Real-time Object Tracking via Compound Similarity Metrics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Video analytics systems face challenges in accurately tracking objects due to occlusion events, leading to incorrect identification of re-emerging targets as new objects, which decreases accuracy and increases resource consumption.
Innovation Solution
A technique that associates lost objects with newly detected objects in real-time by estimating future motion and comparing visual appearance descriptors, using a compound similarity metric to determine re-association, and employing a visual appearance descriptor pool for efficient storage and computation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If object detection models are used to detect and track objects in video analytics systems, then object tracking capability is provided, but accuracy decreases when occlusion events occur causing lost targets to be incorrectly identified as new objects
Solution Approach 1:
The system implements feedback mechanisms by continuously comparing compound similarity metrics (combining motion similarity and visual appearance similarity) between newly detected objects and previously lost targets. This feedback loop enables the system to correctly re-associate lost targets with their original tracking IDs, preventing incorrect identification as new objects and thereby maintaining both tracking reliability and identification precision during occlusion events.
Solution Approach 2:
The patent combines multiple tracking approaches into a composite solution: it integrates motion-based tracking (using tracking trajectories and motion similarity metrics) with appearance-based tracking (using visual appearance descriptors and image similarity metrics). This composite approach creates a more robust object tracking system that maintains accuracy during occlusion events by leveraging the strengths of both motion and appearance information.
2Device complexity
If traditional object tracking methods are used without re-association techniques, then system complexity is reduced, but resource consumption increases due to incorrect tracking of lost targets
Solution Approach 1:
The system performs preliminary actions by maintaining a loss history buffer that stores information about recently lost targets. When new objects are detected, the system proactively compares them against the loss history using compound similarity metrics before final identification is made. This preliminary check prevents incorrect tracking from occurring in the first place, reducing the need for costly post-processing corrections and lowering overall computational resource consumption.
Solution Approach 2:
The patent dynamically adjusts tracking parameters based on detection confidence scores and occlusion detection results. When occlusion is detected or confidence is low, the system activates re-association mode with adjusted similarity thresholds and enables comparison against loss history. This parameter adaptation allows the system to maintain simplicity during normal operation while selectively increasing complexity only when needed to correct potential tracking errors, optimizing resource usage.
3Speed
If real-time object tracking is implemented without re-association of lost targets, then processing speed is maintained, but post-processing time increases due to incorrect tracking corrections
Solution Approach 1:
The system ensures continuity of useful action by performing re-association checks as a continuous part of the real-time detection pipeline rather than as a separate post-processing step. The loss history buffer is continuously updated, and similarity comparisons are continuously performed for each new detection. This integration eliminates the need for separate post-processing correction passes, maintaining real-time processing speed while preventing time loss that would otherwise be required for tracking corrections.
Data Source
AI summary
A first visual appearance descriptor associated with a first object in an environment is obtained based on a first set of images of a first time period. The first object is subsequently absent from the environment in a second set of images of a second time period. A second visual appearance descriptor associated with a second object is obtained based on a third set of images, of a third time period subsequent to the second time period. A compound similarity metric between the first and second objects is obtained in view of visual appearance similarity and motion similarity metrics. The visual appearance similarity metric corresponds to a degree of similarity between the first and second visual appearance descriptors. An identifier associated with the second object is updated to correspond to an identifier associated with the first object in response to determining that the compound similarity metric meets a threshold value.