On behalf of the technical program committee, we are delighted to invite you to participate in the 2024 IEEE 8th International Conference on Smart Internet of Things (IEEE SmartIoT 2024), to be held in Shenzhen, China on November 14-16, 2024.
With the rapid advancement of computer technol-ogy, an increasing number of intelligent devices are being widely applied across various societal domains. Human motion prediction, in particular, holds significant value in human-computer interaction and intelligent surveillance, effectively enhancing the intelligence of these devices. Diversity prediction, a type of human motion prediction, seeks to forecast various potential future movements of the human body. Given the randomness in generation associated with Variational Autoencoders (VAEs), we employed a Recursive Neural network (RNN)-based VAE for future predictions. Additionally, considering the lack of correlation between random Gaussian samples during resampling, we proposed a novel sampling method to generate more reliable and diverse future predictions. Extensive experiments were conducted on the Human 3.6M [30] and HumanEva-I [31] datasets, yielding quite impressive results.
Scarab beetles, with over 35,000 species, significantly impact agriculture globally. Accurate and timely identification is crucial to control their damage to crops and orchards. This study introduces a method for detecting and identifying scarab species, using a dataset of 588 images covering five common species: Blackgill beetle, Verdigris beetle, Proagopertha lucidula, Brown beetle, and Tawny beetle. We developed an automatic scarab detection method using the Mask-RCNN network with the CBAM module, improving feature extraction and reducing background noise interference. The model's accuracy is further enhanced by integrating multi-source IoT data fusion, leveraging diverse data from IoT devices to refine detection. These devices capture a wide array of environmental variables and scarab behaviors, enriching the model's dataset for improved predictive accuracy. The performance of the enhanced C_Resnet algorithm, which includes these integrations, was thoroughly evaluated. Results indicate a significant improvement in detection accuracy, with an average rate of 98.6%, marking a 14% increase over the standard Mask-RCNN algorithm. Comparative analysis with other prevalent models like Yolov3, Solov2, and GNNET shows mAP improvements of 11.7%, 44.8%, and 5.6%, respectively. These findings underscore C_Resnet's superior capability in distinguishing between common beetle species.
Human activity recognition is becoming a research hotspot in recent years due to its wide range of application prospects in surveillance, healthcare, human-computer interaction, etc. A wide variety of sensors has been applied in human activity recognition including cameras, Bluetooth/WiFi receivers, embedding sensors in smartphones/smartwatches, etc. Radar can transmit electromagnetic wave signals and receive target echoes; echo signals contain human motion characteristics, which can be used for human activity recognition. Compared to other sensors, radar has the characteristics of a wide detection range, strong penetration ability, non-contact detection ability, etc. It is not affected by light and is not privacy-intrusive to the observed person. Radar-based human activity recognition is gradually attracting the attention of academia and industry. In this paper, we reviewed current research on radar-based human activity recognition from the aspects of applications and challenges. This paper aims to provide comprehensive and detailed guidance for the foreseen research of radar-based activity recognition.
This study addresses a significant gap in the current literature regarding the impact of hyperparameters on the Mamba model's running time and accuracy. To bridge this gap, a series of comprehensive experiments were designed and conducted to assess the effects of various initial hyperparameter settings, specifically focusing on the learning rate, batch size, and SSM state expansion factor, on the Mamba model's performance. Three widely recognized datasets, SST2, IMDB, and AG_NEWS, were utilized for these experimental evaluations. The primary objective of this research is to provide an indepth analysis of how these hyperparameters influence the running time and accuracy of the model. By varying the initial values of the learning rate, the study explored its impact on both the convergence speed and ultimate accuracy of the model. Adjustments to the batch size were examined to understand their implications for computational efficiency and overall model performance. Furthermore, modifications to the SSM state expansion factor were investigated to determine their effects on model capacity and accuracy. Through these experiments, this research aims to provide valuable insights and guidelines for researchers and practitioners working with the Mamba model. Ultimately, this study seeks to achieve an optimal balance between computational efficiency and model accuracy by contributing to a more nuanced understanding of hyperparameter optimization.
