Predicting the future location of mobile objects reinforces location-aware services with proactive intelligence and helps businesses and decision-makers with better planning and near real-time scheduling in different applications such as traffic congestion control, location-aware advertisements and monitoring public health and well-being. Recent developments in smartphone and location sensors technology and the prevalence of using location-based social networks alongside the improvements in AI and machine learning techniques provide an excellent opportunity to exploit massive amounts of historical and real-time contextual information to recognise mobility patterns and achieve more accurate and intelligent predictions. This unique survey provides a comprehensive overview of the next useful location prediction problem with context-awareness and the related studies. First, we explain the concepts of context and context-awareness and define the next location prediction problem. Then we analyse more than 30 studies in this field concerning the prediction method, the challenges addressed, the datasets and metrics used for training and evaluating the model and the types of context incorporated. Finally, we discuss the advantages and disadvantages of different approaches, focusing on the usefulness of the predicted location and identifying the open challenges and future work on this subject.

In Smart City applications, urban mobility involves complex interactions between traffic infrastructure, diverse road users, and physical environment. This paper addresses the limitations of conventional scene modeling methods that often fail to capture the varied and volatile nature of urban road scenes, particularly in representing the dynamic situations that unfold within them. Inaccurate situation representation hinders precise depiction of scene evolution, limiting our ability to understand and respond effectively to complex urban situations. This paper addresses these challenges by presenting the novel concept of the Context-Aware Scene Graph (CSG) for representing situations used in reasoning applications for enhancing safety and efficiency in urban environments, particularly for bicycle riders. CSG integrates multi-modal data, including ontological knowledge, sensor data, and images, to provide a comprehensive representation of urban road situations, enabling informed decision-making. This paper also validates the effectiveness of the proposed approach using real- world IoT datasets and camera images, with a focus on the bicycle dooring use case. The results outperform existing scene modeling methods by accurately representing situations, including those previously overlooked. The approach also ensures consistent representation, completeness, and captures transitions between situations, including causal relations. These findings highlight our approach's effectiveness in improving road safety, efficiency, and urban life quality through enhanced scene understanding.

Predicting situations in real-time applications is non-trivial. Fusing and incorporating the plethora of heterogeneous context information from many sources in the ecosystem that a user resides in to derive their situation is an expensive and time-consuming process. Yet context is useful only when a user can effectively make use of it in time and reliably. In this paper, using a proactive cyclist hazard alerting scenario, we propose a mechanism to proactively cache context information, so that cyclists are alerted of impending hazards before they might even occur. Our novel approach, which is capable of caching reliable predictive context information has significantly reduced the time to deliver context by 91% and the cost by 80%. We ensure the reliability of predictive cached context using a cross-verification routine that the false-positive rate tends to zero. The context cache is structured hierarchically such that our novel proactive context caching mechanism is capable of caching all low-level to high-level pieces of context, unlike any previous approaches.

Real-time applications increasingly rely on context information to provide relevant and dependable features. Context queries require large-scale retrieval, inferencing, aggregation, and delivery of context using only limited computing resources, especially in a distributed environment. If this is slow, inconsistent, and too expensive to access context information, the dependability and relevancy of real-time applications may fail to exist. This paper argues, transiency of context (i.e., the limited validity period), variations in the features of context query loads (e.g., the request rate, different Quality of Service (QoS), and Quality of Context (QoC) requirements), and lack of prior knowledge about context to make near real-time adaptations as fundamental challenges that need to be addressed to overcome these shortcomings. Hence, we propose a performance metric driven reinforcement learning based adaptive context caching approach aiming to maximize both cost- and performance-efficiency for middleware-based Context Management Systems (CMSs). Although context-aware caching has been thoroughly investigated in the literature, our approach is novel because existing techniques are not fully applicable to caching context due to (i) the underlying fundamental challenges and (ii) not addressing the limitations hindering dependability and consistency of context. Unlike previously tested modes of CMS operations and traditional data caching techniques, our approach can provide real-time pervasive applications with lower cost, faster, and fresher high quality context information. Compared to existing context-aware data caching algorithms, our technique is bespoken for caching context information, which is different from traditional data. We also show that our full-cycle context lifecycle-based approach can maximize both cost- and performance-efficiency while maintaining adequate QoC solely based on real-time performance metrics and our heuristic techniques without depending on any previous knowledge about the context, variations in query features, or quality demands, unlike any previous work. We demonstrate using a real world inspired scenario and a prototype middleware based CMS integrated with our adaptive context caching approach that we have implemented, how realtime applications that are 85% faster can be more relevant and dependable to users, while costing 60.22% less than using existing techniques to access context information. Our model is also at least twice as fast and more flexible to adapt compared to existing benchmarks even under uncertainty and lack of prior knowledge about context, transiency, and variable context query loads.

