Adaptive Object Recognition Model Switching for Tracking Resumption
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object recognition systems face challenges in accurately detecting objects when tracking is paused and resumed in a scene with a different background, due to biased learning towards background patterns, leading to erroneous detection.
Innovation Solution
An object recognition apparatus and method that employs two recognition models: a first model with a high-weight negative example for known object regions and a second model with a low-weight or zero negative example for unknown regions, allowing for adaptive learning and accurate detection in changing backgrounds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If tracking is paused and resumed in a scene with a different background, then object detection should be performed, but erroneous detection occurs due to biased learning towards background patterns
Solution Approach 1:
The patent applies dynamics by making the recognition model adaptive and adjustable based on the tracking state. When tracking is paused and resumed in a different scene, the system dynamically switches between using the first recognition model (with background negative examples) for known object regions and the second recognition model (without background negative examples) for unknown regions, allowing the system to adapt to different background conditions and avoid erroneous detection.
Solution Approach 2:
The patent changes the parameter of recognition model selection based on the presence or absence of previous object region information. By switching between two different recognition models with different negative example weights, the system adjusts its learning parameters to accommodate different background scenarios, thereby maintaining high detection accuracy across varying environmental conditions.
2Measurement precision
If a first recognition model with high-weight negative example is used for known object regions, then detection precision is improved, but the system becomes less adaptable to new backgrounds after tracking pause
Solution Approach 1:
The patent segments the recognition task into two distinct models: the first recognition model for known object regions with high-weight negative examples for precision, and the second recognition model for unknown regions with low or zero negative example weights for adaptability. This segmentation allows each model to be optimized for its specific purpose, resolving the contradiction between precision and adaptability.
Solution Approach 2:
The patent applies local quality by using different recognition models for different spatial regions based on whether they are known or unknown. Known object regions use the first model with strong background discrimination, while unknown regions use the second model that is more flexible and less biased toward previous backgrounds, allowing each area to have the appropriate recognition characteristics.
3Reliability
If online learning is performed using background patterns as negative examples, then tracking accuracy is improved, but learning becomes biased toward background patterns when tracking is paused
Solution Approach 1:
The patent prepares two different recognition models in advance: the first model for when object region information is available (with background negative examples for accurate tracking), and the second model for when tracking is paused (without background negative examples to avoid bias). This preliminary preparation allows the system to select the appropriate model before the actual recognition occurs, preventing learning bias while maintaining tracking accuracy.
Data Source
AI summary
An object recognition apparatus comprises: an extraction unit configured to extract a partial region from an image and extract a feature amount; a recognition unit configured to recognize whether the partial region is a target object based on the feature amount and one of a first recognition model including a feature amount of a positive example indicating the target object and a negative example indicating a background and a second recognition model including that of the positive example; an updating unit configured to update the first recognition model by adding the feature amount; and an output unit configured to output an object region recognized as being the target object, wherein the recognition unit performs recognition based on the first recognition model if the object region was output for a previous image, and based on the second recognition model if not.