Recent improvements in pose estimation has opened up the possibility of new areas of application. One of them is gait recognition, the task of identifying persons based on their unique style of walking, which is increasingly being recognized as an important method of biometric indentification. This thesis has explored the possibilities of using a pose estimation system, OpenPose, together with deep Recurrent Neural Networks (RNNs) in order to see if there is sufficient information in sequences of 2D poses to use for gait recognition. For this to be possible, a new multi-camera dataset consisting of persons walking on a treadmill was gathered, dubbed the FOI dataset. The results show that this approach has some promise. It achieved an overall classification accuracy of 95,5 % on classes it had seen during training and 83,8 % for classes it had not seen during training. It was unable to recognize sequences from angles it had not seen during training, however. For that to be possible, more data pre-processing will likely be required.

Yes / Aims: To create a gait database with optimum accuracy of joint rotational data and an accu-rate representation of 3D volume, and explore the potential of using the database in studying the relationship between static and dynamic features of a human’s gait. Study Design: The study collected gait samples from 38 subjects, in which they were asked to walk, run, walk to run transition, and walk with a bag. The motion capture, video, and 3d measurement data extracted was used to analyse and build a correlation between features. Place and Duration of Study: The study was conducted in the University of Bradford. With the ethical approval from the University, 38 subjects’ motion and body volumes were recorded at the motion capture studio from May 2011- February 2013. Methodology: To date, the database includes 38 subjects (5 females, 33 males) conducting walk cycles with speed and load as covariants. A correlation analysis was conducted to ex-plore the potential of using the database to study the relationship between static and dynamic features. The volumes and surface area of body segments were used as static features. Phased-weighted magnitudes extracted through a Fourier transform of the rotation temporal data of the joints from the motion capture were used as dynamic features. The Pearson correlation coefficient is used to evaluate the relationship between the two sets of data. Results: A new database was created with 38 subjects conducting four forms of gait (walk, run, walk to run, and walking with a hand bag). Each subject recording included a total of 8 samples of each form of gait, and a 3D point cloud (representing the 3D volume of the subject). Using a Pvalue (P<.05) as a criterion for statistical significance, 386 pairs of features displayed a strong relationship. Conclusion: A novel database available to the scientific community has been created. The database can be used as an ideal benchmark to apply gait recognition techniques, and based on the correlation analysis, can offer a detailed perspective of the dynamics of gait and its relationship to volume. Further research in the relationship between static and dynamic features can contribute to the field of biomechanical analysis, use of biometrics in forensic applications, and 3D virtual walk simulation.

Biometrics is a unique physical or behavioral characteristic of a person. This unique attribute, such as fingerprints or gait, can be used for identification or verification purposes. Gait is an emerging biometrics with great potential. Gait recognition is based on recognizing a person by the manner in which they walk. Its potential lays in that it can be captured at a distance and does not require the cooperation of the subject. This advantage makes it a very attractive tool for forensic cases and applications, where it can assist in identifying a suspect when other evidence such as DNA, fingerprints, or a face were not attainable. Gait can be used for recognition in a direct manner when the two samples are shot from similar camera resolution, position, and conditions. Yet in some cases, the only sample available is of an incomplete gait cycle, low resolution, low frame rate, a partially visible subject, or a single static image. Most of these conditions have one thing in common: static measurements. A gait signature is usually formed from a number of dynamic and static features. Static features are physical measurements of height, length, or build; while dynamic features are representations of joint rotations or trajectories. The aim of this thesis is to study the potential of predicting dynamic features from static features. In this thesis, we have created a database that utilizes a 3D laser scanner for capturing accurate shape and volumes of a person, and a motion capture system to accurately record motion data. The first analysis focused on analyzing the correlation between twenty-one 2D static features and eight dynamic features. Eleven pairs of features were regarded as significant with the criterion of a P-value less than 0.05. Other features also showed a strong correlation that indicated the potential of their predictive power. The second analysis focused on 3D static and dynamic features. Through the correlation analysis, 1196 pairs of features were found to be significantly correlated. Based on these results, a linear regression analysis was used to predict a dynamic gait signature. The predictors chosen were based on two adaptive methods that were developed in this thesis: "the top-x" method and the "mixed method". The predictions were assessed for both for their accuracy and their classification potential that would be used for gait recognition. The top results produced a 59.21% mean matching percentile. This result will act as baseline for future research in predicting a dynamic gait signature from static features. The results of this thesis bare potential for applications in biomechanics, biometrics, forensics, and 3D animation.

