Unmanned Ariel Vehicles (UAVs) require identifying water surfaces during flight maneuvers, mainly for safety in execution and its applications. We introduce two novel techniques to identify water surfaces from frontfacing and downward-facing cameras mounted on a UAV. The first method - UNet-RAU, a unique architecture based on UNet and Reflection Attention Units, segments water pixels from front-facing camera views, utilizing the reflection property of water surfaces. On the On-Road and Off-Road datasets of Puddle-1000, UNet-RAU improved its performance by 2% over the state-of-the-art FCN-RAU. Additionally, the UNet-RAU generated an F1-score of 80.97% on our Drone-Water-Front dataset. The second method - Dense Optical Flow based Water Detection (DOF-WD), detects water surfaces in videos of downward-facing cameras. This method utilizes downwash-generated ripples and natural texture features on a water surface to identify water in low and high altitudes, respectively. We empirically validated the performance of the DOF-WD method using our Drone-Water-Down dataset.

Elephants generate infrasonic vocalisations that traverse through the air for long distances. Utilising this phenomenon, a previous work proposed a system, called Eloc, to localise and track elephants in the wild. The Eloc system has been demonstrated to be accurate in calculating the location of infrasonic sources. However, it still lacks the capability to accurately distinguish elephant infrasonic calls from various other infrasonic sources using limited computing power on board. Addressing this problem, the work presented in this paper introduces an approach to distinguish elephant infrasonic calls with a high accuracy on low-resourced hardware. Firstly, a sequence of operations are performed to reduce the effect of noise in the infrasonic signal captured by an Eloc node. Secondly, a wavelet-based signal reconstruction technique is applied to extract spectral features from the infrasonic signal. Finally, the extracted features are fed to a pre-trained machine learning classifier to distinguish the infrasonic vocalisations of elephants. The experimental evaluation using Asian elephant (Elephas Maximus Maximus) infrasonic vocalisation datasets demonstrates that the proposed approach is capable of accurately distinguishing elephant infrasonic calls on low-resourced hardware platform of the Eloc system, with accuracy levels over 82% under varying environmental conditions.

This paper presents a novel method to identify unseen Asian elephants that are not previously captured or identified in available data sets and re-identify previously seen Asian elephants using images of elephant ears, leveraging a semi-supervised learning approach. Ear patterns of unseen elephants are learnt for future re-identification. To aid our process, elephant ear patterns are used as a biomarker to uniquely identify individual Asian elephant, each of which is attached a descriptor. The main challenge is to learn and use a clustering technique to identify new classes (i.e., elephants) in unlabelled elephant ear image sets and leveraging this data in verifying the labelled images. This study proposes a systematic approach to address the problem to uniquely identify elephants, where we developed: (a) a self-supervised learning approach for training the representation of labelled and unlabelled image data to avoid unWanted, bias labelled data, (b) rank statistics for transferring the models’ knowledge of the labelled classes when clustering the unlabelled images, and, (c) improving the identification accuracy of both the classification and clustering algorithms by introducing a optimization problem when training with the data representation on the labelled and unlabelled image data sets. This approach was evaluated on seen (labelled) and unseen (unlabelled) elephants, where we achieved a significant accuracy of 86.89% with an NMI (Normalized Mutual Information) score of 0.9132 on identifying seen elephants. Similarly, an accuracy of 54.29% with an NMI score of 0.6250 was achieved on identifying unseen elephants from the unlabelled Asian elephant ear image data set. Findings of this research provides the ability to accurately identify elephants without having expert knowledge on the field. Our method can be used to uniquely identify elephants from their herds and then use it to track their travel patterns Which is greatly applicable in understanding the social organization of elephant herds, individual behavioural patterns, and estimating demographic parameters as a measure to reducing the human-elephant conflict in Sri Lanka.