Human action recognition plays a pivotal role in the field of computer vision, representing significant technological advancements. Through the analysis of videos or continuous image sequences, accurate identification of human actions and behaviors can be achieved, thereby fostering innovation in modern society. Despite challenges such as data diversity, individual differences, and environmental factors, researchers continuously strive to innovate for more precise methodologies. This article utilizes the YOLOv7-w6 model and algorithm to achieve precise detection of key points in images and showcases remarkable recognition rates and accuracy levels, providing a valuable technical reference for further research and practical applications.
Unmanned Aerial Vehicles (UAVs) have been widely applied in various fields due to their flexibility, low cost, and rapid deployment capabilities. UAV s can be equipped with electromag-netic jamming devices to disrupt target areas, even in complex or hazardous environments. In this paper, we propose a comprehen-sive optimization scheme for the problem of multi-UAV-assisted electromagnetic jamming. This scheme models the problem as a Markov game and employs a multi-agent reinforcement learning (MARL) algorithm to schedule UAVs, aiming to achieve superior electromagnetic jamming effects while ensuring safer positional deployment. We first construct a model for the UAV-assisted electromagnetic jamming scenario based on realistic conditions. Then, we address the problem of electromagnetic jamming on targets by a UAV swarm and propose an improved algorithm based on existing MARL algorithm. This improvement effectively mitigates the issue of agent inertia, making the algorithm more suitable for our specific problem. Experimental results show that our proposed algorithm outperforms existing algorithms in optimizing the electromagnetic jamming effectiveness and reducing the risk factor.
The rapid proliferation of the Internet of Things (IoT) has resulted in an exponential surge in data generation, necessitating robust and secure platforms for data transactions. Blockchain technology, characterized by its immutability and de-centralized architecture, emerges as a promising solution offering enhanced transparency and security. This survey provides an in-depth exploration of blockchain-based incentive mechanisms for IoT applications, systematically categorized into Shapley value, Stackelberg game, and auction model. Each category is examined through its theoretical underpinnings, analytical methodologies, and specific advantages within IoT. By discussing the unique challenges and opportunities at the convergence of blockchain and IoT, this paper seeks to furnish a comprehensive guide for future endeavors in blockchain-enabled IoT ecosystems.
Clock synchronization is a critical technology in wireless sensor networks, which provides a foundation for various operations between network nodes. The accuracy of synchronization and network energy consumption are essential factors that synchronization protocols must consider. The accuracy of clock synchronization can be improved through information exchange between nodes. However, frequent information exchange will increase the communication overhead of the network as the number of nodes increases. Therefore, researchers have proposed implicit synchronization, a low-energy synchronization mechanism. In implicit synchronization, nodes only need to listen to the information transmission between the reference node and the active node to synchronize their clocks to the reference node, thereby reducing energy consumption. However, in current implicit synchronization scenarios, the timestamp is mainly used for synchronization, which reduces clock synchronization accuracy due to the random delay of timestamps. To address this issue, the paper proposes a joint estimation algorithm for clock skew and offset that combines correlation detection and implicit synchronization. The algorithm ensures high clock synchronization accuracy while reducing energy consumption as the entire process is completed at the physical layer. The simulation results show that the proposed algorithm is superior to the traditional algorithm regarding clock synchronization accuracy in the implicit synchronization scenario.
Vision Transformer (ViT) architectures have become widely adopted in the field of computer vision. However, pure ViT models often underperform on small-scale datasets due to their reliance on large amounts of data for meaningful feature learning. To tackle this limitation, we propose an effective method to enhance ViT performance on small datasets by incorporating locality information from Convolutional Neural Networks (CNN). Given the strength of CNN in local feature extraction and ViT in global feature learning, we introduce a fusion framework that leverages pre-trained CNN features to guide ViT models. This integration combines CNN's local inductive bias with ViT's global receptive field, resulting in improved performance on small-scale datasets. Specifically, our approach enhances the locality awareness of ViT without compromising their global feature representation capabilities. We provide a detailed implementation of this fusion method and evaluate its effectiveness through experiments on the Tiny ImageNet dataset [3]. Our results demonstrate significant performance improvements across various ViT models, including an 11.30% improvement for ConViT, a 20.82% improvement for T2T, and a 19.62% improvement for PiT.
