Sensor Tracking with Machine Learning Predictors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing tracking systems using sensor data face challenges in accuracy and speed due to complex computations required for minimizing parameter values, often resulting in local solutions rather than global ones, especially when selecting a good starting point is difficult.
Innovation Solution
The use of multiple machine learning predictors trained to predict diverse parameter values, with an optional optimizer refining predictions and a selector choosing the best output to improve tracking accuracy and speed, particularly in applications like camera pose tracking and object pose estimation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If minimization computations are used to compute parameter values, then tracking accuracy may be improved, but computation time and complexity increase significantly
Solution Approach 1:
The system pre-computes and stores a lookup table of parameter values and their corresponding cost values before tracking operations begin. During actual tracking, the system queries this pre-computed table rather than performing minimization computations in real-time, significantly reducing computation time while maintaining tracking accuracy.
Solution Approach 2:
The invention creates a simplified copy of the complex minimization problem by pre-computing the relationship between parameters and costs, storing this information in a lookup table. This copy allows the system to retrieve pre-determined optimal values without re-solving the complex optimization problem during time-critical tracking operations.
2Measurement precision
If gradient descent search is used to find parameter minima, then parameter optimization may be achieved, but the system may converge to local minima rather than global minima
Solution Approach 1:
The system pre-computes the cost function values for multiple parameter combinations and stores them in a lookup table before tracking begins. This allows the system to identify global minima in advance without relying on gradient descent methods that may get trapped in local minima during real-time operations.
Solution Approach 2:
The invention replaces the mechanical gradient descent search process with a lookup-based retrieval system. Instead of iteratively adjusting parameters based on gradients, the system directly queries pre-computed optimal values from the lookup table, eliminating the risk of local minimum convergence.
3Measurement precision
If multiple machine learning predictors are used to predict diverse parameter values, then tracking accuracy improves, but device complexity increases
Solution Approach 1:
The system combines multiple machine learning predictors into a unified tracking framework where each predictor contributes to estimating different aspects of the tracked object's state. The results from multiple predictors are integrated through a selection mechanism that chooses the most reliable prediction, achieving improved accuracy without proportionally increasing system complexity.
Solution Approach 2:
The invention introduces a selection mechanism as an intermediary between multiple machine learning predictors and the final tracking output. This mediator evaluates predictions from multiple sources and selects the most reliable one, allowing the system to leverage diverse predictors while managing complexity through structured selection rather than direct integration of all predictors.
4Productivity
If parallelization is used with machine learning predictors, then computation speed increases, but device complexity and resource requirements increase
Solution Approach 1:
The system divides the tracking computation into independent parallel segments, each handled by a separate machine learning predictor processing different aspects of the data. These segmented computations can execute simultaneously on multiple processors or cores, increasing computation speed while maintaining manageable complexity through modular design.
Solution Approach 2:
The invention implements dynamic parallelization where the degree of parallel execution adapts based on available computational resources and timing requirements. The system can dynamically adjust the number of active predictors and their execution modes, allowing efficient utilization of parallel resources without permanently increasing system complexity.
Data Source
AI summary
Tracking using sensor data is described, for example, where a plurality of machine learning predictors are used to predict a plurality of complementary, or diverse, parameter values of a process describing how the sensor data arises. In various examples a selector selects which of the predicted values are to be used, for example, to control a computing device. In some examples the tracked parameter values are pose of a moving camera or pose of an object moving in the field of view of a static camera; in some examples the tracked parameter values are of a 3D model of a hand or other articulated or deformable entity. The machine learning predictors have been trained in series, with training examples being reweighted after training an individual predictor, to favor training examples on which the set of predictors already trained performs poorly.