Timothy Tzen Vun Yap

MMU GASPFA: A COTS multimodal biometric database

This paper describes the baseline corpus of a new multimodal biometric database, the MMU GASPFA (Gait-Speech-Face) database. The corpus in GASPFA is acquired using commercial off the shelf (COTS) equipment including digital video cameras, digital voice recorder, digital camera, Kinect camera and accelerometer equipped smart phones. The corpus consists of frontal face images from the digital camera, speech utterances recorded using the digital voice recorder, gait videos with their associated data recorded using both the digital video cameras and Kinect camera simultaneously as well as accelerometer readings from the smart phones. A total of 82 participants had their biometric data recorded. MMU GASPFA is able to support both multimodal biometric authentication as well as gait action recognition. This paper describes the acquisition setup and protocols used in MMU GASPFA, as well as the content of the corpus. Baseline results from a subset of the participants are presented for validation purposes.

Modelling of Direction-dependent Systems using Bilinear Models

The modelling of first and second order direction-dependent systems using bilinear models is considered, for perturbation using periodic signals. Equivalence between the two systems can be obtained for first order systems under binary perturbation. In most other cases, a close match can be obtained. The relationship between the parameters of a direction-dependent system and those of a bilinear system is investigated.

Identification of the static characteristics of a multizone furnace

The identification of the static characteristics of an electric multizone furnace is considered. A gray box approach is utilized. The furnace is divided into segments both horizontally and radially, and physical equations are used to describe heat transfers between segments. The identification strategy seeks to estimate an accurate temperature profile of the furnace at steady-state using a smaller number of sensors. Results indicate that the correct trend of the profile is captured. However, the accuracy is limited by the presence of nonlinear distortion as is confirmed by averaged step tests.

Autonomous Road Potholes Detection on Video

This research work explores the possibility of using deep learning to produce an autonomous system for detecting potholes on video to assist in road monitoring and maintenance. Video data of roads was collected using a GoPro camera mounted on a car. Region-based Fully Convolutional Networks (RFCN) was employed to produce the model to detect potholes from images, and validated on the collected videos. The R-FCN model is able to achieve a Mean Average Precision (MAP) of 89% and a True Positive Rate (TPR) of 89% with no false positive.

Residential Neighbourhood Security using WiFi

This paper focuses on the design of a WiFi-based tracking and monitoring system that can detect people’s movements in a residential neighbourhood. The proposed system uses WiFi access points as scanners that detect signals transmitted by the WiFi-enabled smartphones that are carried by most people. Our proposed system is able to track these people as they move through the neighbourhood. We implement our WiFi-based tracking system in a prototype and demonstrate that it is able to detect all WiFi devices in the vicinity of the scanners. We describe the implementation details of our system as well as discuss some of the results that we obtained.

Daily Activities Classification on Human Motion Primitives Detection Dataset

The study is to classify human motion data captured by a wrist worn accelerometer. The classification is based on the various daily activities of a normal person. The dataset is obtained from Human Motion Primitives Detection [1]. There is a total of 839 trials from 14 activities performed by 16 volunteers (11 males and 5 females) ages between 19 to 91 years. A wrist worn tri-axial accelerometer was used to accrue the acceleration data of X, Y and Z axis during each trial. For feature extraction, nine statistical parameters together with the energy spectral density and the correlation between the accelerometer readings are employed to extract 63 features from the raw acceleration data. Particle Swarm Organization, Tabu Search and Ranker are applied to rank and select the positive roles for the later classification process. Classification is implemented using Support Vector Machine, k-Nearest Neighbors and Random Forest. From the experimental results, the proposed model achieved the highest correct classification rate of 91.5% from Support Vector Machine with radial basis function kernel.

Identification of Road Surface Conditions using IoT Sensors and Machine Learning

The objective of this research is to collect and analyse road surface conditions in Malaysia using Internet-of-Things (IoT) sensors, together with the development of a machine learning model that can identify these conditions. This allows for the facilitation of low cost data acquisition and informed decision making in helping local authorities with repair and resource allocation. The conditions considered in this study include smooth surfaces, uneven surfaces, potholes, speed bumps, and rumble strips. Statistical features such as minimum, maximum, standard deviation, median, average, skewness, and kurtosis are considered, both time and frequency domain forms. Selection of features is performed using Ranker, Greedy Algorithm and Particle Swarm Optimisation (PSO), followed by classification using k-Nearest Neighbour (k-NN), Random Forest (RF), and Support Vector Machine (SVM) with linear and polynomial kernels. The model is able to achieve an accuracy of 99%, underlining the effectiveness of the model to identify these conditions.

Automatic Classification and Retrieval of Brain Hemorrhages

In this work, Computed Tomography (CT) brain images are adopted for the annotation of different types of hemorrhages. The ultimate objective is to devise the semantics-based retrieval system for retrieving the images based on the different keywords. The adopted keywords are hemorrhagic slices, intraaxial, subdural and extradural slices. The proposed approach is consisted of three separated annotation processes are proposed which are annotation of hemorrhagic slices, annotation of intra-axial and annotation of subdural and extradural. The dataset with 519 CT images is obtained from two collaborating hospitals. For the classification, support vector machine (SVM) with radial basis function (RBF) kernel is considered. On overall, the classification results from each experiment achieved precision and recall of more than 79%. After the classification, the images will be annotated with the classified keywords together with the obtained decision values. During the retrieval, the relevant images will be retrieved and ranked correspondingly according to the decision values.

Comparison of perturbation signals for ill-conditioned systems and their effectiveness in model-based control

Due to the effect of directionality, accurate identification of multivariable ill-conditioned systems is difficult without the use of specially designed perturbation signals. A comparison of three different designs of perturbation signals for the identification of these systems is presented, namely the virtual transfer function between inputs, modified zippered spectrum and rotated inputs designs. The effect of the variation of the condition number with frequency on the designs is illustrated via two simulation examples, where only one has a varying condition number. The performance of the designs for application in model-based control such as model predictive control is evaluated. For the system with the unvarying condition number, rotated inputs design is shown to perform best. The virtual transfer function between inputs design is shown to be superior when the condition number varies with frequency.