The prediction of basketball game outcomes, particularly for the NCAA (National Collegiate Athletic Association) tournament, has garnered significant attention from researchers due to its complexity and practical applications for fans, bookmakers, and team management. Traditional prediction methods often rely on expert opinions and statistical models, but recent advancements in machine learning and big data have shifted the focus towards data-driven approaches. This paper explores the application of the TabNet algorithm, a novel deep learning model designed specifically for tabular data, to predict the outcomes of NCAA 2024 basketball games. By leveraging its unique attention mechanism for feature selection, TabNet demonstrates improved prediction performance. The study includes comprehensive data analysis and feature engineering to ensure high-quality input features, thereby enhancing the accuracy of the predictions. Our results highlight the potential of TabNet in sports analytics, offering valuable insights into model interpretability and practical utility.
Optimizing the underlying continuous volumetric scene function using sparse input view collections is crucial for applications in modern industrial production and virtual reality technologies. However, existing technologies in this domain continue to exhibit significant shortcomings in specific areas. Therefore, this paper proposes a method that leverages neural radiance fields as a scene representation, employing an efficient and robust backend penalty loss algorithm to supervise model convergence. This approach achieves high-quality 3D reconstruction from images captured from surrounding views, surpassing existing methods that rely on explicit volumetric representations. Additionally, CL-NeRF incorporates a straightforward tracking and mapping system that adjusts based on the underlying point cloud representation of the neural radiance field. This method is independent of scene size and avoids issues related to sub-map capacity, making it suitable for reconstructing larger scenes. CL-NeRF offers several advantages over previous models, including faster rendering and higher-quality optimization.
China's rapid economic development in recent years has led to a rise in the consumption level of the population. Concurrently, the demand for tobacco products has also increased, resulting in an expansion of the tobacco retail sector. This has given rise to the issue of retail store location. This paper proposes a CNN (convolutional neural network) model that takes into account the spatial features of site selection factors. The model considers four factors: population, POI (points of interest), traffic, and competition. Ultimately, the model predicts the market demand of the region, a value that can be used to assist in site selection decisions. In this paper, the proposed CNN model is compared with several classical machine learning models. The data from Zhanjiang was chosen to train the model in the experiment. The results indicate that the CNN model is the most effective. The model offers insights and methodologies for addressing the site selection problem.
In this research, we delve into the integration of BERT, a state-of-the-art pre-trained language model, with Word2Vec, a traditional word embedding technique, to augment sentiment analysis capabilities. Our approach involves leveraging BERT's contextual understanding for sentiment classification and combining it with Word2Vec embeddings to potentially boost performance. The methodology encompasses data preparation involving the extraction and preprocessing of 58,634 mobile phone reviews from JD.com, hybrid model training, and rigorous evaluation using metrics such as accuracy, precision, recall, and F1 scores. The results indicate that BERT demonstrates excellent standalone sentiment analysis performance, and its integration with Word2Vec embeddings leads to improved accuracy and provides a more nuanced sentiment classification. This study contributes to the field by showcasing the effectiveness of hybrid models in capturing the complexities of textual sentiment.
As robots become increasingly integrated into daily life, the demand for natural and effective social interactions between humans and robots continues to rise. However, existing research on human-robot interaction predominantly focuses on humanoid robots, with limited attention given to the emotional expression capabilities of non-humanoid robots. This study proposes a motion-based emotional expression framework tailored to non-humanoid robots, particularly wheeled robots. By integrating motion behaviors with emotional cues, such as kaomoji, the framework explores the potential for these robots to express emotions. Experimental results indicate that motor behaviors alone can convey emotions with an average accuracy rate of 36.81 % across seven distinct classes. Additionally, Kaomoji are limited in their ability to convey complex or ambiguous emotions, with a accuracy rate of only 40 %. However, combining motor behaviors with kaomoji significantly enhances both the accuracy and diversity of emotional expression, achieving an average accuracy rate of 80 %. This combination not only improves accuracy but also adds variety to emotional expressions, providing an effective method for conveying emotions without relying on facial or vocal expressions. The proposed framework addresses a critical gap in the research on emotional expression in non-humanoid robots, providing a fresh perspective on emotional communication technologies.
