| Popis: |
Rapid advances in biometrics technology makes their use for person‘s identity more acceptable in a variety of applications, especially in the areas of the interest in security and surveillance. The upsurge in terrorist attacks in the past few years has focused research on biometric systems that have the ability to identify individuals from a distance, and this is spearheading research interest in Gait biometric due to being unobtrusive and less dependent on high image/video quality. Gait biometric is a behavioral trait that aims to identify individuals from image sequences based on their walking style. The growing list of possible civil as well as security applications for various purposes is paralleled by the emergence of a variety of research challenges in dealing with a various external as well as internal factors influencing the performance of Gait Recognition (GR) in unconstrained recording conditions. This thesis is concerned with Gait Recognition in unconstrained scenarios aims to address research questions covering (1) The selection of sets of features for a gait signature; (2) The effects of gender and/or recoding condition case (neutral, carrying a bag, coat wearing) on the performance of GR schemes; (3) Integrating gender and/or case classifications into GR; and (4) The role of emerging Kinect sensor technology, with its capability of sensing human skeletal features in GR and applications. Accordingly, our objectives will focus on investigating, developing and testing the performance of using a variety of gait sequencefeatures for the various components/tasks and their integration. Our tests are based on large number of experiments based on CASIA B database as well as an in-house database of Kinect sensor recording. In all experiments, we use different dimension reduction and feature selection methods do reduce the dimensions in these proposed feature vectors, such as Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA) and Fisher Score, followed by different classification methods like; k-nearest-neighbour (k-NN), Support Vector Machine (SVM), Naive Bayes and linear discriminant classifier (LDC), to test the performance of the proposed methods. The initial part is focused on reviewing existing background removal for indoor and outdoor scenarios and developing more efficient versions primarily by adopting the work for wavelet domain rather than the traditional spatial domain based schemes. These include motion detection by frame differencing and Mixture of Gaussians, the latter being more reliable for outdoor scenarios. Subsequently, we investigated a variety of features that can be extractedfrom various subbands of wavelet-decomposed frames of different body parts (partitioned according to the golden ratio). We gradually built sets of features, together with their fused combinations, that can categorized as hybrid of model-based and motion-based models. The first list of features developed to deal with Neutral Gait Recognition (NGR) includes: Spatio-Temporal Model (STM), Legs Motion Detection Feature (LMD), and the Statistical model of the approximation LL-wavelet subband images (AWM). We shall demonstrate that fusing these features achieves accuracy of 97%, which is comparable to the state of the art. These features will be shown to achieve 96% accuracy in gender classification (GC), and we shall establish that the NGR2 scheme that integrates GC into NGR improves the accuracy by a noticeable percentage. Testing the performance of these NGR schemes in recognising non-neutral cases revealed the challenges of Unrestricted Gait Recognition (UGR). The second part of the thesis is focused on developing UGR schemes. For this, first a new statistical wavelet feature set extracted from high frequency subbands, called Detail coefficients Wavelet Model (DWM) was added to the previous list. Using different combinations of these schemes, will be shown to significantly improve the performance for non-neutral gait cases, but to less extent in the coat wearing case. We then develop a Gait Sequence Case Detection (GSCD) which has excellent performance. We will show that integrating GSCD and GC together into UGR improves the performance for all cases. We shall also investigate the different UGS scheme that generalizes existing work on Gait Energy and Gait Entropy images (GEI and GEnI) features but in the wavelet domain and in different body parts. Testing these two schemes, and their fusion, post the PCA dimension reduction yield much improved accuracy for the non-neutral cases compared to existing scheme GEI and GEnI schemes, but are significantly outperformed by the last scheme. However, by fusing the UGS scheme with the GSCD+GC+UGR scheme above we will get best accuracy that outperform the state of the art in GR specially in the non-neutral cases. The thesis ended by conducting a rather limited investigation on the use of the Kinect sensors for GR. We develop two sets of features: Horizontal Distance Features and Vertical Distance Features from small set of skeleton point trajectories. The experimental result on neutral was very successful but for the unrestricted gait recognition (with the 5 case variations) satisfactory but not optimal performance relies on the gallery including balanced number of samples from all cases. |
