Object Tracking Feature Learning for Occlusion and Look-Alike Targets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object tracking methods using Deep Neural Networks fail to accurately distinguish between a tracking target object and similar-looking objects, especially under conditions of attitude changes or occlusion.
Innovation Solution
An image processing apparatus and method that calculates a loss function based on feature distance and overlap rates between tracking and non-tracking objects, using a learning model to enhance object discrimination and tracking precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a Deep Neural Network is used for object tracking, then the tracking speed and automation are improved, but the accuracy of distinguishing between similar-looking objects deteriorates
Solution Approach 1:
The patent changes the parameters used for object discrimination by incorporating not only feature amounts from Deep Neural Networks but also template matching scores and luminance/color information. By multipling these different parameter types and combining them in a comprehensive loss function, the system achieves both high tracking speed and accurate discrimination between similar-looking objects.
Solution Approach 2:
The patent creates a composite tracking system that combines multiple different tracking approaches (Deep Neural Network feature amounts, template matching results, and luminance/color information) into a unified tracking determination. This composite approach leverages the strengths of each method while mitigating their individual weaknesses, particularly in distinguishing similar-looking objects.
2Ease of operation
If feature amounts from Deep Neural Network are used for tracking, then the ease of operation is improved, but the reliability of tracking under occlusion or attitude changes deteriorates
Solution Approach 1:
The patent performs preliminary action by calculating template matching scores and preparing multiple feature representations (luminance, color, and Deep Neural Network features) in advance. These pre-computed values are then integrated in the loss function to reliably determine tracking even when occlusion or attitude changes occur, maintaining both automation and reliability.
Solution Approach 2:
The patent introduces template matching scores and luminance/color information as intermediary elements that bridge the gap between Deep Neural Network feature amounts and reliable tracking determination. These intermediaries provide additional discriminative power that helps maintain tracking reliability under challenging conditions while preserving the automation benefits of Deep Learning.
3Device complexity
If only Deep Neural Network feature amounts are used, then the device complexity is reduced, but the measurement precision of object discrimination deteriorates
Solution Approach 1:
The patent merges multiple different tracking approaches (Deep Neural Network feature extraction, template matching, and luminance/color analysis) into a unified tracking determination system. By combining these methods and integrating their results through a comprehensive loss function that multiplies and sums their respective contributions, the system achieves high object discrimination precision without requiring a single overly complex device structure.
Data Source
AI summary
An image processing apparatus including a non-transitory computer-readable storage medium storing instructions, and a processor executing the instructions to calculate a loss based on a distance between features of tracking target objects in respective images, a distance between features of a tracking target object and a non-tracking target object in an image, and an overlapping rate between an image region of a tracking target object and an image region of a non-tracking target object in one or more images, to perform learning of a model for extracting, based on the loss, a feature amount of an object from an image, and to assign an ID to a second object according to a distance between a first feature amount, based on the completed learning model, of a first object assigned the ID in a first image, and a second feature amount, based on the learning model, of a second image object.