Next-generation wireless networks demand scalable and low-latency distributed learning frameworks that can operate across diverse communication environments. This paper proposes CFVLC, a novel large-scale Federated Learning (FL) framework over a resource-constrained wireless network, involving outdoor users connected through Cell-Free massive MIMO (CF-mMIMO) system, and indoor users distributed across multiple environments supported by Visible Light Communication (VLC) technology. To address the extreme system heterogeneity from different environments, CFVLC introduces an optimization problem that jointly performs device selection and resource allocation to minimize training latency across the network. We propose a two-stage heuristic solution that intelligently balances user participation across environments with varying resources and channel conditions. Extensive simulations demonstrate that our approach significantly improves training latency and model performance compared to conventional schemes, highlighting the benefits of integrating outdoor CF-mMIMO and indoor VLC technologies for large-scale FL.

In Singapore's polyclinics, touch-based self-service kiosks are widely used for administrative functions such as appointment scheduling and billing. However, these systems pose accessibility challenges for elderly patients, often resulting in increased staff workload and longer wait times. This paper presents VoCare AI, a voice-first conversational assistant designed to streamline healthcare administrative tasks through multiagent workflows powered by Large Language Models (LLMs). The system integrates a graph-based orchestration framework (LangGraph) with specialized agents that manage NRIC-based identity verification, appointment handling, billing queries, and general FAQ responses. It features a fine-tuned Automatic Speech Recognition (ASR) model adapted for Singapore-accented English (Singlish), trained on the IMDA National Speech Corpus to handle accent variability, code-switching, and elderly speech patterns. Evaluation results show that the best-performing ASR model achieved a Word Error Rate (WER) of 22.22%, while the Retrieval-Augmented Generation (RAG) module demonstrated strong performance with 1.00 groundedness and 0.92 retrieval relevance. By focusing on Singapore's linguistic diversity and incorporating localized authentication mechanisms, this work demonstrates how AI-driven conversational agents can enhance accessibility and operational efficiency in public healthcare environments. The complete source code for this work is publicly available at: https://github.com/DerrickLJH2000/langgraph-voice-assistant.

Person re-identification (ReID) aims to retrieve images of the same individual across disjoint camera views in surveillance networks, where resource constraints and privacy concerns often make traditional deep learning approaches impractical. This paper presents a lightweight, training-free ReID framework based on structural semantic graphs, designed for deployment in resource-constrained environments. Each person is represented by a multi-layer graph encoding body parts, fashion items, and dominant colors, extracted using pretrained pose estimation, fashion detection, and color clustering models. Without requiring any task-specific fine-tuning or GPU acceleration, our method enables interpretable matching through a hybrid similarity function that blends cosine similarity over weighted features with Jaccard similarity over attribute sets. Evaluated on the DukeMTMC-reID dataset, our method achieves 27.06% Rank-1 accuracy and 8.10% mAP, outperforming attribute-based baseline and random retrieval method. Further experiments validate that our hybrid similarity computation enhances retrieval performance and confirm that our bag-of-features graph encoding offers an effective solution for training-free ReID in resource-constrained environments.

This paper proposes a learnable metric to measure the reconstruction quality of hyperspectral images obtained by computational hyperspectral imaging. Computational hyperspectral imaging aims to obtain low-cost hyperspectral images through consumer camera. While many hyperspectral reconstruction models have been developed for this purpose, conventional image and spectral quality metrics are insufficient to measure the scientific value of the reconstructed HSI cube. This paper proposes an adaptive quality fusion metric (AQF), adaptively aggregating the quality measures from point-wise, spatial-wise and spectral-wise aspects to assess the scientific value preserved by the reconstructed HSI. The proposed AQF metric uses weight parameters generated by a modified hypernetwork to determine the contribution for the three aspects given paired of groundtruth HSI and reconstructed HSI. Experimental results show its compatibility with existing metrics while accurately measuring the scientific information retained by the reconstructed HSI for hyperspectral applications.

