Repository logo

Doctoral Degrees (Computer Engineering)

Permanent URI for this collection


Recent Submissions

Now showing 1 - 9 of 9
  • Item
    Candidate generation and validation techniques for pedestrian detection in thermal (infrared) surveillance videos.
    (2022) Oluyide, Oluwakorede Monica.; Walingo, Tom Mmbasu.; Tapamo, Jules-Raymond.
    Video surveillance systems have become prevalent. Factors responsible for this prevalence include, but are not limited to, rapid advancements in technology, reduction in the cost of surveillance systems and changes in user demand. Research in video surveillance is majorly driven by rising global security needs which in turn increase the demand for proactive systems which monitor persistently. Persistent monitoring is a challenge for most video surveillance systems because they depend on visible light cameras. Visible light cameras depend on the presence of external light and can easily be undermined by over-, under, or non-uniform illumination. Thermal infrared cameras have been considered as alternatives to visible light cameras because they measure the intensity of infrared energy emitted from objects and so can function persistently. Many methods put forward make use of methods developed for visible footage, but these tend to underperform in infrared images due to different characteristics of thermal footage compared to visible footage. This thesis aims to increase the accuracy of pedestrian detection in thermal infrared surveillance footage by incorporating strategies into existing frameworks used in visible image processing techniques for IR pedestrian detection without the need to initially assume a model for the image distribution. Therefore, two novel techniques for candidate generation were formulated. The first is an Entropy-based histogram modication algorithm that incorporates a strategy for energy loss to iteratively modify the histogram of an image for background elimination and pedestrian retention. The second is a Background Subtraction method featuring a strategy for building a reliable background image without needing to use the whole video frame. Furthermore, pedestrian detection involves simultaneously solving several sub-tasks while adapting each task with IR-speci_c adaptations. Therefore, a novel semi-supervised single model for pedestrian detection was formulated that eliminates the need for separate modules of candidate generation and validation by integrating region and boundary properties of the image with motion patterns such that all the _ne-tuning and adjustment happens during energy minimization. Performance evaluations have been performed on four publicly available benchmark surveillance datasets consisting of footage taken under a wide variety of weather conditions and taken from different perspectives.
  • Item
    Contributions into holistic human action recognition.
    (2020) Toudjeu, Tchangou Ignance.; Tapamo, Jules-Raymond.
    In this thesis we holistically investigate the interpretation of human actions in both still images and videos. Human action recognition is currently a research problem of great interest both in academia and industry due to its potential applications which include security surveillances, sports annotation, human-computer interaction, and robotics. Action recognition, being a process of labelling actions using sensory observations, can be defined as a sequence of movements engendered by a human during an executed task. Such a process, when considering visual observations, is quite challenging and faces issues such as background clutter, shadows, illumination variations, occlusions, changes in scale, changes in the person performing the action, and viewpoint variations. Although many approaches to development of human action recognition systems have been proposed in the literature, they focused more on recognition accuracy while ignoring the computational complexity accompanying the recognition process. However, a human action recognition system which is both effective and efficient and can be operated real-time is needed. Firstly, we review, evaluate and compare the most prominent state-of-the-art feature extraction representations categorized between handcrafted feature based techniques and deep learning feature based techniques. Secondly, we propose holistic approaches in each of the categories. The first holistic approach takes advantage of existing slope patterns in the motion history images, which are a simple two dimensional representation of video, and reduces the running time of action recognition. The second one based on circular derivative local binary patterns outperforms the LBP based state-of-the-art techniques and addresses the issues of dimensionality by producing feature descriptor with minimal dimension size with less compromise on the recognition accuracy. The third one introduces a preprocessing step in a proposed 2D-convolutional neural network to deal with the same issue of dimensionality differently in the deep learning techniques. Here the temporal dimension is embedded into motion history images before being learned by a two dimensional convolutional neural network. Thirdly, three datasets (JAFFE, KTH and Pedestrian Action dataset) were used to validate the proposed human action recognition models. Finally, we show that better performance in comparison to the state-of-the-art methods can be achieved using holistic feature based techniques.
  • Item
    Error performance analysis of cross QAM and space-time labeling diversity for cross QAM.
    (2019) Kamdar, Muhammad Wazeer.; Xu, Hongjun.
