor data corresponding to different sensor positions and orientations on the machine. By generating and training on these copied data samples with varied poses, the model learns to accurately represent environments from different machine perspectives, achieving adaptability without physically installing sensors in multiple locations.
Solution Approach 2:
The system systematically varies sensor position and orientation parameters during data augmentation, generating synthetic training data that reflects different sensor configurations. This allows the model to learn robust environmental representations that are invariant to sensor placement, achieving adaptability without increasing physical sensor complexity.
Data Source
AI summary
In various examples, ultrasonic data augmentations for autonomous and/or semi-autonomous systems and applications are described herein. Systems and methods described herein may use sensor data generated using one or more ultrasonic sensors to generate augmented input data for training one or more machine learning models to generate one or more representations (e.g., one or more maps) of an environment. As described herein, the sensor data may be augmented using one or more techniques such that the augmented input data corresponds to various driving environments (e.g., different driving surfaces), various poses on machines (e.g., different locations and/or orientations), and/or includes additional information associated with the ultrasonic sensor(s) and/or the sensor data. The systems and methods described herein may further use a new architecture to generate input data for the machine learning model(s), where the input data better represents the environment surrounding a machine executing the machine learning model(s).