Deepfake technology, which utilizes advanced artificial intelligence to create or manipulate multimedia content, presents significant challenges by obscuring the distinctions between reality and fiction. This phenomenon can lead to severe consequences such as misinformation and deception, particularly in culturally diverse regions like Southeast Asia. In response, this paper aims to enhance deepfake detection capabilities specifically for the Southeast Asian context, with a focus on Singapore, utilizing the Trusted Media Challenge (TMC) dataset. We employ transfer learning to fine-tune existing models with region-specific data and explore various layer freezing strategies to optimize performance. Additionally, we assess the effectiveness of transfer learning against the complete retraining of models to identify the most resource-efficient practices for improving detection capabilities. Our findings reveal that targeted fine-tuning of specific layers can enhance model performance in identifying regional deepfake, providing a balance between computational efficiency and detection accuracy. This research contributes to the development of robust, region-specific deepfake detection methods, which are crucial for combating the evolving threats posed by deepfake technology. We have developed a web application using our trained model, the web application and the source code are available at https://github.com/ict-at-sit/deepfake-detection-app.

This paper introduces a fine-tuned Large Language Models (LLMs) for detecting scam calls, a response to the growing severity of telecommunication fraud in Singapore, which has seen a consistent increase in scam incidents over recent years. Our approach significantly enhances scam detection capabilities through the strategic augmentation of existing datasets with generative AI. This process involves consolidating multiple datasets and employing generative models to synthesize culturally relevant scam scenarios specific to the Singapore context. We use the augmented dataset to fine tune pretrained LLMs, including GPT-2 and Llama. We conducted extensive experiments to compare the effectiveness of traditional machine learning (ML) models, deep learning (DL) techniques, transformers, and our fine-tuned LLMs. The results clearly show that our fine-tuned LLM outperforms other models in terms of inference accuracy while maintaining acceptable inference times, thereby establishing itself as a practical tool for real-time scam detection. The source code of this work is made publicly accessible at https://github.com/ict-at-sit/ScamDetector.

We introduce Hyper-Skin, a hyperspectral dataset covering wide range of wavelengths from visible (VIS) spectrum (400nm - 700nm) to near-infrared (NIR) spectrum (700nm - 1000nm), uniquely designed to facilitate research on facial skin-spectra reconstruction.By reconstructing skin spectra from RGB images, our dataset enables the study of hyperspectral skin analysis, such as melanin and hemoglobin concentrations, directly on the consumer device. Overcoming limitations of existing datasets, Hyper-Skin consists of diverse facial skin data collected with a pushbroom hyperspectral camera. With 330 hyperspectral cubes from 51 subjects, the dataset covers the facial skin from different angles and facial poses.Each hyperspectral cube has dimensions of 1024$\times$1024$\times$448, resulting in millions of spectra vectors per image. The dataset, carefully curated in adherence to ethical guidelines, includes paired hyperspectral images and synthetic RGB images generated using real camera responses. We demonstrate the efficacy of our dataset by showcasing skin spectra reconstruction using state-of-the-art models on 31 bands of hyperspectral data resampled in the VIS and NIR spectrum. This Hyper-Skin dataset would be a valuable resource to NeurIPS community, encouraging the development of novel algorithms for skin spectral reconstruction while fostering interdisciplinary collaboration in hyperspectral skin analysis related to cosmetology and skin's well-being. Instructions to request the data and the related benchmarking codes are publicly available at: https://github.com/hyperspectral-skin/Hyper-Skin-2023.

Stress is an inevitable part of modern life. While stress can negatively impact a person's life and health, positive and under-controlled stress can also enable people to generate creative solutions to problems encountered in their daily lives. Although it is hard to eliminate stress, we can learn to monitor and control its physical and psychological effects. It is essential to provide feasible and immediate solutions for more mental health counselling and support programs to help people relieve stress and improve their mental health. Popular wearable devices, such as smartwatches with several sensing capabilities, including physiological signal monitoring, can alleviate the problem. This work investigates the feasibility of using wrist-based electrodermal activity (EDA) signals collected from wearable devices to predict people's stress status and identify possible factors impacting stress classification accuracy. We use data collected from wrist-worn devices to examine the binary classification discriminating stress from non-stress. For efficient classification, five machine learning-based classifiers were examined. We explore the classification performance on four available EDA databases under different feature selections. According to the results, Support Vector Machine (SVM) outperforms the other machine learning approaches with an accuracy of 92.9 for stress prediction. Additionally, when the subject classification included gender information, the performance analysis showed significant differences between males and females. We further examine a multimodal approach for stress classifications. The results indicate that wearable devices with EDA sensors have a great potential to provide helpful insight for improved mental health monitoring.

