- Study at Deakin
- Campus life
- Industry and community
- About Deakin
We propose a framework to use human gesture as input to trigger events within a DI-Guy simulation scenario in real-time, which could greatly help users to control events and avatar reactions in the scenario. We use Microsoft Kinect device for motion capture and write our own plug-in for gesture recognition and associate gestures with various commands in scene. The framework is a distributed system in which different modules are communicating and synchronizing through data streams. This provides a scalable loosely coupled highly cohesive modular framework where any component can be altered or modified without redesigning the whole system. The proposed framework could provide fast, affordable yet reliable environments for real time interactive crowd simulations. It supports direct user gesture input and allow for direct and collateral avatar reactions based on artificial intelligence. The gesture libraries provide most commonly used gesture recognitions and it could be easily expanded with more requirements.
|Figure 1: Skeleton tracking using Kinect (top) Intuitive interfacing with crowd simulation package (DI-Guy) (bottom)|
The vast majority of image fusion cases take place by having high frequencies information in a source image at spatial coordinates where the other image holds low frequencies. While this is quite common in fusing infrared and thermal images with visual images (at night), expanding image fusion to accept multi-source multi-modal images raises concerns of saturating the resulting fused image and highlights the need to study the fusion capacity in order to minimize the overlapping of high frequencies causing fusion artefacts.
Figure 1: Fusing a natural scene with saturated (8-bits) images
Figure 2: Fusion capacity maps using localised mutual information
Natural scene imaging statistics suggest that the probability to find a uniform image tends to zero. Therefore, the fusion capacity of an image can be approximated by measuring the distance between the normalised histogram of the examined image and the uniform distribution using mutual information. Using the normalised local mutual information measure fusion capacity can be formulated as follows;
where x is the source image with normalised histogram X and u is the uniformly distributed image with a normalised uniform histogram U.
In nature, the fish adjusts itself to hydrodynamics environment and becomes a perfect swimming expert during years of evolution, which drives many researchers to study its body structure and swimming characteristics. With the development of propulsive theories and robotic technologies, the research on a biomimetic robot fish with high velocity, high efficiency and high manoeuvrability has been a hotspot. This helps human being have more chances of exploring the mysterious underwater world and develop higher efficient propellers for ships or underwater vehicles. We designed a new free-swimming biomimetic robot fish (Fig. 1) with a biomimetic tail to simulate the carangiform tail, a barycenter-adjustor for descending/ascending motions and multiple sensors. The robot fish can communicate with the outside by an information relay system on water. We also proposed a three-dimensional (3D) computational fluid dynamic simulation of the biomimetic robot fish by Fluent. User-defined function (UDF) is used to define the movement of the robot fish and a Dynamic Mesh is used to mimic the fish swimming in water. The hydrodynamic analysis (for example: Fig. 2) of the robot fish helps us get comparative data about hydrodynamic properties and guides us to improve the design, remote control and flexibility of the underwater robot fish. Figure 2 shows pressure contours at 4 locomotion times in one fish tail swing period (N*m-2) (a) approaching the left maximal rotation position (t=0.1s); (b) intermedial position during oscillating motion (t=0.25s); (c) going on intermedial position during oscillating motion (t=0.4s); (d) approaching right maximal rotation position (t=0.55s).
Figure 1: The functional prototype of the biomimetic robot fish
One well acknowledged drawback of traditional parallel kinematic machines (PKMs) is that the ratio of accessible workspace to robot footprint is small for these structures. This is most likely a contributing reason why relatively few PKMs are used in industry today. The SCARA-Tau structure is a parallel robot concept designed with the explicit goal of overcoming this limitation and developing a PKM with a workspace similar to that of a serial type robot of the same size. Research done at CISR shows how a proposed variant of the SCARA-Tau PKM can improve the usability of this robot concept further by significantly reducing the dependence between tool platform position and orientation of the original concept. The inverse kinematics of the proposed variant has been derived and a comparison made between this structure and the original SCARA-Tau concept, both with respect to platform orientation changes and workspace.
Fig. 1. The SCARA-Tau prototype (left). A model of the same structure seen from the top (middle) and a model of the modified structure seen from the top (right).
Parallel kinematic machines (PKMs) are receiving increasing attention, both in academia and industry. One rapidly growing use for PKMs is in the solar photovoltaic (PV) industry, where a large number of DELTA robots are used. There is a rising demand for fast robots with larger workspace than what is available today. The SCARA-Tau robot is a novel PKM with a large workspace that would be highly suitable for applications in the PV industry. Present research efforts at CISR are aimed at evaluating the feasibility of using this robot in solar cell manufacturing applications. Adapting the SCARA-Tau robot to solar cell manufacturing involves finding optimal structural parameters for this application. This is a multi-purpose optimization problem. An optimal workspace should have a large volume below the lowest upper arm of the robot, while the isotropic properties and dexterity of the structure should be kept high.
