Across-camera object tracking using a conditional random field model

To ensure safety, most public spaces now deploy monitoring systems. However, in most scenarios, the tracking operations of these monitoring systems are performed manually. These operations should be automated. This paper proposes using a conditional random field (CRF) to formulate the automatic exec...

Full description

Saved in:
Bibliographic Details
Published in:The Journal of supercomputing 2021-12, Vol.77 (12), p.14252-14279
Main Authors: Cheng, Sheng-Tzong, Hsu, Chih-Wei, Horng, Gwo-Jiun, Chen, Sz-Yu
Format: Article
Language:English
Subjects:
Algorithms
Artificial neural networks
Cameras
Compilers
Computer Science
Conditional random fields
Interpreters
Minimum cost
Monitoring
Monitoring systems
Normal distribution
Object recognition
Pedestrian crossings
Pedestrians
Processor Architectures
Programming Languages
Psychology
Public spaces
Tracking
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:To ensure safety, most public spaces now deploy monitoring systems. However, in most scenarios, the tracking operations of these monitoring systems are performed manually. These operations should be automated. This paper proposes using a conditional random field (CRF) to formulate the automatic execution problem as a cost minimization problem. The appearance of pedestrians and the time taken by them to cross the view of a camera are used to solve the automatic execution problem. Crowd psychology is used to define constraints and construct a CRF graph. A Siamese convolutional neural network is employed to recognize pedestrian appearance. The time spent by pedestrians crossing the view of a camera is modeled using a normal distribution. The results of two models are considered as the costs of nodes and edges. The proposed algorithm is applied under constraints to determine matches at the minimum cost. The accuracy of the proposed method is compared with that of other methods by using common datasets and benchmarks. Superior results are obtained when both appearance and spatiotemporal information are employed for solving the automatic execution problem than when using appearance alone.
ISSN:0920-8542
1573-0484
DOI:10.1007/s11227-021-03862-w