Este trabalho propõe um novo enfoque em visão computacional aplicado a sequências de vídeo, de pessoas em movimento, para reconhecê-las por meio da marcha. O movimento humano carrega diferentes informações, considerando-se diferentes maneiras de analisá-lo. O esqueleto carrega as informações do movimento global de articulações do corpo humano e como se comportam durante a caminhada e a silhueta carreia informações referentes ao comportamento global do contorno do corpo humano. Além disso, imagens binárias e em escala de cinza possuem diferentes informações sobre o movimento humano. O método proposto considera o conjunto de frames segmentados de cada indivíduo como uma classe e cada frame como um objeto desta classe. A metodologia aplica o Modelo de Mistura de Gaussianas (GMM) para subtração de fundo, redução de escala realizada por meio de técnicas de multiresolução baseadas na Transformada Wavelet (TW) e a extração dos padrões por meio da Análise dos Componentes Principais (PCA). São propostos e ensaiados quatro novos modelos de captura de movimentos globais do corpo humano durante a marcha: o modelo Silhouette-Gray-Wavelet (SGW) captura o movimento baseado nas variações em nível de cinza; o modelo Silhouette-Binary-Wavelet (SBW) captura o movimento baseado nas informações binárias da silhueta; o modelo Silhouette-Edge-Wavelet (SEW) captura o movimento baseado nas informações contidas na borda das silhuetas e o modelo Silhouette-Skeleton-Wavelet (SSW) captura o movimento baseado do esqueleto humano. As taxas de classificações corretas obtidas separadamente a partir destes quatro diferentes modelos são então combinadas utilizando-se uma nova técnica de fusão. Os resultados demonstram excelente desempenho e mostraram a viabilidade para reconhecimento de pessoas. / This paper proposes a novel computer vision approach that processes video sequences of people walking and then recognises those people by their gait. Human motion carries different information that can be analysed in various ways. The skeleton carries motion information about human joints, and the silhouette carries information about boundary motion of the human body. Moreover, binary and gray-level images contain different information about human movements. Our proposed method considers the set of the segmented frames of each individual as a distinct class and each frame as an object of this class. The methodology applies background extraction using the Gaussian Mixture Model (GMM), a scale reduction based on the Wavelet Transform (WT) and feature extraction by Principal Component Analysis (PCA). We propose four new schemas for motion information capture: the Silhouette-Gray-Wavelet model (SGW) captures motion based on grey level variations; the Silhouette-Binary-Wavelet model (SBW) captures motion based on binary information; the Silhouette-Edge-Binary model (SEW) captures motion based on edge information and the Silhouette-Skeleton-Wavelet model (SSW) captures motion based on skeleton movement. The classification rates obtained separately from these four different models are then merged using a new proposed fusion technique. The results suggest excellent performance in terms of recognising people by their gait.

Este trabalho propõe um novo enfoque em visão computacional aplicado a sequências de vídeo, de pessoas em movimento, para reconhecê-las por meio da marcha. O movimento humano carrega diferentes informações, considerando-se diferentes maneiras de analisá-lo. O esqueleto carrega as informações do movimento global de articulações do corpo humano e como se comportam durante a caminhada e a silhueta carreia informações referentes ao comportamento global do contorno do corpo humano. Além disso, imagens binárias e em escala de cinza possuem diferentes informações sobre o movimento humano. O método proposto considera o conjunto de frames segmentados de cada indivíduo como uma classe e cada frame como um objeto desta classe. A metodologia aplica o Modelo de Mistura de Gaussianas (GMM) para subtração de fundo, redução de escala realizada por meio de técnicas de multiresolução baseadas na Transformada Wavelet (TW) e a extração dos padrões por meio da Análise dos Componentes Principais (PCA). São propostos e ensaiados quatro novos modelos de captura de movimentos globais do corpo humano durante a marcha: o modelo Silhouette-Gray-Wavelet (SGW) captura o movimento baseado nas variações em nível de cinza; o modelo Silhouette-Binary-Wavelet (SBW) captura o movimento baseado nas informações binárias da silhueta; o modelo Silhouette-Edge-Wavelet (SEW) captura o movimento baseado nas informações contidas na borda das silhuetas e o modelo Silhouette-Skeleton-Wavelet (SSW) captura o movimento baseado do esqueleto humano. As taxas de classificações corretas obtidas separadamente a partir destes quatro diferentes modelos são então combinadas utilizando-se uma nova técnica de fusão. Os resultados demonstram excelente desempenho e mostraram a viabilidade para reconhecimento de pessoas. / This paper proposes a novel computer vision approach that processes video sequences of people walking and then recognises those people by their gait. Human motion carries different information that can be analysed in various ways. The skeleton carries motion information about human joints, and the silhouette carries information about boundary motion of the human body. Moreover, binary and gray-level images contain different information about human movements. Our proposed method considers the set of the segmented frames of each individual as a distinct class and each frame as an object of this class. The methodology applies background extraction using the Gaussian Mixture Model (GMM), a scale reduction based on the Wavelet Transform (WT) and feature extraction by Principal Component Analysis (PCA). We propose four new schemas for motion information capture: the Silhouette-Gray-Wavelet model (SGW) captures motion based on grey level variations; the Silhouette-Binary-Wavelet model (SBW) captures motion based on binary information; the Silhouette-Edge-Binary model (SEW) captures motion based on edge information and the Silhouette-Skeleton-Wavelet model (SSW) captures motion based on skeleton movement. The classification rates obtained separately from these four different models are then merged using a new proposed fusion technique. The results suggest excellent performance in terms of recognising people by their gait.

