Adaptive multi-object tracking algorithm based on split trajectory

Multi-object tracking (MOT) has wide-ranging applications in unmanned vehicles, military reconnaissance, and video surveillance. However, real-world scenes present significant challenges to the tracking process. Objects often occlude each other, leading to inaccuracies in position determination. As...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:The Journal of supercomputing Ročník 80; číslo 15; s. 22287 - 22314
Hlavní autoři: Sun, Lifan, Li, Bingyu, Gao, Dan, Fan, Bo
Médium: Journal Article
Jazyk:angličtina
Vydáno: New York Springer US 01.10.2024
Springer Nature B.V
Témata:
Accuracy
Adaptive algorithms
Algorithms
Associations
Compilers
Computer Science
Computer vision
Confidence
Deep learning
Interpreters
Multiple target tracking
Object recognition
Occlusion
Processor Architectures
Programming Languages
Real time
Surveillance
Unmanned vehicles
Multi-object tracking
Lost trajectory
Split trajectory
Appearance features updated
ISSN:0920-8542, 1573-0484
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Multi-object tracking (MOT) has wide-ranging applications in unmanned vehicles, military reconnaissance, and video surveillance. However, real-world scenes present significant challenges to the tracking process. Objects often occlude each other, leading to inaccuracies in position determination. As a result, issues like object loss, trajectory discontinuity, and erroneous associations occur, negatively impacting tracking accuracy and robustness. In the field of MOT, successfully associating objects when they reappear is a vital challenge. To tackle these challenges, this research introduces an adaptive MOT algorithm based on sub-trajectories. We propose the concept of trajectory segmentation, which involves distinguishing between lost trajectories and active trajectories for association. This approach allows for the adoption of the most suitable association strategy depending on different trajectories. Regarding the association of lost trajectories, we introduce an adaptive weight updating module that assigns distinct association reference weights based on the degree of the trajectory occlusion. This improvement enhances the resilience of object tracking during extended occlusion periods. Additionally, we present a novel strategy for updating appearance features to mitigate the impact of occlusion on object appearance, thereby enhancing overall tracking performance. We conducted comparative experiments using the MOT15 and MOT17 datasets, demonstrating that our approach improves tracking accuracy and meets real-time requirements when compared to existing algorithms.
Bibliografie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0920-8542
1573-0484
DOI:10.1007/s11227-024-06285-5