    Abstract available in the PDF
  • Item
    Correcting inter-sectional accuracy differences in drowsiness detection systems using generative adversarial networks (GANs)
    (2020) Ngxande, Mkhuseli.; Tapamo, Jules-Raymond.; Burke, Michael.
    oad accidents contribute to many injuries and deaths among the human population. There is substantial evidence that proves drowsiness is one of the most prominent causes of road accidents all over the world. This results in fatalities and severe injuries for drivers, passengers, and pedestrians. These alarming facts are raising the interest in equipping vehicles with robust driver drowsiness detection systems to minimise accident rates. One of the primary concerns of motor industries is the safety of passengers and as a consequence they have invested significantly in research and development to equip vehicles with systems that can help minimise to road accidents. A number research endeavours have attempted to use Artificial intelligence, and particularly Deep Neural Networks (DNN), to build intelligent systems that can detect drowsiness automatically. However, datasets are crucial when training a DNN. When datasets are unrepresentative, trained models are prone to bias because they are unable to generalise. This is particularly problematic for models trained in specific cultural contexts, which may not represent a wide range of races, and thus fail to generalise. This is a specific challenge for driver drowsiness detection task, where most publicly available datasets are unrepresentative as they cover only certain ethnicity groups. This thesis investigates the problem of an unrepresentative dataset in the training phase of Convolutional Neural Networks (CNNs) models. Firstly, CNNs are compared with several machine learning techniques to establish their superior suitability for the driver drowsiness detection task. An investigation into the implementation of CNNs was performed and highlighted that publicly available datasets such as NTHU, DROZY and CEW do not represent a wide spectrum of ethnicity groups and lead to biased systems. A population bias visualisation technique was proposed to help identify the regions, or individuals where a model is failing to generalise on a picture grid. Furthermore, the use of Generative Adversarial Networks (GANs) with lightweight convolutions called Depthwise Separable Convolutions (DSC) for image translation to multi-domain outputs was investigated in an attempt to generate synthetic datasets. This thesis further showed that GANs can be used to generate more realistic images with varied facial attributes for predicting drowsiness across multiple ethnicity groups. Lastly, a novel framework was developed to detect bias and correct it using synthetic generated images which are produced by GANs. Training models using this framework results in a substantial performance boost.
  • Item
    Improvements of local directional pattern for texture classification.
    (2017) Shabat, Abuobayda Mohammed Mosa.; Tapamo, Jules-Raymond.
    The Local Directional Pattern (LDP) method has established its effectiveness and performance compared to the popular Local Binary Pattern (LBP) method in different applications. In this thesis, several extensions and modification of LDP are proposed with an objective to increase its robustness and discriminative power. Local Directional Pattern (LDP) is dependent on the empirical choice of three for the number of significant bits used to code the responses of the Kirsch Mask operation. In a first study, we applied LDP on informal settlements using various values for the number of significant bits k. It was observed that the change of the value of the number of significant bits led to a change in the performance, depending on the application. Local Directional Pattern (LDP) is based on the computation Kirsch Mask application response values in eight directions. But this method ignores the gray value of the center pixel, which may lead to loss of significant information. Centered Local Directional Pattern (CLDP) is introduced to solve this issue, using the value of the center pixel based on its relations with neighboring pixels. Local Directional Pattern (LDP) also generates a code based on the absolute value of the edge response value; however, the sign of the original value indicates two different trends (positive or negative) of the gradient. To capture the gradient trend, Signed Local Directional Pattern (SLDP) and Centered-SLDP (C-SLDP) are proposed, which compute the eight edge responses based on the two different directions (positive or negative) of the gradients.The Directional Local Binary pattern (DLBP) is introduced, which adopts directional information to represent texture images. This method is more stable than both LDP and LBP because it utilizes the center pixel as a threshold for the edge response of a pixel in eight directions, instead of employing the center pixel as the threshold for pixel intensity of the neighbors, as in the LBP method. Angled Local directional pattern (ALDP) is also presented, with an objective to resolve two problems in the LDP method. These are the value of the number of significant bits k, and to taking into account the center pixel value. It computes the angle values for the edge response of a pixel in eight directions for each angle (0◦,45◦,90◦,135◦). Each angle vector contains three values. The central value in each vector is chosen as a threshold for the other two neighboring pixels. Circular Local Directional Pattern (CILDP) isalso presented, with an objective of a better analysis, especially with textures with a different scale. The method is built around the circular shape to compute the directional edge vector using different radiuses. The performances of LDP, LBP, CLDP, SLDP, C-SLDP, DLBP, ALDP and CILDP are evaluated using five classifiers (K-nearest neighbour algorithm (k-NN), Support Vector Machine (SVM), Perceptron, Naive-Bayes (NB), and Decision Tree (DT)) applied to two different texture datasets: Kylberg dataset and KTH-TIPS2-b dataset. The experimental results demonstrated that the proposed methods outperform both LDP and LBP.