Submitted by Marlene Santos (marlene.bc.ufg@gmail.com) on 2016-08-10T19:31:03Z No. of bitstreams: 2 Dissertação - Weder Cabral Mendes - 2016.pdf: 1214460 bytes, checksum: 14588573f8f81fe4836a9945adacf37d (MD5) license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) / Approved for entry into archive by Luciana Ferreira (lucgeral@gmail.com) on 2016-08-15T13:25:32Z (GMT) No. of bitstreams: 2 Dissertação - Weder Cabral Mendes - 2016.pdf: 1214460 bytes, checksum: 14588573f8f81fe4836a9945adacf37d (MD5) license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) / Made available in DSpace on 2016-08-15T13:25:32Z (GMT). No. of bitstreams: 2 Dissertação - Weder Cabral Mendes - 2016.pdf: 1214460 bytes, checksum: 14588573f8f81fe4836a9945adacf37d (MD5) license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) Previous issue date: 2016-03-31 / Coordenação de Aperfeiçoamento de Pessoal de Nível Superior - CAPES / Gait recognition via computer vision attracted increasing interest for its noninvasive characteristic and mainly for your advantage of recognizing people at distance. Recognition is performing extracting features included in gait, this features are extracted from images sequence of people walking. The main challenges of gait recognition is to extract characteristics with unique information for each person, in additional, the use of accessories and clothes difficult the feature extraction process. This paper proposes a gait recognition method using information of people’s contours transformed in domain frequence by Discrete Fourier Transform. A lot of data are generated from the contours, thereby, three different techniques for dimensionality reduction CDA (Class Discrimination Ability), PCA (Principal Component Analysis) and PLS (Partial Least Squares) are employed to reduce the dimensionality of data and generate characteristics that are relevant to the recongnition system. Two classifiers, KNN (K-Nearest Neighbor) and LDA (Linear Discriminant Analysis) classify the characteristics that are returned by the dimensionality reduction methods. The accuracy are achieved by the combination of the dimensionality reduction methods and classifiers, the highest accuracy was 92:67%, which was achieved with the combination between the LDA and PCA (LDAPCA). Therefore, the results show that the information contained in the contours of silhouette are discriminant to recognize people by their gait. / O reconhecimento de pessoas através da marcha humana via visão computacional tem ganhado destaque por ser uma técnica biométrica não invasiva e principalmente por sua vantagem de reconhecer pessoas à distância. O reconhecimento é realizando extraindo características contidas na marcha de cada pessoa, essas características são extraídas de sequências de imagens da pessoa caminhando. Os principais desafios dessa técnica biométrica está em extrair as características com informações que consigam diferenciar uma pessoa da outra, além disso, o uso de acessórios e vestimentas dificultam o processo de extração de características. Este trabalho propõe um método de reconhecimento baseado na marcha humana utilizando informações dos contornos das pessoas transformados para o domínio da frequência por meio da Transformada Discreta de Fourier. Como são geradas muitos dados a partir dos contornos, três técnicas diferentes de redução de dimensionalidade CDA (Class Discrimination Ability), PCA (Principal Component Analysis) e PLS (Partial Least Squares) são empregadas para reduzir a quantidade de dados e gerar características que sejam relevantes para o sistema de reconhecimento. Dois classificadores, KNN (K-Nearest Neighbor) e LDA (Linear Discriminant Analysis) classificam as características retornadas pelos métodos de redução de dimensionalidade. As taxas de acurácia são obtidas pelos resultados gerados entre a combinação dos métodos de redução de dimensionalidade e os classificadores, a maior taxa de acurácia foi de 92;67%, a qual foi alcançada com a combinação entre o LDA e PCA (LDAPCA). Dessa forma, conclui-se que as informações contidas no contorno da silhueta no domínio da frequência são discriminantes para reconhecer pessoas através da marcha.

The aim of this thesis is to recognize a walking person in a sequence of images by defining his or her reference points to compare the course of their movement and then to identify the scanned person. Methods „k-means“ and „mean shift“ are used to obtain the silhouette of the person. However “environment model estimation” method is used here before those mentioned above. It is a type of a difference method and it helps to specify the scanning area and shortens the time of segmentation. During the search for the reference points the thesis focuses on three areas: the centre of the head and both ankle joints. Those points are later determined on the previous image sequence and compared with the real locations of the centre of the head and ankle joints marked by the user. The thesis also focuses on comparing the movement courses of those points and tries to identify the people whose walks are being scanned. Problematic situations which occurred during the whole process are analyzed in the end. The result of the thesis is an algorithm which can locate a moving person in an image sequence (video) and determine the reference points (centre of the head and ankles) to compare them and identify the scanned person.