Drone Pose Estimation via Semantic Score Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
State-of-the-art visual odometry systems rely on hand-crafted features that are easy to detect but lack effectiveness for pose estimation, particularly due to high uncertainty in distant points and uninformative textures from objects like trees and buildings.
Innovation Solution
A novel CNN architecture that computes score maps considering semantics and motions to select beneficial features for pose estimation, allowing for accurate prediction without the need for an inertia measurement unit and handling ambiguous features through deep supervision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If hand-crafted features (SIFT, FAST, ORB) are used for visual odometry, then feature detection and tracking are easy, but pose estimation accuracy deteriorates due to high uncertainty in distant points and uninformative textures
Solution Approach 1:
The patent changes the fundamental parameter of feature selection from hand-crafted geometric descriptors to deep learning-based semantic features. The CNN architecture transforms image features through multiple convolutional layers, changing the representation parameters from simple gradient orientations to semantic feature vectors that capture meaningful object information, thereby improving pose estimation accuracy while maintaining ease of operation through automated feature extraction
Solution Approach 2:
The patent substitutes the mechanical hand-crafted feature extraction process with a neural network-based system. Instead of manually designing feature detectors like SIFT and ORB, the system uses a CNN to automatically learn and extract features that are optimized for pose estimation, replacing the mechanical feature detection pipeline with an intelligent learning-based approach
2Ease of operation
If distant points are used for tracking, then tracking is easy due to small motions in images, but pose estimation uncertainty increases
Solution Approach 1:
The patent applies local quality by making the feature selection process adaptive to different spatial locations. The CNN processes features locally across the image and generates confidence scores for each feature point, allowing the system to weigh and select features based on their local quality and reliability. This enables the system to automatically prefer nearby points with high information content while still utilizing distant points when they provide reliable tracking signals
3Quantity of substance
If features from trees or buildings are used, then feature quantity is abundant, but pose estimation quality deteriorates due to ambiguous textures
Solution Approach 1:
The patent introduces an intermediary confidence scoring mechanism between feature extraction and pose estimation. The CNN processes features from abundant sources like trees and buildings, then generates confidence scores that act as intermediaries to filter and weight these features. This intermediary layer allows the system to utilize the quantity of available features while filtering out ambiguous textures through learned confidence assessments, improving overall pose estimation quality
Data Source
AI summary
A computer-implemented method, system, and computer program product are provided for a stabilization system utilizing pose estimation in an aerial drone. The method includes receiving, by a pose estimation system, a plurality of images from one or more cameras. The method also includes predicting, by the pose estimation system, a pose from the score map and a combined feature map, the combined feature map correlated from a pair of the plurality of images. The method additionally includes moving, by a propulsion system, the aerial drone responsive to the pose.