Stereo Vehicle Pose Estimation With Selective Depth Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for estimating the pose of obstacles in autonomous driving are computationally intensive and difficult to perform on-the-fly, hindering effective obstacle avoidance in dynamic environments.
Innovation Solution
A system and method for estimating the pose of objects using stereo images from multiple cameras on an ego vehicle, involving object detection, feature point extraction, orientation estimation, and pose ambiguity resolution to dynamically determine the pose of nearby moving vehicles.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional depth map construction from stereo images is used, then depth information can be obtained, but computational complexity increases significantly
Solution Approach 1:
The patent extracts only the essential depth information needed for pose estimation rather than constructing a complete depth map of the entire scene. By selectively obtaining depth data at specific locations (object positions) rather than globally, the computational burden is reduced while maintaining the necessary measurement precision for obstacle pose estimation.
Solution Approach 2:
The system performs preliminary object detection and position identification before depth estimation. By first locating objects of interest in the stereo images and then only computing depth at those specific positions, the system avoids the computationally intensive process of generating depth information for the entire scene, thus reducing overall computational complexity.
2Reliability
If comprehensive pose estimation is performed, then obstacle avoidance accuracy improves, but processing time increases
Solution Approach 1:
The patent implements partial pose estimation by focusing only on the essential pose parameters needed for obstacle avoidance (position, orientation of detected objects) rather than computing complete pose information for all scene elements. This selective approach maintains sufficient accuracy for safety-critical decisions while significantly reducing processing time.
Solution Approach 2:
The system uses the detected object features and positions to directly compute pose information without requiring extensive additional processing. The pose estimation leverages the already-extracted object detection results, making the system self-sufficient and reducing redundant computations that would increase processing time.
3Measurement precision
If multiple sensors are deployed for pose estimation, then measurement accuracy improves, but system complexity increases
Solution Approach 1:
The patent makes the stereo camera system multi-functional by using it for both object detection and depth/pose estimation. Instead of adding separate depth sensors (like time-of-flight cameras or structured light sensors), the system extracts depth information from the existing stereo image pairs, allowing the same hardware to serve multiple purposes and improving pose estimation accuracy without increasing sensor system complexity.
Solution Approach 2:
The system merges the object detection function and depth estimation function into a unified processing pipeline. By combining these functions that could be performed by separate sensor systems into a single stereo vision-based approach, the patent achieves accurate pose estimation while avoiding the complexity of integrating and calibrating multiple independent sensor systems.
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
The present teaching relates to method, system, medium, and implementation of estimating a pose of an object in autonomous driving. Stereo images are obtained from a scene via at least two cameras. Objects present in the scene are then detected from the stereo images. For each object detected, a plurality of feature points are identified and depth information associated with the feature points are obtained. An orientation of each object is then estimated based on the feature points and depth information of the object and the pose of the object is then estimated based on the orientation and additional features of the object.