  • Item
    Feature regularization and learning for human activity recognition.
    (2018) Osayamwen, Festus Osazuwa.; Tapamo, Jules-Raymond.
    Feature extraction is an essential component in the design of human activity recognition model. However, relying on extracted features alone for learning often makes the model a suboptimal model. Therefore, this research work seeks to address such potential problem by investigating feature regularization. Feature regularization is used for encapsulating discriminative patterns that are needed for better and efficient model learning. Firstly, a within-class subspace regularization approach is proposed for eigenfeatures extraction and regularization in human activity recognition. In this ap- proach, the within-class subspace is modelled using more eigenvalues from the reliable subspace to obtain a four-parameter modelling scheme. This model enables a better and true estimation of the eigenvalues that are distorted by the small sample size effect. This regularization is done in one piece, thereby avoiding undue complexity of modelling eigenspectrum differently. The whole eigenspace is used for performance evaluation because feature extraction and dimensionality reduction are done at a later stage of the evaluation process. Results show that the proposed approach has better discriminative capacity than several other subspace approaches for human activity recognition. Secondly, with the use of likelihood prior probability, a new regularization scheme that improves the loss function of deep convolutional neural network is proposed. The results obtained from this work demonstrate that a well regularized feature yields better class discrimination in human activity recognition. The major contribution of the thesis is the development of feature extraction strategies for determining discriminative patterns needed for efficient model learning.
  • Item
    Power-line insulator defect detection and classification.
    (2018) Iruansi, Usiholo.; Tapamo, Jules-Raymond.; Davidson, Innocent Ewean.
    Faulty insulators may compromise the electrical and mechanical integrity of a power delivery system, leading to leakage currents owing through line supports. This poses a risk to human safety and increases electrical losses and voltage drop in the power grid. Therefore, it is necessary to monitor and inspect insulators for damages that could be caused by degradation or any accident on the power system infrastructure. However, the traditional method of inspection is inadequate in meeting the growth and development of the present power grid, hence automated systems based on computer vision method are presently being explored as a means to solve this problem speedily, economically and accurately. This thesis proposes a method to distinguish between defectuous and nondefectuous insulators from two approaches; structural inspection to detect broken parts and a study of hydrophobicity of insulators under wet conditions. For the structural inspection of insulators, an active contour model is used to segment the insulator from the image context, and thereafter the insulator region of interest is extracted. Then, di erent feature extraction methods such as local binary pattern, scale invariant feature transform and grey-level co-occurrence matrix are used to extract features from the extracted insulator region of interest image and then fed into classi ers, such as a support vector machine and K-nearest neighbour for insulator condition classi cation. For the hydrophobicity study of the insulator, an active contour model is used to segment water droplets on the insulator, and thereafter the geometrical characteristics of the water droplets are extracted. The extracted geometrical features are then fed into a classi er to assess the insulator condition based on the hydrophobicity levels. Experiments performed in this research work show that the proposed methods outperformed some existing state-of-the-art methods.
  • Item
    Investigation of feature extraction algorithms and techniques for hyperspectral images.
    (2017) Adebanjo, Hannah Morenike.; Tapamo, Jules-Raymond.
