Object Normalization via High-Order Model Fitting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computer vision techniques fail to reliably track objects across image sequences from different cameras and viewpoints due to variations in object properties like size and orientation, requiring manual calibration and lacking an automated normalization procedure.
Innovation Solution
The method involves determining normalization parameters by fitting a high-order model, such as a least squares fit of a second-order polynomial, to classification results from multiple viewpoints, allowing for the computation of normalized features and training data that are independent of camera viewpoint.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual calibration procedures are used to normalize objects across different viewpoints, then measurement precision can be achieved, but device complexity and time consumption increase significantly
Solution Approach 1:
The system performs self-calibration by automatically determining normalization parameters through high-order model fitting to classification results from multiple viewpoints, eliminating the need for manual calibration procedures while achieving accurate object property measurement across different camera viewpoints
Solution Approach 2:
The patent transforms the calibration approach by changing from fixed manual parameter setting to dynamic parameter determination through mathematical modeling, where normalization parameters are automatically computed by fitting high-order models to observed classification data across multiple viewpoints
2Measurement precision
If manual calibration procedures are used for object normalization, then measurement precision improves, but productivity decreases due to time-consuming calibration
Solution Approach 1:
The system performs preliminary automated calibration by pre-determining normalization parameters through high-order model fitting during system initialization or training phase, so that subsequent object classification and measurement operations can proceed rapidly without repeated manual calibration steps
Solution Approach 2:
The patent replaces manual mechanical calibration procedures with automated computational methods, using high-order mathematical models and least squares fitting algorithms to automatically determine normalization parameters, thereby eliminating time-consuming manual operations while maintaining measurement precision
3Reliability
If existing computer vision techniques are used for tracking across same camera, then tracking reliability is maintained, but adaptability to different cameras and viewpoints is limited
Solution Approach 1:
The patent creates a universal normalization framework that works across multiple cameras and viewpoints by determining viewpoint-specific normalization parameters through high-order model fitting, enabling the same object classification system to reliably track objects across different camera configurations without requiring viewpoint-specific manual calibration
Solution Approach 2:
The system introduces dynamic adaptability by automatically adjusting normalization parameters based on the specific viewpoint and camera characteristics, using high-order model fitting to adapt to different geometric configurations while maintaining consistent object property measurement across varying viewing conditions
Data Source
AI summary
Methods and apparatus are provided for normalizing objects across a plurality of image viewpoints. A set of classification results are obtained for a given object class across a sequence of images for each of a plurality of viewpoints. The classification results are each comprised of a position of one of the objects in the image, and at least one projected property of the object at that position. Normalization parameters are then determined for each of the viewpoints by fitting a high order model to the classification results to model a change in the projected property. The high order model may implement a least squares fit of a second order polynomial to the classification results. The normalization parameters may be used to compute normalized features and normalized training data for object classification.