Bluetooth Low Energy (BLE) beacons are designed to operate for years on a coin-cell battery. However, the formation of a mesh network, overlaying on the existing Bluetooth Low Energy (BLE) beacons infrastructure, can severely degrade the lifetime of underlying beacons owing to the excessive current drawn by the scanning event. Even though we can sustain the lifetime of the underlying beacon with duty-cycle scanning, such duty-cycle scanning imposes another challenge to the overlay mesh in disseminating the packet. To this end, this paper proposes a novel overlay protocol that: 1) employs duty-cycle scanning to guarantee the lifetime of the underlying beacon, while 2) defining a set of scanning policies to increase the packet dissemination rate through the overlay mesh network. The duty-cycle scanning defines the scanning time slot based on the lowest feasible duty cycle unveiled through a comprehensive analysis of energy consumed by advertising and scanning events. The scanning policies, on the other hand, allow each node to explore all possible time slots before locking their scanning event to a particular time slot that is most likely to hear the incoming packet. Extensive experiments with practical implementation demonstrate the feasibility of our proposed overlay mesh for real-world use cases.

This paper presents a nonlinear location estimation to infer the position of a user holding a smartphone. We consider a large location with$M$number of grid points, each grid point is labeled with a unique fingerprint consisting of the received signal strength (RSS) values measured from$N$number of Bluetooth Low Energy (BLE) beacons. Given the fingerprint observed by the smartphone, the user’s current location can be estimated by finding the top-k similar fingerprints from the list of fingerprints registered in the database. Besides the environmental factors, the dynamicity in holding the smartphone is another source to the variation in fingerprint measurements, yet there are not many studies addressing the fingerprint variability due to dynamic smartphone positions held by human hands during online detection. To this end, we propose a nonlinear location estimation using the kernel method. Specifically, our proposed method comprises of two steps: 1) a beacon selection strategy to select a subset of beacons that is insensitive to the subtle change of holding positions, and 2) a kernel method to compute the similarity between this subset of observed signals and all the fingerprints registered in the database. The experimental results based on large-scale data collected in a complex building indicate a substantial performance gain of our proposed approach in comparison to state-of-the-art methods. The dataset consisting of the signal information collected from the beacons is available online.

This paper presents a human activity recognition (HAR) system with wearable devices. While various approaches have been suggested for HAR, most of them focus on either 1) the inertial sensors to capture the physical movement or 2) subject-dependent evaluations that are less practical to real world cases. To this end, our work integrates sensing in-puts from physiological sensors to compensate the limitation of inertial sensors in capturing the human activities with less physical movements. Physiological sensors can capture physiological responses reflecting human behaviors in executing daily activities. To simulate a realistic application, three different evaluation scenarios are considered, namely All-access, Cross-subject and Cross-activity. Lastly, we propose a Hierarchical Deep Learning (HDL) model, which improves the accuracy and stability of HAR, compared to conventional models. Our proposed HDL with fusion of inertial and physiological sensing inputs achieves 97.16%, 92.23%, 90.18% average accuracy in All-access, Cross-subject, Cross-activity scenarios, which confirms the effectiveness of our approach.

The recent pandemic has brought tremendous changes to everyone’s life, causing stress about losing loved ones, losing jobs, and having changes in sleep or eating habits. This study investigates the feasibility of utilizing Electrodermal Activity (EDA) collected from wearable devices to detect people’s stress. EDA can quantify the changes in sympathetic dynamics by measuring sweat produced by our sweat glands. Currently, the adoption of EDA sensors to commercially off-the-shelf smart-watches is still in the infancy stage, and only a few brands have the EDA sensors implemented into their smartwatch. To facilitate our feasibility study, we need the datasets that contain the EDA signals collected from wearable devices. This paper uses two publicly available datasets containing the EDA signals collected from research-grade wearable devices. We cast the stress detection problem as a binary classification problem and trained the classifiers with three popular machine learning methods: K-Nearest Neighbor, Logistic Regression, and Random Forests. According to experimental results, Random Forests achieves an accuracy of 85.7% to classify stress from non-stress status. The results verified that wearable devices with EDA sensors have the potential to predict stress status.

