In the context of exoskeleton rehabilitation for patients in the later stages of recovery, reducing unnecessary interactive torque and enhancing exoskeleton transparency is important, where patients are encouraged to engage in voluntary movement. This study introduces a novel method known as Periodic Kernelized Movement Primitives (PerKMP) to address this issue. PerKMP is an advanced iteration of the original Kernelized Movement Primitives (KMP) that integrates periodic kernel functions for heightened adaptability. Two distinct PerKMP modules have been developed: the desired trajectory modulation PerKMP, which dynamically plans the desired trajectory according to the walking state, and the joint torque estimation PerKMP, which accurately estimates total joint torque in real-time and adjusts controller stiffness accordingly. A series of experiments with different walking speeds and different subjects were conducted to verify the validity of the method. Through the combined implementation of two PerKMP modules, the average absolute value of the interactive torque and energy per unit distance (EPUD) were reduced effectively. This research broadens the application of imitation learning methods in exoskeletons, and enables real-time control adjustments, thereby presenting a pioneering approach to enhancing the adaptability of exoskeleton rehabilitation systems.

As a part of computer vision, object detection is crucial for traffic management, emergency response, autonomous driving vehicles, and smart cities. Despite significant progress in object detection, detecting small objects remains challenging due to their low resolution, which results in little visual information, difficulty in extracting discriminative features, and susceptibility to interference from environmental factors. To address these challenges, we propose SF-DETR, a new model designed specifically for scenarios with small targets. Firstly, We designed a new backbone network that uses partial channel self-attention to replace the backbone network of RTDETR, which can capture both local and global contextual information for extracting and enhancing input features, thereby enhancing the perception of small targets. Secondly, in order to enhance the feature representation ability of the model and better preserve the details of small objects, we proposed Feature Enhancement and Refinement (FER) module, which incorporates a bidirectional fusion mechanism between high-resolution and low-resolution features, allowing for more comprehensive information transfer between features and further improving the effect of multi-scale feature fusion. Finally, we introduce an efficient IoU method (PIoU) which simplifies the computation, speeds up the convergence, and improves the detection accuracy. SF-DETR significantly improves the detection of small targets, outperforming widely used models on various metrics, while significantly reducing model parameters and computational costs compared to RT-DETR. Compared to RTDETR-R34, our model has improved the mAP@50 and [email protected]:0.95 by 3.3% and 2.4% respectively on the Visdrone2019 test set.

In skeleton-based action recognition, accurately recognizing human actions is challenging due to complex movements, keypoints noise and action similarity. This paper introduces the PoseConv3D-Tafimodel, which combines PoseConv3D with the new TSCA, IFA modules and an integrated learning strategy. The TSCA module improves the temporal modeling of joints and the recognition of complex actions. The IFA module exploits inter-frame differences to distinguish similar actions better. The integrated learning strategy improves the robustness and generalization of the model. Experiments show that PoseConv3D-TAFI improves the top-1 accuracy by 1.4% and 2.02% on the UCF101 and HMDB51 datasets, respectively, and outperforms existing methods in handling complex and similar actions. The model shows strong potential for practical applications of skeleton-based action recognition.

Accurate detection of students' behaviors in class-rooms through video surveillance is crucial for improving teaching quality. This paper presents LAF -DETR, a novel algorithm designed to address challenges such as dense student interactions and multi-scale variations in complex classroom environments. Key contributions include a lightweight convolution module to enhance the extraction of multi-scale features, a feature fusion module to reduce background noise and improve multi-scale feature integration, and a downsampling module to capture compre-hensive information across various behavior scales. Experiments on three public datasets (SCB-Dataset3-S, CrowdHuman, Smart-Classroom-Student-Behavior) show that LAF-DETR outperforms the baseline model by 2.8%, 1.9%, and 2.1 % in Average Precision (AP), respectively, demonstrating its effectiveness in handling occlusion and multi-scale detection challenges.

