Multi-object tracking with dynamic feature extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing multi-object tracking (MOT) techniques face challenges with noisy feature extraction, particularly in cases of overlapping adjacent objects, background objects, and ambiguous objects, which can lead to inefficient tracking and high computational requirements.
Innovation Solution
The proposed solution involves determining an adjacency overlap (AdO) score for detected objects in a frame based on overlap with other objects, and extracting features only if the AdO score is below a threshold, thereby improving feature extraction and tracking accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing MOT techniques extract features from all detected objects, then tracking coverage is maximized, but feature extraction quality deteriorates due to noise from overlapping objects and background
Solution Approach 1:
The patent applies local quality by differentiating the treatment of objects based on their spatial relationships. Objects with high adjacency overlap scores (indicating potential overlap with other objects or background) are excluded from feature extraction, while objects with low overlap scores are processed normally. This selective approach improves feature quality for critical objects without sacrificing overall tracking coverage.
Solution Approach 2:
The patent introduces an intermediary mechanism - the adjacency overlap score calculation - that mediates between the detected objects and the feature extraction process. By computing overlap scores with other detected objects and background regions, the system determines which objects should have their features extracted, effectively filtering out noisy objects before feature extraction occurs.
2Measurement precision
If heavy backbone networks are used for feature extraction, then tracking accuracy improves, but computational requirements increase
Solution Approach 1:
The patent applies partial action by selectively applying feature extraction only to a subset of objects (those with low adjacency overlap scores) rather than processing all detected objects. This reduces the total computational load while maintaining tracking accuracy for the most critical objects that are less likely to be obscured by overlaps.
Solution Approach 2:
The patent performs preliminary action by calculating adjacency overlap scores before executing the computationally intensive feature extraction process. This preliminary filtering step identifies which objects require full feature extraction and which can be skipped, optimizing the balance between computational resources and tracking accuracy.
3Quantity of substance
If features are extracted from overlapping objects, then complete object information is captured, but tracking reliability decreases due to noisy features
Solution Approach 1:
The patent converts the harmful effect of object overlap into a beneficial filtering criterion. Instead of treating overlap as a problem to be solved by complex processing, the system uses the overlap score as a direct indicator to identify objects that should be excluded from feature extraction. This transforms the overlap information from a source of noise into a useful filtering signal that improves tracking reliability.
Data Source
AI summary
Techniques and systems are provided for shape estimation. For instance, a process can include determining a first adjacency overlap (AdO) score of a first detected object in a first frame based on overlap between the first detected object and a second detected object in the first frame; and extracting features of the first detected object based on a comparison of the AdO score and a threshold AdO score.