Object Tracking via Feature and Color Score Map Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object tracking technologies using neural networks trained offline face challenges with low accuracy when encountering unlearned targets due to the lack of current target characteristics, leading to incorrect tracking of similar features or objects, especially when objects deform or occlude.
Innovation Solution
A processor-implemented method that acquires feature and color histograms from template images, generates feature and color score maps from search images, and combines these to determine a final score map and bounding box, using color weights to enhance discrimination between foreground and background, thereby improving tracking accuracy by integrating feature and color information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If only feature information from neural network is used for tracking, then the system complexity is low, but tracking accuracy deteriorates when encountering unlearned targets or deformed objects
Solution Approach 1:
The patent combines feature information from neural network with color histogram information to create a composite tracking system. The feature score map from neural network and color score map from color histograms are merged through weighted addition to produce a final score map, thereby improving tracking accuracy while maintaining reasonable system complexity through the use of established algorithms
2Measurement precision
If color information is always used for tracking, then discrimination between foreground and background is improved, but computational load and processing time increase
Solution Approach 1:
The patent dynamically adjusts the contribution of color information based on the calculated color weight, which is determined by the ratio of foreground to background pixel values. When color information is highly discriminative, it is given higher weight; when less useful, its weight is reduced. This dynamic adjustment optimizes both discrimination accuracy and processing efficiency in real-time
Solution Approach 2:
The patent changes the parameter of color weight dynamically based on the current frame's color distribution characteristics. By calculating the ratio of foreground to background pixels and adjusting the color weight accordingly, the system adapts to varying scene conditions, improving discrimination when color is useful and reducing computational overhead when it is not
3Adaptability or versatility
If neural network is trained offline with fixed data, then training time is reduced, but adaptability to new targets and scenarios deteriorates
Solution Approach 1:
The patent introduces color histograms as an intermediary mechanism that provides additional discriminative information without requiring complex online retraining of the neural network. The color histogram calculator processes color information independently and merges it with feature information, enabling the system to adapt to new targets and scenarios in real-time without modifying the pre-trained neural network
Data Source
AI summary
A method, apparatus, and system with tracking are disclosed. The apparatus is configured to acquire a feature map of a template image, a color histogram of a foreground of the template image that has an object, and a color histogram of a background of the template image other than the foreground, acquire a feature score map, and a bounding box map corresponding to the feature score map, based on the feature map of the template image and a feature map of a search image, acquire a color score map based on the color histogram of the foreground, the color histogram of the background, and a color value of the search image, acquire a final score map, and a bounding box map corresponding to the final score map, based on the feature score map, the color score map, and the bounding box map, and output the corresponding bounding box.