Multi-Camera Target Tracking With Adaptive Re-Identification Thresholds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing multi-camera tracking systems face significant performance drops in real-world scenarios due to 'domain drift', leading to high false positives and negatives, necessitating excessive human operator interaction.
Innovation Solution
Adjust the human threshold inversely proportional to the number of reference tracklets, using a larger number of tracklets to reduce operator interaction and improve tracking performance by minimizing false positives and negatives.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning models are used for multi-camera tracking, then object recognition capability is improved, but false positives and false negatives increase due to domain drift
Solution Approach 1:
The patent dynamically adjusts the human threshold parameter based on the number of reference tracklets available. When few reference tracklets are available, a higher threshold is used to reduce false positives. As more reference tracklets accumulate, the threshold is lowered to improve detection sensitivity. This adaptive parameter adjustment resolves the contradiction by optimizing the trade-off between false positives and false negatives based on available data.
2Measurement precision
If a low human threshold is used to reduce false positives, then tracking precision is improved, but operator interaction increases
Solution Approach 1:
The patent implements a dynamic threshold adjustment mechanism where the human threshold is not fixed but adapts based on the number of reference tracklets. The system transitions from a static to a dynamic thresholding approach, automatically optimizing the balance between precision and operator workload as more tracking data becomes available.
Solution Approach 2:
The system performs self-optimization by automatically adjusting the human threshold based on accumulated reference tracklets without requiring manual operator intervention. The tracking system serves itself by learning from accumulated data and autonomously optimizing its operational parameters.
3Ease of operation
If a high human threshold is used to reduce operator interaction, then ease of operation is improved, but false positives increase
Solution Approach 1:
The patent dynamically adjusts the human threshold parameter based on the number of reference tracklets available. When few reference tracklets are available, a higher threshold is used to reduce false positives. As more reference tracklets accumulate, the threshold is lowered to improve detection sensitivity. This adaptive parameter adjustment resolves the contradiction by optimizing the trade-off between false positives and false negatives based on available data.
4Reliability
If the number of reference tracklets is increased to improve tracking performance, then false positives and false negatives are reduced, but system complexity increases
Solution Approach 1:
The patent accumulates reference tracklets in advance before formal tracking begins. By gathering and preprocessing reference data beforehand, the system builds a robust foundation for tracking without adding complexity during the actual tracking operation. This preliminary data collection phase separates the complexity of data gathering from the simplicity of tracking execution.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
The invention relates to a method (100) for real-time, multi-camera tracking of a target object comprising the following steps carried out for at least one new tracklet captured by at least one camera: - calculation (106) of at least one first distance between said new tracklet and at least one reference tracklet belonging to said target object; - when the smallest first calculated distance is between a first predetermined threshold value (SR), called the re-identification value, and a second predetermined threshold value (SH), called the human threshold, presentation (114) of said new tracklet to said operator, for assignment or not of said new tracklet to said target object, manually, by said operator; characterized in that the human threshold (SH) is a function of the number of reference tracklets in said reference list.It also relates to a computer program, an object re-identification device and an object re-identification system implementing such a process.