Context awareness adds intelligence to and enriches data for applications, services and systems while enabling underlying algorithms to sense dynamic changes in incoming data streams. Context-aware machine learning is often adopted in intelligent services by endowing meaning to Internet of Things(IoT)/ubiquitous data. Intelligent transportation systems (ITS) are at the forefront of applying context awareness with marked success. In contrast to non-context-aware machine learning models, context-aware machine learning models often perform better in traffic prediction/classification and are capable of supporting complex and more intelligent ITS decision-making. This paper presents a comprehensive review of recent studies in context-aware machine learning for intelligent transportation, especially focusing on road transportation systems. State-of-the-art techniques are discussed from several perspectives, including contextual data (e.g., location, time, weather, road condition and events), applications (i.e., traffic prediction and decision making), modes (i.e., specialised and general), learning methods (e.g., supervised, unsupervised, semi-supervised and transfer learning). Two main frameworks of context-aware machine learning models are summarised. In addition, open challenges and future research directions of developing context-aware machine learning models for ITS are discussed, and a novel context-aware machine learning layered engine (CAMILLE) architecture is proposed as a potential solution to address identified gaps in the studied body of knowledge.

The success of advanced Background Subtraction (BGS) algorithms for dynamic backgrounds is mostly in land scenes such as those in CDNet benchmarks; few handle underwater scenes, since existing underwater video datasets are either in low resolution or with only static backgrounds. Consequently, the lack of reliable BGS support makes supervised Moving-Objects Segmentation (MOS) algorithms much harder to adapt to unknown underwater scenes because of the diversities of the aquatic environments. For example, those trained by the latest underwater image dataset, SUIM, are ineffective in the underwater videos of our experiments.The underwater waving objects (e.g., plants) often render existing BGS algorithms inaccurate due to three types of errors: (a) incompletely identified MOs (Moving Objects), (b) missing MOs, and (c) falsely identified MOs. In this paper, we propose a novel Clustering-Based Multi-State Background Representation (CBMSBR) model to learn and represent the repeatable patterns of waving movements in k background states (i.e., color ranges) per pixel, and thus accurately subtract the background waving objects to reduce these errors. In addition, we further develop a CBMSBR+ model to remove the more challenging background objects in unusually large magnitudes of wavings. Both models come from a basic observation: the video pixels in the waving zones repeatedly switch among multiple background states; e.g., a pixel switches among water state, plant 1 state, and plant 2 state. To test our proposed models, we create experiments using three types of challenging scenarios that each often covers at least two error types, i.e., the scattered MOs scenario covering (b) and (c), the crowded MOs scenario covering (a) - (c), and the slow MOs scenario covering (a) and (c). Experiments on these scenarios demonstrate the accuracy, effectiveness, and efficiency of our models and their applications in MOS improvements.

Smart energy management based on the Internet of Things (IoT) aims to achieve optimal energy utilization through real-time energy monitoring and analyses of power consumption patterns in IoT networks (e.g., residential homes and offices) supported by wireless technologies - this is of great significance for the sustainable development of energy. Energy disaggregation is an important technology to realize smart energy management, as it can determine the power consumption of each appliance from the total load (e.g., aggregated data). Also, it gives us clear insights into users’ daily power-consumption-related behaviours, which can enhance their awareness of power-saving and lead them to a more sustainable lifestyle. This paper reviews the state-of-the-art algorithms for energy/power disaggregation and public datasets of power consumption. Also, potential use cases for smart energy management based on IoT networks are presented along with a discussion of open issues for future study.

Summary Individuals' right to privacy includes control over access to their location information. With the advent of location‐based services and personal transport services (such as ridesharing), the risk of location privacy breaches is increased greatly. The potential negative effects of location privacy leakages include spam location‐based service flooding, threats to personal safety (such as physical attacks), and intrusion related to access to private places (such as homes and hospitals). Therefore, protecting the privacy of users' real locations is becoming increasingly important. This is often achieved using a pseudo‐location near the real location, but existing pseudo‐location generators, such as NRand and the uniform random method, suffer from statistical inference, which can infer the obfuscation domain to cover the real location. In this paper, we propose an intelligent pseudo‐location recommendation (IPLR) method to reduce the risk of a statistical inference attack. In IPLR, we generate a random substitute of the real location to attract the adversary and thus hide the real location. Then, the pseudo‐location is generated in the neighborhood of the random substitute location following a normal distribution; the random substitute location is changed frequently to confuse attackers. In particular, we define three levels of location privacy, ie, address level, street level, and district level, to evaluate the effectiveness of the IPLR method. Our experimental study using simulation data demonstrates that the proposed IPLR method achieves lower risk of location privacy leakage and higher probabilities of safety in all three levels of location privacy than NRand and the random method. It also demonstrates the effectiveness of the proposed IPLR to balance location privacy and service quality.

Summary The increasing amount of solid waste is becoming a significant problem that needs to be addressed urgently. The reliable and accurate classification method is a crucial step in waste disposal because different types of wastes have different disposal ways. The existing waste classification models driven by deep learning are not easy to achieve accurate results and still need to be improved due to the various architecture networks adopted. Their performance on different datasets is varied, and there is also a lack of specific large‐scale datasets for training. We propose a new combination classification model based on three pretrained CNN models (VGG19, DenseNet169, and NASNetLarge) for processing the ImageNet database and achieve high classification accuracy. In our proposed model, the transfer learning model based on each pretrained model is constructed as a candidate classifier, and the optimal output of three candidate classifiers is selected as the final classification result. The experiments based on two waste image datasets demonstrate that the proposed model achieves 96.5% and 94% classification accuracy and outperforms several counterpart methods.