Visual Target Tracking Using Inference Tensors for Similar Objects
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing tracking techniques using deep neural networks are prone to erroneous tracking when similar objects are present, as they calculate similarities independently for each feature subset, leading to incorrect tracking.
Innovation Solution
An information processing apparatus and method that utilizes first and second computation units to obtain features of a tracking target and search region, respectively, and generates an inference tensor and map to accurately determine the tracking target's position, using a fully convolutional neural network with trained weight parameters to reduce erroneous tracking.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If template matching or deep neural network techniques are used to track a target, then tracking precision is improved, but erroneous tracking occurs when similar objects are present due to high cross-correlation values
Solution Approach 1:
The patent segments the feature extraction process into multiple independent branches (e.g., shape features, color features, texture features) that process different aspects of the target independently. Each branch extracts specific types of features and passes them to separate computation units, preventing similar objects with different feature profiles from being erroneously tracked.
Solution Approach 2:
The patent introduces a new dimensional approach by computing tracking scores not just based on single feature similarities, but by combining multiple feature dimensions (shape, color, texture, etc.) through a weighted summation mechanism. This multi-dimensional feature space allows the system to distinguish between targets that may be similar in one aspect but differ in others.
2Productivity
If similarities are calculated independently for each feature subset, then computational efficiency is improved, but erroneous tracking occurs when objects are similar in some aspects
Solution Approach 1:
The patent merges the results from multiple independent feature computation units by summing their respective tracking scores. Each computation unit independently processes its specific feature type and generates a score, then these scores are combined through weighted addition to produce a final tracking determination. This merging approach maintains the computational efficiency of independent processing while achieving accurate tracking through combined multi-feature analysis.
Data Source
AI summary
An information processing apparatus comprises a first computation unit configured to obtain first features of an image of a tracking target, a second computation unit configured to obtain second features of an image of a search region, a third computation unit configured to obtain an inference tensor representing likelihoods that the tracking target is present at respective positions of the image of the search region, using the first features and the second features, and a fourth computation unit configured to obtain an inference map representing a position of the tracking target in the image of the search region, using the inference tensor.