Learning Data Collection for Unexpected Hazard Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current technologies for recognizing dangerous objects using learning models, such as neural networks, face challenges in accurately identifying objects that may cause accidents, as there is a gap between human-defined dangers and actual hazards, making it difficult to learn all potential danger objects effectively.
Innovation Solution
A collection device and learning system that determines a moving body's dangerous state using sensor output values, specifying images from captured video as learning data based on the timing of the dangerous state, allowing for the collection of learning data that includes objects causing the specific state, which can include unexpected hazards.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If learning data is collected using only human-defined dangerous objects, then the learning model can be trained with predefined safety standards, but it fails to recognize unexpected hazards that fall outside human-defined categories
Solution Approach 1:
The system performs preliminary action by collecting learning data from situations where the moving body actually fell into a dangerous state, rather than relying solely on pre-defined dangerous objects. This allows the model to learn from real-world hazard scenarios including unexpected hazards that were not previously defined by humans.
Solution Approach 2:
The system uses feedback from sensor outputs that detect when the moving body is in a specific dangerous state to identify and collect relevant learning images. This feedback mechanism enables the system to continuously improve by learning from actual danger scenarios encountered during operation, expanding recognition beyond predefined categories.
2Reliability
If all potential danger objects are manually defined in advance, then comprehensive coverage of known hazards is achieved, but the system cannot adapt to new or unexpected hazard types
Solution Approach 1:
The system performs self-service by automatically collecting and organizing learning data from actual dangerous states encountered during operation. Instead of requiring continuous manual definition of new hazard types, the system autonomously identifies dangerous situations through sensor feedback and accumulates corresponding images for model training.
Solution Approach 2:
The system prepares learning data in advance by collecting images from dangerous states as they occur during normal operation, rather than requiring manual preparation of training data for each new hazard type. This preliminary collection builds a comprehensive dataset that includes both known and unexpected hazards.
3Quantity of substance
If learning data is collected from all captured images, then comprehensive training data is obtained, but the data collection becomes inefficient and includes irrelevant images
Solution Approach 1:
The system extracts only the relevant learning images from the captured image sequence by identifying the specific timing when the moving body falls into a dangerous state. Instead of collecting all captured images, it selectively extracts and stores only those images that correspond to actual dangerous situations, significantly improving data collection efficiency.
Solution Approach 2:
The system performs preliminary identification of dangerous states using sensor outputs during normal operation, marking the timing of dangerous events for later image extraction. This preliminary detection enables efficient retrieval of only relevant learning images when needed for model training.
Data Source
AI summary
A collection device of learning data of a learning model for detecting a danger of a moving body, includes a hardware processor that determines whether the moving body is in a specific state related to a danger of the moving body by using an output value of a sensor that detects the specific state, and specifies a part of images of an image group used for the danger detection as a learning image in which a cause of falling into the specific state is estimated to be captured on a basis of a timing at which the moving body is determined to be in the specific state.