Purish S. Machine learning methods for human gait recognition

A study of the most significant works developed in recent years on gait recognition, highlighting methods based on visual features, was conducted. The study allowed us to identify the problems and challenges facing researchers in this area. The analysis also revealed the advantages of gait-based systems compared to other systems that use common biometric characteristics. The problem of the presence of gait covariants, which significantly affect the effectiveness of gait recognition methods and the practical feasibility of developing biometric identification systems based on gait recognition, is analyzed. To consider in detail the covariates that have the greatest impact on the efficiency of gait recognition and the proposed methods that are robust to the influence of these covariates. To reduce the time spent on human recognition using the HoG feature descriptor, a method was proposed that uses the popular HoG feature descriptor and Haralick texture features with sum variance to reduce the number of HoG features. It is shown that the Haralick texture features can be used together with HoG to extract the most characteristic low-dimensional feature vectors that can uniquely identify people. As a result of the study, it was determined that the proposed method is, in most cases, more effective than analogous methods. For example, on the CASIA A dataset, the proposed method showed an average recognition accuracy of 97.77%, which exceeds the performance of the analogous methods. On the CASIA B dataset, the method showed the best accuracy for almost all camera viewing angles (the advantage over the closest analog ranged from 3.8% to 6.4%, depending on the viewing angle). For the CMU MoBo dataset, the proposed method showed 82.00% accuracy in human gait recognition, which exceeds the performance of the nearest analog method by 5%. To recognize a person by gait under the conditions of wearing different clothes, a method for recognizing a person by gait based on the local descriptor LR2P was proposed. The proposed method uses the gait energy image and exploits the local distinctive features of regions in the GEI, which improves the gait recognition performance. To confirm the effectiveness of the proposed method, an experimental study was conducted on several datasets. The accuracy rate on the OU-ISIR B dataset was 86.22%, which exceeds the accuracy of the closest analog method by 4.30%. The average accuracy rate for all 32 clothing styles on the OU-ISIR B dataset was 91.78%, which indicates that the proposed method can minimize the problem of clothing variations. To recognize a person by gait in a frontal camera projection, a recognition method based on the analysis of the gait energy image contour was proposed, consisting of the following steps. An experimental study on the CMU MoBo dataset showed that the proposed method demonstrates a 6.41% improvement in recognition accuracy over the analog method. A study on the CASIA B dataset for zero viewing angle showed that the proposed method achieved an accuracy rate of 99.70%, which exceeds the accuracy of the closest analog method by 5.30%. To recognize a person by gait under conditions of different gait speeds, a recognition method based on the use of GEI spatial dynamics was proposed, consisting of the following steps. To confirm the effectiveness of the proposed method, an experimental study was conducted on several datasets. The study on the CASIA C dataset showed that the proposed method has an accuracy rate ranging from 0.5% to 6.1%, depending on the study parameters. The average accuracy on the OU-ISIR A dataset was 98.69%, which exceeds the accuracy of the closest analog method by 9.2%. A software module for recognizing people by their gait has been developed. The software module can be part of a reliable and efficient biometric identification system that uses a person's gait as a biometric identifier. This module can be used in various areas where it is necessary to perform authentication and access control procedures. It can be integrated into existing video surveillance systems and used to provide various security and identification processes. The main functional and non-functional requirements for the developed module are defined. A deployment diagram was formed and a data schema was provided. To determine the technologies used, a description of the technology stack of the program module was provided. The structure of the program module project is provided, the main software packages of the developed software are identified and described. In order to ensure the quality of the developed software module, functional and unit testing was carried out. The methods developed in the dissertation have been implemented in the activities of Dobrodiya Foods LLC, PJSC Vetropack Gostomel Glassworks, as well as in the educational process of the Odesa Polytechnic National University.