This paper presents a novel bluetooth low energy (BLE) protocol enabling a BLE node to perform RF fingerprint acquisition by measuring the received signal strength (RSS) from its neighboring nodes and simultaneously broadcast the acquired fingerprint via its advertising packet. However, the fingerprint acquisition and broadcast process in a dense BLE network is very challenging owing to: 1) the likelihood of packet collision; and 2) the length-constrained packet. To this end, we exploit a compressive sensing (CS) framework allowing each node to acquire no more than$M$measurements from a very dense network, in which the number of nodes$N$is far greater than$M$. By aggregating the$M$-dimensional compressed fingerprint vector from$s

This paper presents a novel Q-Learning method in forwarding the proximity sensing information to the remote server through low-power mesh network overlays on Bluetooth Low Energy (BLE) technology. Even though proximity sensing information can be easily monitored with pervasive smartphones, it is almost impossible to remotely monitor this information in a harsh location where it is not easy to access the Internet. With our overlay mesh network, each node should decide to either forward the packet or continue with their own activity when receiving the packet forwarding request, so as to minimize the end-to-end packet delivery latency but maximize the utilization of underlying infrastructures. Reinforcement learning (RL) is employed to train each node to make the above decision. Despite extensive upfront training, there is a high possibility that each node might still encounter an unseen state owing to the network dynamics. However, our novel Q-learning is able to deal with above challenges by constructing a Q-table during online learning, and then use the Q-table as input data for offline training. The experimental results indicate the substantial performance gain of our proposed approach in comparison to the existing Q-learning methods.

While many technologies (e.g., RFID, QR code, etc.) have been developed for items identification, they fail to provide continuous monitoring for items in transit. This paper introduces a Bluetooth Low Energy (BLE) beacon-based system, which can be deployed easily with any off-the-shelf smartphone without modification on the existing infrastructures. However, it is an elusive challenge to achieve a fast item identification and counting involving massive items stacked up inside a confined space (e.g., a container), resulting in an ultra-dense beacon network (UDBN). To this end, we propose a novel beaconing solution capable of informing the receiver about their own presence as well as the presence of their neighboring beacons for identification purpose. Specifically, our proposed solution provides a well-designed yet innovative protocol data unit (PDU) which allows the beacon to encapsulate its neighboring information into its own advertising packet. A prototype consisting of 300 beacons is implemented to demonstrate the feasibility of our proposed solution for real-world applications. The extensive experiment confirm the superiority of our proposed solution in delivering a fast item identification and counting in UDBN.

Bluetooth low-energy (BLE) beacon is a key enabler for smart interaction between the user device and the physical thing, in which the physical thing can actively engage users for interaction via its advertising packet. However, reliability is always an issue for the beacon-based interaction since the beacon employs an unreliable broadcasting approach which provides no way to check if the user device has received the correct packet. We define the sparse observation to describe the phenomenon where the number of packets received by the user device within an arbitrarily small time duration is less than the number of deployed beacons. This article studies the sparse observation causing by the following two factors: 1) the unpredictable environmental variations and 2) the uncontrollable operating conditions of a beacon. An analysis is provided to investigate the interaction reliability in connection with the above two factors. Motivated by the above challenges, a novel solution, which exploits the ambient RF fingerprinting to address the sparse-observation issues, is proposed to enhance the interaction reliability. Our proposed solution is validated with extensive experiments consisting of real data collected from both indoor and outdoor environments. Finally, the feasibility of our proposed solution is demonstrated with a proof-of-concept prototype implemented over multiple physical things.

Lately, there are many works exploited the radio frequency (RF) fingerprint for occupancy detection. However, most works suffer severe performance variations owing to the unreliable received signal strength (RSS). In this paper, we propose a deep learning approach to occupancy detection: 1) an unsupervised denoising-contractive autoencoder (DCAE) is built to learn a robust fingerprint representation from the raw RSS measurements, and 2) a supervised softmax function is added at the last layer for classification. A real testbed with Bluetooth Low Energy (BLE) beacons was built such that we can collect real-world RSS data for experiments. The data were collected via different devices at different times to better reflect environmental variations. The experimental results show that our proposed approach achieves a substantial performance gain in comparison to the conventional machine learning approaches. Specifically, our proposed DCAE is able to reconstruct the noisy and always changing data with less than 0.047 mean square error. Overall, our occupancy detection combining DCAE and softmax classifier achieves sub-room level accuracy for at least 99.3% of the time.