Fig. 1. Parallel kinematic robots used for solar cell manufacturing (left) and the SCARA-Tau prototype (right)
Stereo correspondence estimation in one of the most active research areas in the field of computer vision and number of techniques has been proposed and developed, possessing both advantages and shortcomings. Among the techniques reported, multiresolution analysis based stereo correspondence estimation has gained a lot of research focus in recent years. Although the most widely employed medium for multiresolution analysis is wavelets and multiwavelets bases, however, relatively little work has been reported in this context. In this research we have tried to address some of the issues regarding the work reported in this domain and the shortcomings involved. While addressing the shortcomings of the existing algorithms, we also propose a new technique to overcome some of the flaws that could have significantly negatively impacted on the algorithm performance and has not been addressed in the earlier propositions.
Our algorithm uses multi-resolution analysis enforced with wavelets/multiwavelets transform modulus maxima to establish correspondences between the stereo pairs of images. A variety of wavelets and multiwavelets bases, possessing distinct properties such as orthogonality, approximation order and shapes are employed to analyse their effect on the performance of correspondence estimation. The idea is to provide knowledge base to understand and establish relationships between wavelets and multiwavelets properties and their effect on the quality of stereo correspondence estimation. In addition, comparative performance analysis of the proposed algorithm, with eight existing famous algorithms, is also performed to provide an insight of the capabilities of the proposed algorithm as well as the potential of wavelets and multiwavelets theories in stereo vision.
(RMS) for number of selected images with wavelets/multiwavelets basis highlighted with different colours
In many imaging applications, such as medical imaging and surveillance operations, it is beneficial to extract key details from captured images. However, in such applications the imaging technology used often results in low quality images, making it difficult to extract meaningful information. For example, a surveillance camera may capture a wide field of view, at low resolution, but if a forensic team needs to identify a suspect's face or a car number plate it is often not possible.
Super-resolution is a method of enhancing images so that features of interest can be extracted in fine detail, while still using low resolution imaging hardware. Multi-frame super-resolution fuses information from a series of low resolution images to create an image, or series of images, with a higher spatial resolution. As well as increasing resolution, this concept can be used to extend the field of view, remove moving objects and correct degradations inherent in the imaging process, such as noise, blur and spatial-sampling errors. Input images for multi-frame super-resolution algorithms must be captured from slightly different perspectives of a scene, resulting in a sub-pixel shift between images, so that detail not captured by the original imaging system can be recovered.
Current super-resolution algorithms make a range of assumptions about the low resolution input images to simplify the scenario. Super-resolution algorithms generally use a simple affine image transform, permit only minute displacement between images and assume all images belong to a planar scene. In removing these restrictions it is possible to greatly increase the number of possible applications for this technology. Allowing significant displacement between images and using a projective, rather than affine transform, means super-resolution may be used to create a high resolution view of a 3-Dimensional scene.
An image fusion metric does not suit all fusion algorithms. Some algorithms have to be evaluated with specially designed metrics. This section discusses, through counter examples, the compatibility challenges facing image fusion metrics with respect to a certain fusion algorithm.
Figure 1 illustrates different incompatibility challenges between image fusion algorithms and metrics. The proposed duality index depends on performing fusion experiments where we know exactly how the result should be. Developing such an index is fairly simple although it has its solid mathematical background listed in abstract algebra literature. The duality index depends mainly on estimating if the fusion operator measures the actual transferred data from a source image with an non-informative, also called a zero, image. The image fusion algorithm/metric duality index is then defined as DI0,:⊕x→R where ⊕ is the set of all fusion algorithms and is the set of all fusion metrics. The duality index is estimated as the average error in Equation below for all images in the domain.
While single-domain image fusion systems aim to add information from source images into one fused image, the main objective of multi-domain image fusion is to incorporate captured information from sensors observing a certain phenomena from different viewpoints. This allows the observer to understand the whole situation. However, not all sensors are as accurate, low cost, portable and readily available. In fact, some of these sensors have limitations according to the nature of captured data, power consumption and cost of the device. Most of these specially designed sensors adopt a color map that best demonstrates the captured information. This is because most non-visual sensors are too complex to be equipped with large memory modules, faster codecs or modules to support higher bandwidth. This work values image properties during the fusion process.
where F⊕ is the set of fusable images with their properties and Pi⊕i is the ith image property
Color-map and Histogram Fusion
Multi-sensor data fusion refers to the process of combining data from different resources to improve the quality of the information and accuracy of measurements. It is in the heart of a widespread range of applications including military, smart buildings and bridges, satellites and industry. The main problem with multi-sensor data fusion is that it is highly dependent on the quality of measurements obtained from each sensor. And since these sensors operate in real environments, there is no guarantee that the outputs of the sensors are accurate. The inaccuracies can be because of the health of the sensors, i.e. the battery charge condition, or the high noise nature of the environment. Some errors may also occur due to the faulty communication channels. The result of these situations will be the loss of measurements or packets that may cause the data fusion process to diverge.
The main goal of this research is to incorporate the possibility of missing measurements or packet in the data fusion process and obtain the optimal data fusion for systems suffering from this problem. The research finds the minimum error variance of the resulting system state estimation using a Kalman-like recursive filter. Figures 1 and 2 represent the performance of the newly developed optimal filter against the traditional data fusion techniques in terms of error variance and MSE.
