Radar-Aided Visual Inertial Odometry Moving Object Removal
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Visual inertial odometry systems face large positioning errors when tracking visual features on moving vehicles, as they assume a static environment, leading to outliers and performance degradation in vehicular applications.
Innovation Solution
A RADAR-aided visual inertial odometry system that translates RADAR velocity and depth maps into camera images to detect and remove visual features corresponding to moving objects, refining estimated positions, orientations, and velocities by integrating RADAR-based depth initialization and outlier detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If visual inertial odometry tracks visual features on moving vehicles, then more visual features are available for tracking, but large positioning errors occur due to violation of the static environment assumption
Solution Approach 1:
The patent extracts and removes visual features that belong to moving objects from the set of tracked features. By using RADAR data to identify moving objects and their associated visual features, the system separates harmful moving features from useful static features, allowing selective tracking that maintains reliability while preserving feature quantity.
Solution Approach 2:
The patent introduces RADAR as an intermediary sensor to bridge the gap between visual features and their motion status. The RADAR data serves as a mediator that provides depth and velocity information to determine whether visual features belong to moving or static objects, enabling intelligent feature selection without direct visual-motion coupling.
2Device complexity
If visual inertial odometry assumes a static environment, then the system remains simple and computationally efficient, but positioning accuracy degrades in dynamic environments with moving vehicles
Solution Approach 1:
The patent makes the visual inertial odometry system multi-functional by integrating it with RADAR data processing capabilities. The system can now handle both static and dynamic environments using the same core VIO algorithm, with the added functionality of RADAR-based moving object detection and feature filtering, thereby improving reliability without fundamentally changing the system architecture.
3Measurement precision
If RADAR velocity map is translated to form 3D RADAR velocity image with depth initialization, then moving object detection accuracy improves, but computational complexity increases
Solution Approach 1:
The patent performs depth initialization using RADAR data before the visual feature tracking process. By pre-computing depth information from RADAR measurements and storing it in a 3D velocity image, the system prepares motion status data in advance, which speeds up the subsequent moving object detection process and reduces real-time computational burden.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach significantly reduces positioning errors and improves the accuracy of vehicle tracking by distinguishing moving objects from static features, enhancing the reliability of visual inertial odometry in dynamic environments.
Implementation Method 1
translate a radio detection and ranging (RADAR) velocity map in at least one image plane of at least one camera, to form a three-dimensional RADAR velocity image
Data Source
AI summary
Various embodiments disclose a device with one or more processors which may be configured to translate a RADAR velocity map in at least one image plane of at least one camera, to form a three-dimensional RADAR velocity image. The 3D RADAR velocity image includes a relative velocity of each pixel in the one or more images, and the relative velocity of each pixel is based on a RADAR velocity estimate in the three-dimensional RADAR velocity map. The one or more processors may be configured to determine whether visual features correspond to a moving object based on the relative velocity of each pixel determined, and may be configured to remove the visual features that correspond to a moving object, prior to providing them as an input into a state updater, in a RADAR-aided visual inertial odometer.