Multi-Commodity Flow for Identity Preservation in Crowd Tracking

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing multiple target tracking methods, particularly in crowded scenes and team sports, face challenges in maintaining individual identities when paths intersect and appearance cues are sporadic, leading to identity switches and high computational complexity.

Innovation Solution

The method involves discretizing the area into a grid, partitioning tracked people into groups with separate appearance models, generating a probability occupancy map, and using a K-Shortest Path algorithm on a Direct Acyclic Graph to compute trajectories, followed by a layered tracklet-based multi-commodity flow problem to derive individual trajectories, incorporating sparse appearance information to prevent identity switches.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If tracklet-based methods match tracklets in successive batches of frames, then computational complexity is reduced, but tracking accuracy deteriorates due to sub-optimal local decisions

Engineering Contradiction:
Improvecomputational complexityVSAvoidtracking accuracy
Core Design Contradiction:
Device complexityVSReliability

Solution Approach 1:

The patent transitions from local batch processing to global trajectory optimization by formulating the problem as a Minimum Cost Maximum Flow (MCMF) problem on a spatio-temporal graph. This dimensional transformation allows simultaneous consideration of all tracklets across the entire video sequence, achieving global optimality while maintaining computational feasibility through graph theory techniques.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Solution Approach 2:

The patent introduces a flow network as an intermediary structure between detections and trajectories. By constructing a graph where nodes represent detections and edges represent possible associations, the MCMF algorithm can optimally route flows (tracklets) through the network, mediating between local detection results and global trajectory consistency.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If appearance information is exploited to prevent identity switches, then identity preservation improves, but computational complexity increases due to sparse and intermittent appearance data

Engineering Contradiction:
Improveidentity preservationVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent applies appearance information partially and selectively rather than continuously. By using appearance cues only when available and reliable (e.g., when detection confidence is high or appearance features are clearly observable), the system prevents identity switches at critical moments without incurring continuous computational overhead during periods when appearance data is sparse or unreliable.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The patent dynamically adjusts the weight of appearance information in the cost function based on detection quality and appearance reliability. When appearance data is sparse or uncertain, the system reduces its influence and relies more on motion consistency; when appearance is clear and reliable, it increases weight to strongly constrain identity assignment, thus adapting computational effort to data quality.

Inventive Principle:
Principle #35Parameter changes

3Reliability

If global optimization is applied to tracklet association, then tracking accuracy improves, but computational complexity becomes unmanageable for long video sequences

Engineering Contradiction:
Improvetracking accuracyVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent segments the video sequence into manageable chunks or batches while maintaining global optimization within each segment. By dividing the long sequence into overlapping or contiguous segments and applying MCMF locally to each, the system achieves near-global optimality without the computational burden of processing the entire sequence at once, while tracklet associations across segment boundaries ensure continuity.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP2780871B1Tracklet-based multi-commodity network flow for tracking multiple people
Publication Date: 2015.09.16 SWISS TIMING LTD
  • EP2780871B1 patent drawingFigure 1~2(b)
  • EP2780871B1 patent drawingFigure 3~5
  • EP2780871B1 patent drawingFigure 4a~4d

AI summary

A method for continuously tracking multiple people partitioned into groups while preserving identities under global appearance constraints, wherein people's trajectories may intersect, and wherein only sparse appearance information is available is disclosed. Individual trajectories for each group identity are obtained by solving a layered tracklet-based multi-commodity f low (MCNF) programming problem, wherein tracklets are connected parts of splitted trajectories, wherein each trajectory is split at posit ions which are in the neighborhood of another, wherein said neighborhood encompasses locations within a predefined distance.