The rapid progress in deepfake content generation has now come to a point where it raises significant concerns about the implications for society. Therefore, a new challenge of detecting deepfakes arises to protect individuals from potential misuse. Even though introduced detection algorithms perform well on high-quality deepfakes, detecting low-quality deepfakes has been challenging. As a remedy, researchers try to feed more training data to increase detection ability. HoWever, providing more data and processing them is not always feasible in a practical scenario. Thus, for the first time in this domain, we propose to employ super-resolution (SR) as a preprocessing step instead of feeding more data to improve low-quality deepfake detection. Extensive experiments were conducted on the FaceForensics++ deepfake dataset. Initially, three baseline models, Meso-4, MesoInception-4, and XceptionNet, were trained and tested on the dataset without any preprocessing mechanism. XceptionNet outperformed with 90.54% accuracy revealing deeper networks detect low-quality depfakes adequately. Then those baseline models were trained with SR preprocessing. To do that, we employed two SR networks, called VDSR and RSRGAN. RESRGAN+XceptionNet outperformed the previous baseline models by obtaining 96.05% accuracy, showing SR preprocessing usefulness in low-quality deepfake detection. Further experiments utilizing performance metrics, statistical tests, and visualization of activation maps showed that SR preprocessing is promising when applied to deepfake detection networks and detection algorithms experience a significant performance.

Garbage classification is a vital sub-process of garbage management. Litter differs drastically with space and time, hence visual classification of garbage with deep learning under-performs due to this high variation in the garbage. This paper proposes an active deep learning approach consisting of a novel querying function to determine when to query the labels from a human user and rapidly adapt to user input. The querying function was constructed by first calculating four input features (Least Confidence, Marginal Value, Confidence Ratio, and Entropy Value) and then by learning an XGBoost classifier to predict the mistakes of the DenseNet-169 main-classifier. The research involves three steps: (a) Pretraining the DenseNet169 with TrashNet dataset, (b) Learning an XGBoost classifier to predict mistakes of the DenseNet169 main-classifier to be used as a querying function for active learning, (c) Replacing the last layer of the DenseNet169 with another XGBoost classifier to adapt to human input rapidly, The full-cascade is named ActiveCNN-XGB – Active Convolutional Neural Network with XGBoost. A New Trash Dataset (NTD) was constructed with 1750 in-situ garbage images. As a querying function, the XGBoost classifier predicted mistakes of the DenseNet169 with an accuracy of 80.15% for image streams. Out of 650 test images, the proposed ActiveCNN-XGB gained an accuracy of 74.40% which is an improvement of 21.07% on NTD.

Individual identification is a basic requirement for research in behavior, ecology and conservation. Photographic records are commonly used in situations where individuals are visually distinct. However, keeping track of identities becomes challenging with increasing population sizes and corresponding datasets. There is growing interest in the potential of deep-learning methods for computer vision to assist with automating this task. Here we apply Convolutional Neural Networks, a popular architecture for Artificial Neural Networks used in image classification, to the problem of identifying individual Asian elephants through photographs. We evaluate the performance of five different types of CNN models used in facial recognition (VGG16, ResNet50, InceptionV3, Xception, and Alexnet), on datasets representing three different feature regions (the full body, face, and ears), trained with two techniques (transfer learning vs. training from scratch) for n = 56 elephants. We tested accuracy in matching the top candidate as well as top five candidates. We found that VGG16 trained with the transfer-learning technique outperformed other models on the body and face datasets with accuracies of 21.34% and 42.35%, respectively, in matching the top candidate. Nevertheless, the best performance was achieved by an Xception model trained from the scratch on the ear dataset, with an accuracy of 89.02% for matching the top candidate and 99.27% for including the correct individual among the top five. However, this impressive level of accuracy was obtained with a dataset of 3816 labeled training images of 56 elephants. There are more than 1000 wild elephants in the population under observation, requiring extensive human effort and skill to initially annotate the images used as training data. Therefore, we consider this approach impractical for monitoring large wild populations. Nevertheless this it could be very useful in record keeping and fraud prevention for large captive elephant populations, as well as monitoring animals that have been rehabilitated and released or moved for management purposes.