While Bluetooth Low Energy (BLE) beacons have been massively deployed to broadcast their advertising packets to any receivers in their vicinity, it is relatively difficult, if not impossible, for a beacon to report the packet back to the server in the absence of a receiver. This paper proposes a novel BLE-based overlay mesh (BOM) that enables the mesh functionality to existing beacon networks without introducing new infrastructure. However, it is an elusive challenge to jointly manage the beaconing and flooding events. To this end, BOM employs 1) best-effort scheduling (BES) to minimize the packet collision rate (PCR) while scheduling the time slots for beaconing events, and 2) received signal strength (RSS)-based bounded flooding (RBF) to maximize the packet delivery ratio (PDR) for the advertising packet while forwarding the relaying packet across the BOM network. Extensive simulations indicate the substantial performance gain of our proposed approach in comparison to the legacy approaches. Specifically, BES reduces the PCR to 66.67%, whereas RBF improves the PDR for the advertising packet to 52% while maintaining approximately the same PDR for the relaying packet. The practical experiment with a real network testbed further demonstrates the feasibility of BOM.

Device-free occupancy detection is very important for certain Internet of Things applications that do not require the user to carry a receiver. This paper achieves the device-free occupancy detection with RF fingerprinting, which labels each zone with a 2M-dimensional fingerprint vector. Specifically, the fingerprint vector consists of received signal strength (RSS) values measured from M Bluetooth low energy (BLE) beacons and also their corresponding temporal RSS variations. However, the unreliable RSS values caused two common issues with the fingerprint vector: 1) noise and 2) sparsity. To this end, we propose denoising-contractive autoencoder (DCAE) to jointly deal with these two issues, by learning a robust fingerprint prior to device-free occupancy detection. We validate the performance of our proposed DCAE with large-scale real-world datasets. The experimental results indicate the substantial performance gain of our proposed DCAE in comparison with state-of-the-art autoencoders. In particular, the classifier trained using the fingerprints learned by our proposed DCAE is able to maintain at least 90% accuracy when the noise factor or sparsity ratio increases to 0.6 and 0.5, respectively.

Bluetooth low energy (BLE) beacons have been widely deployed to deliver proximity-based services (PBSs) to user's smartphones when users are in the proximity of a beacon. Conventional proximity detection simply uses the received signal strength (RSS) to infer the proximity, and then retrieves the PBS by mapping the beacon ID with the corresponding service in the cloud database. Such an approach suffers two major issues: 1) the severe RSS fluctuation might confuse the smartphone during the detection and 2) a malicious PBS can be delivered by manipulating the same beacon ID. This paper proposes RF fingerprinting to label a beacon with an N-dimensional fingerprint vector, which consists of N RSS values from N deployed beacons. The contribution of our proposed method is twofold: 1) we infer the proximity based on the fingerprint vector instead of relying solely on the single RSS value and 2) we retrieve the PBS by mapping the fingerprint vector instead of the hard-coded beacon ID. The challenge with our proposed approach is the incomplete fingerprint observation during real-time detection, resulting in an underdetermined proximity detection problem. To this end, we exploit the compressive sensing (CS) approach based on the differential evolutional algorithm to address such an underdetermined problem. Extensive simulations with realworld datasets show that our proposed approach outperforms the legacy machine learning techniques with substantial performance gains.

Lately, Bluetooth Low Energy (BLE) beacon has attracted a lot of interests for its capabilities in enhancing the interaction between smart things in the Internet of Things (IoT) ecosystem via proximity approach. Even though Proximity sensing is capable of delivering a correct interaction, it might have a problem for explicit interaction when exact distance estimation is required. Considering those interactive applications which are distance-dependent, this paper proposed an optimized support vector machine (O-SVM) on the cloud for distance estimation and a Kalman filter (KF) on the edge to obtain a near true RSS value from a list of RSS measurements. Four benchmark functions (i.e., two from Industries and two Machine Learning Techniques) have been used for performance evaluation. Simulation with real signal samples was conducted to verify the performance of our proposed algorithm. Besides examining the performance gain of our proposed solution over the four benchmark functions, we also implemented the proposed solution on a smartphone for practical testing to demonstrate its feasibility. The proposed solution not only outperforms the rest with significant performance gain, i.e., > 50% error reduction compared to the benchmark functions. Furthermore, practical implementation verified that our proposed approach is able to return the estimate distance in less than 1s, such real-time response is desirable for many delay- sensitive applications.

