Road Feature Classification for Visual Inertial Odometry Scale Drift
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Inertial measurement units in visual inertial odometry systems for vehicles face challenges such as scale drift and unobservable camera pose at constant velocities, leading to inaccurate location determination of road features.
Innovation Solution
The system classifies road features by using precalibration data including camera height, pitch, and roll relative to the road plane, performing inverse projections, and applying machine learning models to determine feature depth and align image patches with the road plane normal, thereby providing independent scale measurements and reducing noise in inertial measurements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If visual inertial odometry is used for vehicle navigation, then location determination capability is improved, but scale drift and unobservable camera pose at constant velocities occur leading to measurement inaccuracies
Solution Approach 1:
The patent introduces road features as an intermediary element to establish a reference frame. By detecting and classifying features on the road plane (lane markings, curbs, signs) and using their known geometric relationships, the system creates a mediator reference system that allows accurate pose estimation without suffering from scale drift. The road features serve as a stable reference that breaks the unobservability problem at constant velocities.
Solution Approach 2:
The patent changes the parameter space by incorporating classification information about road features (type, orientation, position relative to vehicle) into the odometry calculation. By using the classified feature data to constrain the solution space and provide geometric constraints, the system transforms the unobservable pose parameters into observable ones through the known relationships between road features and the vehicle.
2Measurement precision
If machine learning models are applied for feature classification, then classification accuracy is improved, but computational complexity and processing time increase
Solution Approach 1:
The patent segments the feature classification process into distinct stages: detecting potential features in the image, classifying them as road-related or non-road-related, and then further classifying road features into specific types (lane markings, curbs, signs). This segmentation allows the system to apply appropriate processing at each stage and reduces the overall computational burden by filtering out non-road features early.
Solution Approach 2:
The patent performs preliminary classification to identify road features before detailed processing. By pre-classifying features as road-related or non-road-related based on initial detection, the system reduces the number of features that require complex machine learning analysis, thereby reducing computational complexity while maintaining accuracy for the critical road features.
Data Source
AI summary
An electronic device is described. The electronic device includes a memory and a processor in communication with the memory. The memory is configured to store precalibration data for a camera mounted on a vehicle, the precalibration data including a camera height determined relative to a road plane the vehicle is configured to contact during operation. The processor is configured to receive a plurality of images. The processor is also configured to classify one or more features in the plurality of images as road features based on the precalibration data.