Object Tracking via Dual Tracker Overlap Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object tracking methods face challenges in adapting to sudden changes in object appearance, requiring significant computational resources and often necessitating human intervention, especially when using multiple trackers, which can drift away from the target and require complex fusion methods.
Innovation Solution
A method and system that compute overlap values between bounding boxes from two trackers to update and maintain an object model, using scores based on motion consistency and visual similarity to determine the correct tracker and adjust bounding boxes, allowing for adaptive tracking without human intervention.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single tracker is used with quick adaptation to handle appearance variations, then tracking reliability improves, but computational resources increase significantly
Solution Approach 1:
The system divides the tracking task into two independent trackers with different characteristics. Tracker 1 uses a generic object model for robustness, while Tracker 2 uses a learned appearance model for adaptability. This segmentation allows each tracker to operate with moderate computational requirements rather than one highly complex tracker, resolving the contradiction between reliability and computational resources.
2Reliability
If multiple trackers are combined to improve tracking robustness, then tracking reliability improves, but device complexity increases
Solution Approach 1:
Instead of having trackers independently process the same data and then fusing results (complex approach), the system inverts the workflow: Tracker 1 generates a generic model, Tracker 2 learns from it to create a specialized appearance model, and they alternately update each other's models. This inverted approach simplifies the system architecture while maintaining multiple-trackers robustness.
3Reliability
If trackers adapt quickly to appearance changes, then tracking reliability improves, but manufacturing precision of tracking accuracy deteriorates due to drift
Solution Approach 1:
The system implements bidirectional feedback between trackers: Tracker 2's learned appearance model feedback updates Tracker 1's generic model when confidence is high, and Tracker 1's stable generic model provides correction feedback to Tracker 2 when drift is detected. This feedback mechanism allows quick adaptation while preventing precision deterioration through mutual correction.
4Reliability
If existent multiple tracker combination methods are used, then tracking reliability improves, but ease of operation deteriorates due to requiring human intervention
Solution Approach 1:
The system implements self-service automation where the trackers autonomously monitor each other's confidence levels and automatically switch between them based on performance. The alternation mechanism and model updating occur without human intervention, with the system self-regulating based on confidence thresholds and drift detection, thereby improving ease of operation while maintaining reliability.
5Reliability
If existent multiple tracker fusion methods are used, then tracking reliability improves, but device complexity increases due to requiring confidence maps
Solution Approach 1:
The system extracts only the essential confidence information from each tracker's output without requiring full confidence maps. By taking out just the critical confidence metric and using it for simple threshold-based switching and model updating, the system achieves multi-tracker robustness while dramatically reducing complexity compared to methods requiring complete confidence map fusion.
Data Source
AI summary
Various aspects of a method and system for object tracking are disclosed herein. The method comprises computation of an overlap value between a first bounding box and a second bounding box, associated with an object in an image frame of a video stream. The overlap values are utilized to track the object in the video stream. The first bounding box encompasses at least a partial first portion of an object. The second bounding box encompasses at least a partial second portion of the object. A first score associated with the first bounding box and a second score associated with the second bounding box is computed based on a comparison of the computed overlap value with a pre-defined threshold value. The first bounding box is updated to a new bounding box that at least partially encompasses the object, based on the computed first score and the computed second score.