Vehicle Image Data Collection for Lane-Change-Based Obstacle Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods require extensive manual effort and time to prepare training data for determining obstacles in road images, especially when the frequency of obstacles is low, making it difficult to generate accurate determination models.
Innovation Solution
A data collection device that captures images of vehicle surroundings and determines if a lane change has occurred based on vehicle history and direction indicator information, using this data to selectively collect images as training data for a model that identifies road obstacles, and transmits this data to an external learning device for model generation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If manual selection and labeling of images is performed to prepare training data, then training data can be obtained for obstacle determination, but it requires extensive time and effort especially when obstacle frequency is low
Solution Approach 1:
The system automatically determines whether images contain obstacles by analyzing lane change patterns and vehicle behavior, eliminating the need for manual labeling. The processor autonomously identifies training data by detecting lane changes through direction indicator information and odometry data, and automatically selects images from periods without lane changes as obstacle-free training data.
Solution Approach 2:
The system pre-identifies candidate training data by analyzing lane change history before the actual training data collection is needed. By determining lane change patterns in advance and pre-selecting images from periods without lane changes, the system prepares training data automatically without requiring manual intervention when the data is actually needed.
2Measurement precision
If images containing obstacles are used for learning, then the determination model can learn obstacle detection, but the low frequency of obstacles makes it difficult to collect sufficient training data
Solution Approach 1:
Instead of directly selecting images containing obstacles for training, the system inverts the approach by selecting images from periods without lane changes as proxy indicators of obstacle-free conditions. By training on negative examples (periods without lane changes) rather than directly on obstacle images, the system indirectly learns to identify obstacles through what is absent.
Solution Approach 2:
The system uses lane change behavior as an intermediary indicator to indirectly identify training data quality. Rather than directly detecting obstacles in images for training purposes, the system uses lane change patterns (detected through direction indicators and odometry) as a mediator to infer whether images contain obstacles, thereby obtaining training data without direct obstacle detection during data collection.
Data Source
AI summary
A data collection device of an embodiment includes a processor. The processor is configured to execute a program to acquire an image obtained by capturing surroundings of a first vehicle, determine whether the first vehicle has performed a lane change during a determination period based on information indicating a traveling history of the first vehicle, and collect the image included in the determination period as training data for a determination model that determines whether there is an obstacle in a road in a case where it is determined that the first vehicle has not performed a lane change.