Appearance-Guided Tracklet Splitting and Merging for Multi-Object Tracking

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing multi-object tracking models often output imperfect tracklets, either containing multiple different objects within one tracklet or having one object in multiple tracklets, leading to reduced accuracy due to factors like occlusion and illumination changes.

Innovation Solution

A method involving split and merge operations on tracklets, including determining appearance features, clustering labels, and updating fragment and image block labels to correct tracklet boundaries, and merging tracklets based on cosine distance matrices to improve accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If tracklets are generated by multi-object tracking models, then tracking coverage is achieved, but tracking accuracy deteriorates due to false splits and merges

Engineering Contradiction:
Improvetracking accuracyVSAvoidtracking efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent segments tracklets into finer granular units based on appearance feature changes. By dividing tracklets at points where appearance features indicate potential object transitions, the system creates more precise tracking segments that reduce false merges while maintaining comprehensive coverage through systematic segmentation of the video sequence.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent merges tracklets that were incorrectly split by combining them back together. Through appearance feature analysis and trajectory consistency checking, the system identifies tracklets that should have remained connected and merges them, thereby reducing false splits and improving overall tracking accuracy without sacrificing the benefits of fine-grained segmentation.

Inventive Principle:
Principle #5Merging (Combining)

2Measurement precision

If tracklets are split to separate multiple objects, then object identification accuracy improves, but false splits occur when single objects are divided

Engineering Contradiction:
Improveobject identification accuracyVSAvoidfalse split rate
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent implements feedback mechanisms that continuously monitor tracklet characteristics during the splitting process. Appearance feature sequences and trajectory patterns are analyzed in real-time to determine whether a split is justified. When feedback indicates that a split would create false separations of a single object, the system adjusts or cancels the split, thereby reducing false split rates while maintaining high object identification accuracy.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent employs dynamic adjustment of tracklet boundaries based on changing appearance features and contextual information. Rather than using fixed splitting criteria, the system dynamically adapts split points based on real-time analysis of appearance feature sequences, allowing flexible response to complex scenarios where objects may temporarily resemble each other or where occlusion occurs.

Inventive Principle:
Principle #15Dynamics

3Manufacturing precision

If appearance features are used to distinguish objects, then tracking precision improves, but computational complexity increases

Engineering Contradiction:
Improvetracking precisionVSAvoidcomputational complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent applies local quality analysis by examining appearance features at specific locations and time points within tracklets rather than processing the entire tracklet uniformly. By identifying and focusing computational resources on critical regions where appearance changes occur, the system achieves high tracking precision while reducing overall computational complexity through targeted processing of only the most informative features.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent employs partial action by selectively processing appearance features based on their relevance and importance. Rather than analyzing all possible features exhaustively, the system identifies and processes only the most discriminative features that provide sufficient information for accurate tracking, thereby achieving adequate precision with reduced computational effort.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4451210B1Method, device, and storage medium for improving multi-object tracking
Publication Date: 2025.10.29 FUJITSU LTD
  • EP4451210B1 patent drawingFigure 1
  • EP4451210B1 patent drawingFigure 2
  • EP4451210B1 patent drawingFigure 3

AI summary

The present disclosure relates to a method, device and storage medium for improving multi-object tracking. According to an embodiment, the method comprises: performing a split operation on a tracklet provided for one object by a multi-object tracking model. The split operation comprises: determining an appearance feature sequence of the tracklet; determining a clustering label set of the appearance feature sequence; determining an image block label sequence; determining a fragment label sequence corresponding to continuous fragments, having the same clustering labels, in the image block label sequence; in a case where a length of the fragment label sequence is greater than the number of types of the clustering labels in the clustering label set, updating the image block label sequence and the fragment label sequence by performing an update operation; and splitting the tracklet based on the updated image block label sequence. The method may further comprise a merge operation.