Wearable devices have rapidly evolved as essential tools for continuous health monitoring. However challenges persist in achieving high-accuracy physiological data acquisition due to issues such as suboptimal sensor placement and environmental interference. To overcome these limitations, we introduce a wearable ear sensor designed to offer comprehensive and reliable physiological monitoring. The sensor integrates different sensing modalities, including electroencephalography (EEG) and electromyography (EMG), within a compact, user-friendly form optimized for long-term wear. Utilizing biocompatible polydimethylsiloxane (PDMS) for in-ear electrodes and a lightweight reference electrode, the device ensures stable contact and signal fidelity. The designed signal acquisition system is able to realize low-noise amplification and adaptive signal processing, enabling accurate detection of low-amplitude bioelectrical signals. Experimental validation demonstrates the sensor's capability to capture and analyze physiological signals associated with various activities, such as eye movements and respiratory rhythms. The findings highlight the potential of ear-worn sensors in advancing wearable health monitoring technologies, with applications in both clinical and consumer health contexts.
Emerging autonomous vehicles require significant augmentation and modifications to enable effective use by passengers with disabilities. We consider the problem of designing autonomous public transportation for passengers with visual impairment. Our proposed architecture augments the vehicule with robotic arms to assist the passenger in embarking and disembarking, and a smartphone app to coordinate the assistance with the vehicular technology. We identify critical challenges involved and our approaches to address them. We develop a simulation framework to demonstrate the viability of the system.
The recognition of sign language is of great significance to remove the communication barrier between the speech/hearing impaired and the public. Current intuitive glove solutions are limited to recognize discrete simple gestures (i.e., numbers, letters, or words), far from satisfying the need of the signers' daily conversation. With cutting-edge artificial intelligence (AI), this paper presents a segmentation-enhanced Convolutional Neural Network (CNN) for continuous and complex sign recognition. The proposed sign language recognition and communication system comprising sensing gloves, deep learning block, and virtual reality (VR) communication interface. Both non-segmentation and segmentation assisted deep learning model achieve the recognition of 20 words (93.25% and 82.81%) and 10 sentences (95.5% and 82%) that are picked up as demonstration. To overcome the incapability of recognizing new sentences, the segmentation approach is hence proposed to split entire sentence signal into word fragments, then the deep learning model memorizes all fragments and inversely reconstructs and recognizes sentences. New/never-seen sentences created by new-order word fragment recombination can be recognized with segmentation, largely improving the practicality of the sign language interpretation system. Finally, the AI-integrated VR communication interface allows the bidirectional interaction between signers and non-signers.
Person re-identification aims to recognize a target pedestrian across non-overlapping camera views based on source information. The Internet of Things (IoT) provides a wide range of application scenarios for pedestrian re-identification technology-smart city management, resource optimization, and multi-source data fusion. It is crucial for IoT applications like intelligent video surveillance but remains challenging due to factors like low image resolution, varying angles, lighting changes, and occlusion. In this paper, we propose a multi-task learning approach that integrates text information to enhance recognition accuracy. Using a dual-stream Transformer encoder, we extract both image and text features. To improve feature interaction and learning, we perform multimodal interaction for fine-grained alignment and share feature for modality-invariant feature representation and learning. Our method, TFTI, outperforms state-of-the-art techniques in person re-identification, as validated on the CUHK-PEDES dataset.
This article introduces Fed-NAS, a novel framework that combines Federated Learning (FL) with Efficient Neural Architecture Search (ENAS) to allow for distributed neural architecture optimization while maintaining privacy. In Fed-NAS, various clients run the ENAS algorithm on their own data and then send the results to a central server, which combines them into a global weight dictionary. This dictionary helps clients transfer weight efficiently, encouraging sharing of knowledge and speeding up training. The framework's effectiveness is illustrated through its performance on the CIFAR-10 dataset, surpassing that of traditional ENAS methods. Furthermore, the future inclusion of blockchain technology will improve the confidentiality and decentralization aspects of the system. Upcoming research will also be concentrated on expanding Fed-NAS to operate on multiple GPUs and devices to assess its performance and scalability, while also comparing it to traditional fixed-architecture Federated Learning systems to confirm the advantages of dynamic architecture adjustment and parameter sharing.
