A New Sparse Low-rank Matrix Decomposition Method and its Application on Train Passenger Abnormal Action Identification

Title:

A New Sparse Low-rank Matrix Decomposition Method and its Application on Train Passenger Abnormal Action Identification

Creator:

Zhang , Y., Cai , G., Sun, J., Wang, Y., and Chen , J.

Identifier:

https://cdk.lib.cas.cz/client/handle/uuid:b01491eb-5695-4342-bbf8-b19cd46a6ff4
uuid:b01491eb-5695-4342-bbf8-b19cd46a6ff4
doi:10.14311/NNW.2015.25.033

Subject:

sparse, low-rank matrix, alternative algorithms, SCAD, and abnormal action identification

Type:

model:article and TEXT

Format:

bez média and svazek

Description:

In the article a new sparse low-rank matrix decomposition model is proposed based on the smoothly clipped absolute deviation (SCAD) penalty. In order to overcome the computational hurdle we generalize the alternating direction method of multipliers (ADMM) algorithm to develop an alternative algorithm to solve the model. The algorithm we designed alternatively renew the sparse matrix and low-rank matrix in terms of the closed form of SCAD penalty. Thus, the algorithm reduces the computational complexity while at the same time to keep the computational accuracy. A series of simulations have been designed to demonstrate the performances of the algorithm with comparing with the Augmented Lagrange Multiplier (ALM) algorithm. Ultimately, we apply the model to an on- board video background modeling problem. According to model the on-board video background, we can separate the video background and passenger's actions. Thus, the model can help us to identify the abnormal action of train passengers. The experiments show the background matrix we estimated is not only sparser, but the computational efficiency is also improved.

Language:

English

Rights:

http://creativecommons.org/publicdomain/mark/1.0/
policy:public

Source:

Neural network world: international journal on neural and mass-parallel computing and information systems | 2015 Volume:25 | Number:6

Harvested from:

CDK

Metadata only:

false