Kalman filtering is a well known technique for state estimation of real-time dynamic systems. It provides the minimum variance error of state estimation when the system incorporates noises from the environment and the sensors. However, the Kalman filters assume perfect modelling of the system and complete knowledge about the system parameters. In real-world situations, these assumptions are not valid and the outputs of the Kalman filters will not be robust. This research addresses the problem of system modelling when the system under consideration suffers from uncertainties. It also provides the robust design of a new filter that provides the optimal filtering for systems under uncertainties without prior knowledge about the uncertainty values.
The uncertainties addressed in this research are the uncertainties in the modelling parameters and the uncertainties in the observations. The first kind may take place due to inaccurate modelling of the systems, linearization of nonlinear systems or model reduction. The second kind of uncertainty is due to poor communication channels, unhealthy sensors, high manoeuvring of tracked objects or high noise in the environment. The results of this research include finding an upper bound on the estimation error covariance and finding the optimal filter parameters that guarantee the estimation error covariance to fall below this upper bound. Figures 1 and 2 compare the result of state estimation of a two state system suffering from uncertainties in the modelling parameters and uncertainties in the observation process using the proposed robust filter and the conventional Kalman filter.
Despite the advantages of MRI, there are problems related to the use of MRI machines. The biggest problem when using MR imaging is that it is a time consuming process where a single scan may take 30 minutes. This creates large patient backlogs that delay their medical diagnosis and proper medication. Another major problem is that the MRI scanning process is sensitive to patients' movements. In some cases it is hard to control the movement of the patient such as in the case of children. In this case the patient is required to do the MRI scan again which increases the backlog. The MRI machines are complex and very expensive equipment. This makes increasing the number of operating MRI machines an impractical solution for reducing the backlogs. The practical solution would be to increase the speed of the scanning process itself. However, the speed of MRI machines is physically limited and cannot be any faster as long as the same number of measurements is required.
This research addresses the problem of increasing the speed of the MRI scanning process using a novel measurements sampling technique that guarantees the minimum number of acquired measurements without compromising the scan quality. The idea is to find the best set of measurements that represent the whole set of measurements before starting the scanning process. This will eliminate any chance of redundancies in the measurements. The new data acquisition technique reduces the time of the scan by one third without introducing any noise. Higher speeds can be achieved in the cost of introducing additive noise. The results of this research also include guaranteed robustness of the image reconstruction in the presence of patient frequent movements. The next figure represents the result of image acquisition from 65% of the measurements and the result after robustly filtering the output image.
This research investigates a methodology for generating a distinctive object representation offline, using short-baseline stereo fundamentals to tri-angulate highly descriptive object features in multiple pairs of stereo images. Several sparse 2.5D perspective views are generated and then registered into a single coordinate space. Having some priori knowledge, such as the proposed sparse-feature model, is very useful when detecting an object and estimating its pose in real-time systems, such as augmented reality.
Figure - 2.5D perspective view of a box edge
Creating a highly programmable surface operating at relatively high speed and in real time is an area of research with many challenges. Such a system has applications in the field of optical telescopes, product manufacturing, and giant 3D-screens and billboards for advertising and artwork. This research investigates various system designs, modularity, programmability and the system control intelligence. A simulation environment was developed to streamline system reconfiguration to translate complex mathematical functions into 3D shapes virtually before being displayed on the physical surface.
World leading operations are increasingly relying on modelling and simulation to develop more efficient systems and to produce higher quality products and services. Modelling and simulation allows scientists and engineers a better understanding of three-dimensional and time-dependent phenomena, as well as providing a platform for predicting future behaviour. Virtual training (VT) systems utilising advanced virtual reality technology is of growing importance to widespread application domains, such as aerospace, engineering, medical and education fields.
The perceived capability and technological affordance for enhancing human abilities to learn abstract concepts and complex procedural tasks of VT systems led to a wide adaptation for training and fast-paced technological advancements. Despite its adaptation for training and fast-paced technological advancements, ways in which to evaluate efficacy of such technology are unclear. In addition, how to better design such technology to achieve effective learning outcomes remains a significant challenge. In particular, consider the complexity of individual differences on human performance, adaptation and acceptance of new software and hardware devices evoked by 3D VT systems.
It is imperative to obtain a solid understanding of important elements that contribute to effective learning via such class of applications. This problem is addressed by developing a new evaluation method that focuses on cognitive, affective and skill-based learning dimensions. Results highlight the contribution of the method in analysing user performance, and understanding of the individual differences in learning ability and user experience within the 3D VT. The benefits of the method imply that it is effective to provide adequate, reliable and valid evaluation results in a comprehensive and systematic fashion. In addition, by exploring spatial knowledge and technical skill acquisition, knowledge visualization and user perceptions in a machine assembly training scenario, this research suggests 3D VT is effective in integrating multimodal system feedback and presents information appropriate to user input. Also the 3D VT is effective to support cognition and performance, as well as induce positive user perception and affect.