Visual Object Tracking via Predicted and Detected Bounding Box Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing visual object tracking techniques, such as those based on Intersection-over-Union (IoU), rely on high-accuracy object detectors, which may not be feasible due to computational constraints or cost limitations, leading to potential tracking accuracy deterioration.
Innovation Solution
A visual object tracking method that estimates bounding boxes and object IDs by combining predicted and detected bounding boxes, using a machine learning evaluation model to generate confidence vectors and association weights, and adjusting model parameters based on a deviation index to improve tracking accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a light-weight object detector is used to reduce computational requirements, then device complexity and computational cost are reduced, but detection accuracy deteriorates leading to tracking accuracy deterioration
Solution Approach 1:
The patent combines predicted bounding boxes (from tracking) and detected bounding boxes (from detection) into a unified set of estimated bounding boxes. This merging allows the system to leverage both the computational efficiency of light-weight detectors and the accuracy benefits of prediction, resolving the contradiction between device complexity and measurement precision
Solution Approach 2:
The patent introduces an evaluation model as an intermediary that processes both predicted and detected bounding boxes. This intermediary component generates confidence scores and association weights, enabling the system to reconcile the outputs from light-weight detection with tracking predictions, thereby maintaining detection accuracy while using computationally efficient detectors
2Productivity
If a light-weight object detector is used to reduce computational requirements, then processing speed is improved, but tracking accuracy deteriorates
Solution Approach 1:
The system merges predicted bounding boxes from tracking (which maintain accuracy over time) with detected bounding boxes from light-weight detectors (which provide fast processing). This combination allows the system to achieve both high processing speed and reliable tracking accuracy by leveraging the strengths of both approaches
Solution Approach 2:
The evaluation model provides feedback by generating confidence scores and association weights that indicate the reliability of each bounding box. This feedback mechanism allows the tracking system to adjust its reliance on detection versus prediction based on current confidence levels, maintaining tracking accuracy while utilizing fast light-weight detectors
Data Source
AI summary
An estimation unit of a visual object tracking apparatus estimates a plurality of estimated bounding boxes and estimated object IDs respectively corresponding to the estimated bounding boxes based on a plurality of predicted bounding boxes and a plurality of detected bounding boxes. For example, the “detected bounding boxes” are “bounding boxes (bounded areas)” detected by a detector in each of a plurality of frames in a time series such as moving images. The “bounding box” is a frame surrounding an image of an object detected in a frame. For example, the “predicted bounding box” is a “bounding box” predicted by a predictor based on an estimated bounding box(es) estimated for one or a plurality of frames in the past.