Vehicle Data Correlation Using Human Cues for Hazard Labeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in distinguishing between relevant and irrelevant sensor data, as artificial neural networks (ANNs) require substantial training to discern between important and unimportant information, similar to novice drivers, and traditional methods struggle to teach ANNs to recognize potential dangers without explicit cues.
Innovation Solution
The use of human speech, gaze, and gestures to label relevant sensor data, where processors analyze keywords and gaze direction to attribute relevance to sensor data, allowing ANNs to learn from human instructors and improve their ability to identify dangerous situations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional training methods are used to teach ANNs to distinguish relevant sensor data, then the ANN can eventually learn to identify important information, but the training process requires substantial time and computational resources
Solution Approach 1:
The patent introduces human instructors as intermediaries who provide labeled sensor data to train ANNs. Human instructors explicitly identify and label relevant objects and situations in sensor data, creating a training dataset that teaches ANNs to distinguish important information without requiring extensive trial-and-error training time.
Solution Approach 2:
Human instructors perform preliminary labeling of sensor data before the ANN processes it for autonomous driving decisions. This pre-labeling action creates a structured training dataset that prepares the ANN in advance, allowing it to learn relevant data patterns more efficiently during the training phase.
2Reliability
If ANNs are trained to recognize all potential dangers in sensor data, then safety improves, but the system complexity and computational load increase significantly
Solution Approach 1:
The patent applies local quality by having human instructors selectively label specific regions and objects in sensor data that are relevant to safety. Instead of requiring the ANN to analyze every pixel and object uniformly, the human-labeled data highlights locally important areas, allowing the ANN to focus computational resources on critical regions while maintaining high safety standards.
3Manufacturing precision
If human instructors provide detailed feedback on sensor data, then the quality of training data improves, but the time and resources required for data collection increase
Solution Approach 1:
The system enables self-service by allowing human instructors to provide feedback in a natural, intuitive manner during normal driving instruction. Instructors label relevant objects and situations as they naturally occur during teaching moments, rather than requiring structured, time-consuming data collection protocols. This approach maintains high data quality while improving collection efficiency through natural interaction.
Data Source
AI summary
A vehicle data relation device includes an internal audio/image data analyzer, configured to receive first data representing at least one of audio from within the vehicle or an image from within the vehicle; identify within the first data second data representing an audio indicator or an image indicator, wherein the audio indicator is human speech associated with a significance of an object external to the vehicle, and wherein the image indicator is an action of a human within the vehicle associated with a significance of an object external to the vehicle; an external image analyzer, configured to receive third data representing an image of a vicinity external to the vehicle; identify within the third data an object corresponding to at least one of the audio indicator or the video indicator; and an object data generator, configured to generate data corresponding to the object.