Zoologists visually recognise each Asian elephant (Elephas maximus), mainly based on their ear patterns. Towards automating this process, the existing methods on African elephants are less instrumental for Asian elephants, because the nick patterns are rare. This paper presents a cascade of Convolutional Neural Networks for uniquely detecting Asian elephants with two steps: (1) an elephant-ear localisation step at a species level, and (2) an ear-patch classification step at an individual level. First, a YOLO CNN with pre-trained weights on ImageNet is retrained with manually cropped elephant ears to localise them in the colour image. Second, these cropped ear patches are learnt by a CNN to classify each elephant by the Zoologist’s labelling; Xception outperformed VGG16, ResNet50, InceptionV3, and AlexNet in this second step on 56 elephants. Xception produced a top-1 accuracy of 88% and top-5 accuracy of 99.27% for reidentification as the best performance. Discriminative regions of elephant ears were visually explained by GradCAM on Xception reidentification classifier.

Fashion design is an art that reshapes the designers’ imagination into visible content which requires a significant amount of time and effort. The assistance provided by the available design tools are limited in the sense of visualizing and fitting of the generated cloth on the human body. We present, ARGAN-an Augmented Reality (AR) based Fashion Design system which is able to generate a new dress when a sketch and a theme image are provided as the input into a Controllable Generative Adversarial Network. Further, this system can visualize the generated virtual 2D apparel in realtime on a real human body using Augmented Reality. To the best of our knowledge, this work is the first attempt at utilizing Deep Generative Models (e.g. GANs) in an Augmented Reality prototype in fashion designing for generate creative fashion content in 2D and exploiting the possibility of Deep Generative Models to generate fashion designs align to a theme. Our findings show that the use of the ARGAN can support fashion designers’ during their designing process.

In intelligent traffic light control, matrices derived from real-time traffic data are paramount for efficiency and performance. The rewards and state representations in previous studies could mislead a Reinforcement Learning agent in some cases. This paper examines the effectiveness of considering the Standard Deviation of vehicle’s Waiting Time (SDWT) on Deep Reinforcement Learning based traffic congestion control with emergency facilitation. The proposed method was self-evaluated by only considering average waiting time under both synthetic and Toronto real-world dataset. It has demonstrated that the proposed method was able to gain a significant impact on performance by considering the SDWT. Moreover, the proposed method was able to reach zero waiting time for emergency vehicles.

Lantana camara is an exotic invasive plant that has been a major threat to the biodiversity of several areas around the world. This paper presents a novel methodology to identify the distribution of Lantana camara flowers in aerial images. The proposed model uses aerial images as inputs and the model consist of three stages. The first step is the detection of possible flower patches in the aerial images using Local Binary Patterns mechanism. The second step is the recognition of Lantana camara flowers from the localized flower patches through a classification process using a Convolutional Neural Network (CNN). The third step is the marking the presence of the flowers of Lantana camara in the original image. Achieved sensitivity rate of this study is 40.71%. The proposed model succeeded in identifying Lantana camara distribution by identifying the presence of Lantana camara flowers in aerial images.

The human-elephant conflict is causing significant damages to the life of human and elephants in Sri Lanka. Minimizing encounters between humans and elephants is hence crucial in alleviating the human-elephant conflict. We have earlier introduced Eloc, a cost-effective localization system that is based on infrasonic emissions from wild elephants. One of the remaining challenges is a resource-efficient method to detect the elephants' infrasonic emissions. We present our efforts on devising a support vector machine (SVM) that is able to detect elephant rumbles on Eloc nodes.

Radio Tomographic Imaging (RTI) enables device free localization of physical objects by using signal attenuation in wireless networks. In this paper, we explore how existing RTI methods can be used in WiFi networks to do tomographic imaging. Moreover we analyze and evaluate the properties that affect the accuracy of WiFi tomographic imaging process.

This paper proposes a new algorithm for the multiple instance learning problem (MIL) and investigates its application for detecting Attention Deficit Hyperactive Disorder (ADHD) from resting-state functional Magnetic Resonance Imaging data. The core component of many kernel-based MIL algorithms is usually an SVM-like batch optimization framework, hence scaling to large datasets like fMRI is often difficult. On the other hand, a family of on-line kernel classification algorithms widely known as “perceptron-like” kernel classifiers demonstrate efficient and accurate solutions. This paper presents MiPAL - Multiple-instance Passive Aggressive Learning algorithm, based on such a perceptron-like kernel classifier. First, MiPAL builds a labeller by inputting negative bags into PA algorithm. Second, this labeller helps to train a separate PA classifier to predict binary class labels such that least-negative instances are regarded as positive. Due to the on-line PA algorithm's fast adaptation, the impact of invalid positive support-vectors could be attenuated by the new, accurate support-set over time. Our experimental results reveal performance gains in several MIL datasets including state-of-the-art performance in Muskl, Fox, and comparable accuracy in the preprocessed ADHD-200 dataset.