While the Internet of Things (IoT) is driving a transformation of current society toward a smarter one, new challenges and opportunities have arisen to accommodate the demands of IoT development. Low power wireless devices are, undoubtedly, the most viable solution for diverse IoT use cases. Among such devices, Bluetooth low energy (BLE) beacons have emerged as one of the most promising due to the ubiquity of Bluetooth-compatible devices, such as iPhones and Android smartphones. However, for BLE beacons to continue penetrating the IoT ecosystem in a holistic manner, interdisciplinary research is needed to ensure seamless integration. This paper consolidates the information on the state-of-the-art BLE beacon, from its application and deployment cases, hardware requirements, and casing design to its software and protocol design, and it delivers a timely review of the related research challenges. In particular, the BLE beacon's cutting-edge applications, the interoperability between packet profiles, the reliability of its signal detection and distance estimation methods, the sustainability of its low energy, and its deployment constraints are discussed to identify research opportunities and directions.

The emergence of Bluetooth low energy (BLE) beacons has promoted the development of proximity-based service (PBS), which is a context-aware application delivered subject to the Proximity of Interest (PoI). Most commercial applications use the sequential proximity detection with a fixed scanning mechanism to identify the target PoI. Such sequential execution, though is able to produce reliable detection, suffers severe performance degradation especially when the number of deployed beacons in the vicinity increases. To understand the effects of dense deployment, we conduct an empirical investigation and derive the statistical properties of both received signal strength (RSS) and signal inter-arrival time. In light of the statistical insights, this paper proposes a high resolution proximity detection using an adaptive scanning mechanism fusion with a spontaneous Differential Evolution (AS+sDE). This novel approach enables the receiver to adapt its scanning duration conditioned on the deployment density and make an almost spontaneous detection in parallel with the scanning. The feasibility of the proposed approach is verified by both simulations and real-world implementations. For a density of$\leq 5\ beacons/m^2$,AS+sDEachieves a superior performance with a high accuracy rate, i.e., on average$<1s$is spent to guarantee at least 90 percent accuracy.

Presents corrections to the paper, “High resolution beacon-based proximity detection for dense deployment", (Ng, P.C., et al), IEEE Trans. Mobile Comput., vol. 17, no. 6, pp. 1369–1382, Jun. 2018.

Existing interactive systems suffer from low user engagement due to their passiveness and steep learning curve. To address these issues, this paper presents an interactive framework, Notify-and-Interact, which leverages the Bluetooth low energy (BLE) beacon infrastructure to notify and a smart-phone to interact, such that it transforms a passive interactive system into an active one. The proposed framework is demonstrated in the Ping Yuan and Kinmay W Tang Gallery, where a series of wildlife artworks are exhibited. Engagement conversion rate is measured, and users' quality of experience (QoE) is surveyed through likert assessment. Artworks with Notify-and-Interact outperforms the QR code with a high engagement conversion rate at the interaction stage, i.e., 86% over 53%, and an average engagement time of 55.67s over 28.69s, respectively. The mean opinion score (MOS) shows that around 80% of the users expressed high satisfaction with the installed Notify-and-Interact framework in the gallery.

A proximity-based service (PBS) leverages the estimated proximity to provide users the accessibility to object or location restricted service. This paper exploits the interaction between Bluetooth Low Energy (BLE) Beacon and smartphone to set forth the fundamental building block of a beacon-based PBS system. In real-world scenarios, a beacon-based PBS system might suffer from sparse conditions when some beacons malfunction or beacons can only be deployed in a few specific positions. Motivated by such limitations, a similarity filter extended with compressive sampling matching pursuit (SF-CoSaMP) is proposed to ensure the reliability of proximity detection under such sparse conditions before smartphone proceed to retrieve the corresponding PBS. An extensive simulation with large volume of collected data has been conducted and the results prove the reliability of the proposed algorithm with high detection accuracy in an environment with sparse deployment.

The meeting of pervasive screens and smart devices has witnessed the birth of screen-smart device interaction (SSI), a key enabler to many novel interactive use cases. Most current surveys focus on direct human-screen interaction, and to the best of our knowledge, none have studied state-of-the-art SSI. This survey identifies three core elements of SSI and delivers a timely discussion on SSI oriented around the screen, the smart device, and the interaction modality. Two evaluation metrics (i.e., interaction latency and accuracy) have been adopted and refined to match the evaluation criterion of SSI. The bottlenecks that hinder the further advancement of the current SSI in connection with this metrics are studied. Last, future research challenges and opportunities are highlighted in the hope of inspiring continuous research efforts to realize the next generation of SSI.