Object Tracking Device Using Discriminator for Obstacle Handling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object tracking methods in radiation therapy face challenges in accurately specifying the position of a tracking object due to obstacles like bones and collimators, which can hide the target, and are invasive or slow, especially when dealing with real-time tracking and varying patient movements.
Innovation Solution
An object tracking device that creates superimposed images by overlaying non-tracking object images on tracking object images, learns image features and position information to create a discriminator, and uses this discriminator to specify the tracking object's position in real-time, even when obstacles are present, by separating and editing non-tracking object images based on size, resolution, and contrast.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If template matching is used to track the target object, then the tracking method is simple and fast, but the tracking accuracy deteriorates when the target object is hidden by obstacles like bones or collimators
Solution Approach 1:
The patent introduces a discriminator as an intermediary component that has been pre-trained to recognize target objects even when partially obscured. The discriminator acts as a mediator between the input image and the tracking output, enabling accurate target identification despite obstacles like bones or collimators that would normally interfere with simple template matching
Solution Approach 2:
The system performs preliminary training of the discriminator using pre-prepared images that include various obstacle conditions. By preparing the discriminator in advance with diverse training data showing target objects obscured by different types of obstacles, the system enables fast real-time tracking without requiring complex obstacle detection during the actual tracking process
2Measurement precision
If metal markers are inserted into the patient's body to increase tracking accuracy, then the tracking accuracy improves, but the method becomes invasive and places a burden on the patient
Solution Approach 1:
The patent creates a virtual copy of the target object's appearance and characteristics through the discriminator model. Instead of inserting physical markers into the patient's body, the system learns to recognize and track the target object based on its natural imaging characteristics from pre-prepared images, providing a non-invasive alternative that achieves similar tracking accuracy
Solution Approach 2:
The system replaces the mechanical approach of inserting physical markers with an information-processing approach using deep learning. The discriminator processes imaging data to identify and track the target object based on learned patterns, substituting the need for physical marker insertion with computational recognition methods
3Measurement precision
If bone suppression processing is performed to remove obstacles from images, then the tracking accuracy improves, but the processing time increases and real-time tracking becomes difficult
Solution Approach 1:
The system performs obstacle handling in advance during the training phase rather than during real-time tracking. The discriminator is trained on pre-prepared images that include various obstacle conditions, so that during actual tracking, the system can directly use the trained model for fast inference without requiring time-consuming obstacle removal processing
Solution Approach 2:
The discriminator learns to copy or replicate the ability to identify target objects despite obstacles by training on diverse images. Instead of physically removing obstacles from images during tracking, the system creates a virtual understanding of target appearance under various obstacle conditions, enabling fast tracking without actual image modification
4Productivity
If the template image is created to match the target object, then the tracking is simple and fast, but the tracking fails when the target object's appearance changes due to breathing or movement
Solution Approach 1:
The system transitions from a static template matching approach to a dynamic deep learning model. The discriminator is trained on multiple images capturing the target object in various states (different breathing phases, positions, and appearances), enabling the model to adapt to appearance changes while maintaining fast tracking performance during inference
Solution Approach 2:
The system prepares the discriminator in advance with diverse training images that capture various target appearances. By pre-training with images from different breathing phases and positions, the system builds robustness to appearance changes before actual tracking begins, enabling reliable tracking without requiring real-time template updates
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An object tracking device includes a superimposed image creation unit (C6) configured to create a plurality of superimposed images (26) in which each of a plurality of non-tracking object images which do not include a tracking object image feature is superimposed on a tracking object section image (23) which includes a tracking object image feature (21); a discriminator creation unit (C9) configured to learn at least one of an image feature and position information of the tracking object, based on the plurality of superimposed images (26) to create a discriminator; and a tracking object specifying unit (C12) configured to specify at least one of the image feature and the position information of the tracking object in a tracked image (28) including the respective image features of the tracking object and an obstacle, based on the discriminator and the tracked image (28). Thereby, it easy to specify a tracking object site without being affected by the obstacle, as compared to an image tracking method by template matching of the prior art.