Intoxicated person identification using Markov chains and neural networks

• Published in: Neural computing & applications 2021-04, Vol.33 (8), p.3459-3467
• Main Author: Koukiou, Georgia
• Format: Article
• Language: English
• Subjects: Artificial Intelligence; Blood vessels; Clusters; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer Science; Data Mining and Knowledge Discovery; Eigenvalues; Forehead; Image Processing and Computer Vision; Intoxication; Markov analysis; Markov chains; Mathematical analysis; Matrix methods; Neural networks; Original Article; Pixels; Probability and Statistics in Computer Science; Subspaces; Temperature distribution
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-020-05219-5

In this work, Markov chains are used to model the statistical behavior of the pixels on the image of the forehead of a person in order to detect intoxication. It is the first time that second-order statistics are used for this purpose. The images were obtained in the thermal infrared region. Intoxication affects blood vessels activity and thus the temperature distribution on the face having a significant effect on the corresponding pixels statistics. The pixels of the forehead images are quantized to 32 Gy levels so that Markov chain models are structured using 32 states. The feature vectors used are the eigenvalues obtained from the first-order transition matrices of the Markov chain models. Since for each person a frame sequence of 50 views is acquired, a cluster of 50 vectors is formed in the 32-dimensional feature space. The feature space is firstly analyzed using projections of the clusters in 3D subspaces of the original 32D feature space. After that, the capability of a simple feed forward neural network to separate the clusters belonging to sober persons from those corresponding to intoxicated persons is investigated. A simple three-layer neural structure has a 98% vector separability success and a 100% cluster separability if the majority voting is considered. Furthermore, the classification problem is faced by excluding from the training procedure either one or five persons and using them in the testing phase. Accordingly, a neural network is trained using all except the excluded data. The obtained neural structure tested with the features of the persons in which it was not trained presents high drunk identification success if the majority voting is considered.