Fusion Processor for Automated Target Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic target recognition (ATR) systems fail to effectively fuse ATR scores across different sensors and poses due to pruning of data and lack of consideration for relative orientation, leading to ambiguity in target type prediction.
Innovation Solution
A method and system that fuse confidence values and azimuth angles from multiple images by aligning and normalizing data, using a fusion processor to produce a fused curve for each target type, predicting the target based on maximum values or areas under the curve within specific azimuth windows.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If ATR systems prune portions of data for efficiency or because specific ATR systems cannot provide accurate predictive data without pruning, then processing efficiency is improved, but data completeness and prediction accuracy deteriorate
Solution Approach 1:
The patent performs preliminary association of regions of interest between multiple images before fusion, establishing correspondence relationships in advance. This allows the fusion processor to efficiently access only relevant associated data during fusion operations, rather than processing all raw data, thus improving efficiency while preserving data completeness.
Solution Approach 2:
The patent introduces an intermediate association structure that links regions of interest across multiple images. This intermediary layer enables selective data retrieval and fusion without requiring all ATR systems to process complete unpruned datasets, resolving the conflict between efficiency and data completeness.
2Device complexity
If ATR systems fuse ATR scores without regard to detailed pose information within the pose space uncertainty, then fusion complexity is reduced, but prediction accuracy and ambiguity reduction deteriorate
Solution Approach 1:
The patent segments the fusion process into distinct stages: first associating regions of interest and their corresponding pose information, then fusing ATR scores while preserving pose details. This segmentation allows the system to handle pose information systematically without overwhelming complexity, while maintaining prediction accuracy through pose-aware fusion.
Solution Approach 2:
The patent extends the fusion process from simple score aggregation to multi-dimensional fusion that incorporates pose space uncertainty as an additional dimension. By treating pose information as a separate dimensional parameter rather than ignoring it, the system achieves more accurate predictions while managing complexity through structured dimensional handling.
3Speed
If ATR systems determine the best ATR evidence over a range of azimuth angles for each look without considering consistency or relationship criteria between pose information, then processing speed is improved, but hypothesis ambiguity and prediction reliability deteriorate
Solution Approach 1:
The patent implements feedback mechanisms where the association of pose information between images provides consistency checks that feed back into the fusion process. This feedback ensures that only consistent pose relationships are used in fusion, improving hypothesis reliability while maintaining processing speed through efficient feedback-based validation rather than exhaustive checking.
Solution Approach 2:
The patent performs preliminary association of pose information between images before the actual ATR score fusion. By establishing consistent pose relationships in advance, the system creates a filtered set of reliable hypotheses that speeds up subsequent fusion operations while ensuring reliability through pre-validated pose consistency.
Data Source
AI summary
A method of predicting a target type in a set of target types from at least one image is provided. At least one image is obtained. A first and second set of confidence values and associated azimuth angles are determined for each target type in the set of target types from the at least one image. The first and second set of confidence values are fused for each of the azimuth angles to produce a fused curve for each target type in the set of target types. When multiple images are obtained, first and second set of possible detections are compiled corresponding to regions of interest in the multiple images. The possible detections are associated by regions of interest. The fused curves are produced for every region of interest. In the embodiments, the target type is predicted from the set of target types based on criteria concerning the fused curve.