Region-Specific Neural Network Models for Indoor Pose Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for estimating a device's pose in a space are inaccurate over time, require expensive and time-consuming calibration, and are resource-intensive, especially when dealing with changes in the environment or indoors where GPS is unreliable.
Innovation Solution
Training region-specific models using images or synthetic data to compute device poses in specific areas, allowing for iterative pose estimation with reduced memory and power requirements, and enabling accurate navigation and pose-related tasks without continuous remote processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a single global model is used to estimate device pose across the entire space, then the system is simpler to deploy, but accuracy deteriorates when the device moves to different regions with varying environmental characteristics
Solution Approach 1:
The patent divides the global space into multiple region-specific zones, each with its own trained model. The system segments the operational domain and assigns specialized models to each segment, allowing each model to optimize for local environmental characteristics while maintaining overall system accuracy across diverse regions
Solution Approach 2:
The patent implements local quality by training separate models for different spatial regions, where each model develops specialized characteristics optimized for its specific region's environmental features. This allows the system to adapt to local variations in lighting, textures, and geometric structures that a global model would average out
2Measurement precision
If region-specific models are used to improve pose estimation accuracy in different areas, then measurement precision improves, but device complexity increases due to multiple models
Solution Approach 1:
The patent implements dynamics by making the model selection process adaptive and iterative. The system dynamically selects and switches between region-specific models based on the device's current location and movement patterns, allowing the model set to evolve and adapt to the device's operational trajectory rather than statically loading all possible models
3Measurement precision
If multiple region-specific models are deployed to cover different areas, then pose estimation accuracy improves across the space, but memory requirements increase
Solution Approach 1:
The patent segments the model set into region-specific components that can be selectively loaded and unloaded based on the device's current location. This segmentation allows the system to maintain only the necessary subset of models in active memory at any given time, reducing overall memory requirements while preserving accuracy where needed
Solution Approach 2:
The patent implements discarding and recovering by unloading region-specific models from memory when they are no longer needed based on the device's movement away from their corresponding regions. The models can be recovered and reloaded when the device returns to those regions, optimizing memory utilization without sacrificing estimation accuracy
4Measurement precision
If iterative model switching is implemented to track device movement across regions, then pose estimation accuracy is maintained over time, but computational cost and processing time increase
Solution Approach 1:
The patent applies preliminary action by pre-training multiple region-specific models offline before deployment. This preliminary training allows the models to be ready for immediate use when needed, eliminating the need for time-consuming training during runtime and reducing the computational overhead associated with iterative model switching
Solution Approach 2:
The patent implements feedback mechanisms that monitor the device's pose estimation quality and regional location in real-time. This feedback allows the system to proactively switch models before accuracy degrades, reducing the need for frequent switching and minimizing the time loss associated with model transitions
Data Source
AI summary
A system for estimating a plurality of device poses in a space, comprising: at least one device, comprising at least one hardware processor connected to at least one sensor and adapted to: in at least one of a plurality of iterations: using at least one model to compute a first estimated device pose in a first identified region of the space in response to at least one input signal captured by the at least one sensor, where the at least one model is one of a plurality of region specific models, each trained to compute an estimated device pose in one of a plurality of regions of the space; providing the first estimated device pose to at least one software object executed by the at least one hardware processor to perform a pose-oriented task; receiving at least one other model trained to compute another estimated device pose in a second.