Cascade Siamese-RPN for Accurate Aerial Visual Localization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing localization methods for Unmanned Aerial Vehicles (UAVs) face challenges in large uncertainty regions due to high computation complexity and limited detection accuracy, particularly in visual-aided localization tasks, and lack efficient optimization of rate and configuration in cascade modeling.
Innovation Solution
A computer-implemented method using a cascade Siamese-RPN architecture with multi-feature pyramid extraction and end-to-end training for optimized classification stages, incorporating parallel paths for different detection modalities, to enhance detection accuracy and reduce computation complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional image retrieval methods (Bag-of-Words, VLAD, Siamese CNN) are used for visual localization, then detection accuracy is improved, but computation complexity increases significantly
Solution Approach 1:
The patent divides the visual localization task into two independent stages: a coarse search stage that identifies candidate regions, and a fine alignment stage that performs high-accuracy 6-DoF estimation. This segmentation allows each stage to be optimized independently, reducing overall computation complexity while maintaining detection accuracy.
Solution Approach 2:
The patent introduces a cascade architecture with multiple stages that dynamically processes candidate regions. The first stage generates candidate regions with lower computational cost, and subsequent stages progressively refine these candidates with increasing accuracy, adapting the computation level to the uncertainty region size.
2Productivity
If cascade architecture with multiple stages is used to reduce computation complexity, then processing speed is improved, but detection accuracy may deteriorate due to information loss at each stage
Solution Approach 1:
The patent implements a nested cascade structure where subsequent stages are embedded within the output of previous stages. The coarse search stage outputs candidate regions that serve as input to the fine alignment stage, with each stage nested within the previous one. This nesting ensures that information is preserved and refined rather than lost, maintaining detection accuracy while improving processing speed.
Solution Approach 2:
The cascade architecture incorporates feedback mechanisms where detection results from each stage are fed back to adjust and refine subsequent stages. The candidate regions identified in the first stage provide feedback to guide the second stage's processing, ensuring that accuracy is maintained while benefiting from the speed improvements of the multi-stage approach.
3Measurement precision
If Siamese-RPN is used for detection, then detection capability is improved, but the shallow nature of feature extraction subnetwork limits performance
Solution Approach 1:
The patent applies preliminary action by performing coarse feature extraction and candidate region identification in the first stage before proceeding to deeper feature analysis in subsequent stages. This preliminary processing reduces the search space and allows the shallower network to achieve effective detection capability without requiring extremely deep feature extraction networks from the start.
Data Source
AI summary
A computer-implemented method for processing a geo-location of an aerial platform, comprises: (a) providing a set of classification stages, which is trained end-to-end to provide a set of cascade decisions, wherein each cascade decision is optimized for a different classification modality and for different spatial resolution; (b) receiving a geo-location query comprising an aerial sensor image, a reference bird's eye view image-map with coordinate system and sensor pose angles measures with respect to the coordinate system; (c) projecting the aerial sensor image onto the coordinate system of the reference image-map using the pose angles measures; (d) matching between the projected sensor image and corresponding image patches in the reference satellite image-map, by applying said set of classification stages in accordance with features extraction from said images and similarity criteria; (e) generating a cascade sequencing, wherein a classification score is aggregated along a cascade path among said set of cascade stages; (f) generating a cascade decision at each stage along the cascade path based on the aggregated classification score and on defined decision logic between successive stages; (g) retrieving a geo-location associated with said aggregated classification decision; and (h) sending the retrieved geo-location to a geo-location requester.