    Hyperspectral images (HSIs) are remote-sensed images that are characterized by very high spatial and spectral dimensions and nd applications, for example, in land cover classi cation, urban planning and management, security and food processing. Unlike conventional three bands RGB images, their high dimensional data space creates a challenge for traditional image processing techniques which are usually based on the assumption that there exists su cient training samples in order to increase the likelihood of high classi cation accuracy. However, the high cost and di culty of obtaining ground truth of hyperspectral data sets makes this assumption unrealistic and necessitates the introduction of alternative methods for their processing. Several techniques have been developed in the exploration of the rich spectral and spatial information in HSIs. Speci cally, feature extraction (FE) techniques are introduced in the processing of HSIs as a necessary step before classi cation. They are aimed at transforming the high dimensional data of the HSI into one of a lower dimension while retaining as much spatial and/or spectral information as possible. In this research, we develop semi-supervised FE techniques which combine features of supervised and unsupervised techniques into a single framework for the processing of HSIs. Firstly, we developed a feature extraction algorithm known as Semi-Supervised Linear Embedding (SSLE) for the extraction of features in HSI. The algorithm combines supervised Linear Discriminant Analysis (LDA) and unsupervised Local Linear Embedding (LLE) to enhance class discrimination while also preserving the properties of classes of interest. The technique was developed based on the fact that LDA extracts features from HSIs by discriminating between classes of interest and it can only extract C 􀀀 1 features provided there are C classes in the image by extracting features that are equivalent to the number of classes in the HSI. Experiments show that the SSLE algorithm overcomes the limitation of LDA and extracts features that are equivalent to ii iii the number of classes in HSIs. Secondly, a graphical manifold dimension reduction (DR) algorithm known as Graph Clustered Discriminant Analysis (GCDA) is developed. The algorithm is developed to dynamically select labeled samples from the pool of available unlabeled samples in order to complement the few available label samples in HSIs. The selection is achieved by entwining K-means clustering with a semi-supervised manifold discriminant analysis. Using two HSI data sets, experimental results show that GCDA extracts features that are equivalent to the number of classes with high classi cation accuracy when compared with other state-of-the-art techniques. Furthermore, we develop a window-based partitioning approach to preserve the spatial properties of HSIs when their features are being extracted. In this approach, the HSI is partitioned along its spatial dimension into n windows and the covariance matrices of each window are computed. The covariance matrices of the windows are then merged into a single matrix through using the Kalman ltering approach so that the resulting covariance matrix may be used for dimension reduction. Experiments show that the windowing approach achieves high classi cation accuracy and preserves the spatial properties of HSIs. For the proposed feature extraction techniques, Support Vector Machine (SVM) and Neural Networks (NN) classi cation techniques are employed and their performances are compared for these two classi ers. The performances of all proposed FE techniques have also been shown to outperform other state-of-the-art approaches.
  • Item
    Energy efficient medium access protocol for DS-CDMA based wireless sesor networks.
    (2012) Thippeswamy, Muddenahalli Nagendrappa.; Takawira, Fambirai.
    Wireless Sensor Networks (WSN), a new class of devices, has the potential to revolutionize the capturing, processing, and communication of critical data at low cost. Sensor networks consist of small, low-power, and low-cost devices with limited computational and wireless communication capabilities. These sensor nodes can only transmit a finite number of messages before they run out of energy. Thus, reducing the energy consumption per node for end-to-end data transmission is an important design consideration for WSNs. The Medium Access Control (MAC) protocols aim at providing collision-free access to the wireless medium. MAC protocols also provide the most direct control over the utilization of the transceiver, which consumes most of the energy of the sensor nodes. The major part of this thesis is based on a proposed MAC protocol called Distributed Receiver-oriented MAC (DRMACSN) protocol for code division multiple access (CDMA) based WSNs. The proposed MAC protocol employs the channel load blocking scheme to reduce energy consumption in the network. The performance of the proposed MAC protocol is verified through simulations for average packet throughput, average delay and energy consumption. The performance of the proposed MAC protocol is also compared to the IEEE 802.15.4 MAC and the MAC without the channel load sensing scheme via simulations. An analytical model is derived to analyse the average packet throughput and average energy consumption performance for the DRMACSN MAC protocol. The packet success probability, the message success and blocking probabilities are derived for the DRMACSN MAC protocol. The discrete-time multiple vacation queuing models are used to model the delay behaviour of the DRMACSN MAC protocol. The Probability Generating Functions (PGF) of the arrivals of new messages in sleep, back-off and transmit states are derived. The PGF of arrivals of retransmitted packets of a new message are also derived. The queue length and delay expressions for both the Bernoulli and Poisson message arrival models are derived. Comparison between the analytical and simulation results shows that the analytical model is accurate. The proposed MAC protocol is aimed at having an improved average packet throughput, a reduced packet delay, reduced energy consumption performance for WSN.