This paper addresses the problem of on-line learning for object tracking. Although a variety of techniques have been proposed in literature, a recent benchmark reveals that none of them can work well in all scenarios due to numerous practical challenges, such as illumination variations, motion blur, etc. These challenges occur at different time frames making it hard to design a tracker. In this paper, a machine learning framework for object tracking is investigated, which can integrate with a variety of feature and kernel engineering techniques in dealing with many challenges in different scenarios. By following the successful tracking-by-detection methodology, this paper proposes OMKT - On-line Multiple Kernel Tracking scheme, which attempts to tackle the object tracking task by exploring recent advances of on-line multiple kernel learning techniques in machine learning. In particular, OMKT sequentially learns the best tracker for each individual kernel via on-line Projectron++ learning, and at the same time attempts to identify the optimal combination of multiple kernel trackers using the Hedge algorithm. In contrast to many existing schemes in literature, OMKT learns both the kernel classifiers and their combination on-line, hence it can adapt to tracking changes faster. Furthermore, the projection strategy in Projectron++ alleviates the difficulty of pre-specifying a budget size for support-set. Promising experimental results on a recent benchmark reveal usefulness of OMKT.

Dengue is one of the deadly and fast spreading diseases in Sri Lanka. The female Aedes mosquito is the dengue vector and these mosquitoes breed in clear and non-flowing water. The Public Health Inspectors (PHIs) are tasked with detecting and eliminating such water collection areas. However, they face the problem of detecting potential breeding sites in hard-to-reach areas. With the technological development, the drones come as one of the most cost effective unmanned vehicles to access the places that a man cannot access. This paper presents a novel approach for identifying mosquito breeding areas via drone images through the distinct coloration of those areas by applying the Histogram of Oriented Gradients (HOG) algorithm. Using the HOG algorithm, we detect potential water retention areas using drone images.

We describe a real time authoring tool that is useful to supervise group of users in an augmented reality environment. The overall system is composed of a server; the Command Center (CC) and several clients; Wearable Systems (WS). The server is used to visualize the virtual representation of the real environment, 3d features map where augmentation are to occur, augmented content that will be rendered and users avatar that evolves on the field. The user's WS is equipped of a Head Mounted Display (HMD) along with sensors (GPS/DRM/Compass) to provide constant pose estimation (position and orientation) to the CC and a vision based tracking that uses natural features and pose estimation allgorithms to render virtual content accordingly. Alternatively a sensor based renderer is used to display AR guidance on the HMD based on the path defined and sent from the CC. The natural features system handles a large number of small 3D features maps. The overall system offers a complete system to train and supervise remote users through an AR environment.

This paper presents "dMarkers", a concept of "dynamic markers" or markers that can change for marker technologies for Augmented Reality (AR) technologies. Here, by dynamic we mean markers that can change or morph into different markers which could open up various new possibilities through technologies such as AR. To impelment this technology we use a combination of thermochromic inks and peltier semiconductor elemnts to achieve a subtly changing animated display for paper and fabric materials. The key feature of this technology lies within its ability to animate the material itself without embedding any separate displays. Thus this technology is used to subtly animate various patterns for markers to be used as dynamic markers. As a proof of concept, we implement the dMarkers for Quick Response (QR) codes. In our implementation we present generating various QR codes for paper and fabric based materials. (Fabric is used in this case to explore the future possibility of implementing the dMarkers with clothes) We present the successful detection of these dMarker QR codes by a third party QR code reader on a smartphone. Upon discussion of these results and some limitations of the current technologies, we move on to discuss various applications including the use of this technology in AR as a novel dimension for interaction in Augmented Reality.