Lightweight human pose estimation (HPE) has garnered significant attention due to its widespread applications in edge and mobile devices. However, lightweight human pose estimation methods demonstrate limited accuracy when detecting complex movements, thus restricting their practical utility. To address this issue, we propose MGDNet, a novel lightweight network featuring three innovative modules: the GA-Bottleneck module, the MAC-Block module, and the Dual-View Feature Enhancement Module (DV-FEM). The GA-Bottleneck module integrates ghost convolution with multi-dimensional adaptive frequency-aware attention to capture multi-scale frequency characteristics, enhancing robustness for complex actions. The MAC-Block combines multi-receptive field depth convolution with frequency domain analysis to achieve precise joint localization. The DV-FEM leverages complementary local-global information to enhance feature representation for complex actions. Extensive experiments conducted on the COCO and MPII datasets demonstrate that MGDNet achieves state-of-the-art performance among lightweight models, attaining an average precision (AP) of 73.0% on COCO val2017, which represents a 1.6% improvement over Greit-HRNet. On the MPII dataset, MGDNet achieves the highest PCKh score of 87.6%. The proposed MGDNet outperforms existing lightweight approaches by addressing the challenge of low detection accuracy in complex motion scenarios while maintaining comparable computational efficiency.

Sharding technology has become an important direction in blockchain scalability research. However, unreasonable account allocation during the sharding process has also brought about several challenges. First, the imbalance in transaction workloads between shards affects transaction efficiency. Shards with excessive workload can lead to congestion. Second, the high ratio of cross-shard transactions increases the communication overhead of the blockchain. In this paper, we propose FennelChain, a blockchain sharding protocol based on a graph partitioning algorithm, which allocates accounts in a rational manner to improve blockchain performance. We model the blockchain system using graph theory, mapping accounts and transactions in the system to vertices and edges in a graph, thus constructing a blockchain transaction graph. Furthermore, we propose the P-Fennel dynamic account partitioning algorithm, which reassigns accounts in the blockchain transaction graph to reduce the ratio of cross-shard transactions and address the problem of transaction load imbalance between shards. Finally, we conduct extensive experiments, and the results show that the proposed protocol outperforms other baselines in terms of throughput, transaction confirmation latency, and the ratio of cross-shard transactions.

Road maintenance is an indispensable part of the transportation system, ensuring road safety and maximizing road efficiency. The rapid and accurate acquisition of road surface information is crucial for effective road maintenance management. With the development of artificial intelligence, computer vision-based road defect detection methods are becoming popular. In this study, we propose a universal road defect detection model named YOLO-RDD, using You Only Look Once version 8 as the basic framework. Firstly, we introduce a new feature extraction and fusion approach to enhance the interaction between shallow and deep feature information, enabling multi-scale defect detection. Secondly, inspired by dynamic snake convolution, as cracks are the major road defects, we propose a new feature extraction module called DSC-C2f to adapt to the elongated and continuous morphology of cracks. Finally, we establish cross-layer connections between the backbone, neck, and prediction stages and integrate a coordinate attention module that considers inter-channel relationships and long-range positional dependencies. This allows the model to selectively focus on relevant parts of shallow and deep semantic information. We evaluate our model using the dataset provided by the crowdsensing-based road damage detection challenge (CRDDC2022) and further assess its performance on a publicly available crack and sealed crack dataset. The experiments demonstrated that our model achieved the best performance compared to the state-of-the-art (SOTA) models. Compared to the baseline model YOLOv8, our model achieved improvements of 1.2% in Precision, 1.9% in Recall, 1.9% in [email protected], 0.8% in [email protected]:0.95, and 1.6% in F1-Score.

Aspect-based sentiment analysis(ABSA) is a fine-grained sentiment classification task that aims to identify the sentiment polarity of specific aspects in a sentence. Graph convolutional networks(GCN) have recently been widely applied for modeling the associations between aspects and opinion words. However, most studies treat the relationships between all words in the graph equally through dependency analysis, without considering their dependency types, which may lead to the inability to distinguish important relationships. On the other hand, dependency trees can only reveal relationships between words and cannot simulate complex sentence relationships, such as conditional, parallel, and contrast relationships, which are crucial for capturing more accurate sentiment relationships for different aspects. To address these challenges, this paper proposes a Collaborative computation model based on Dependency Types and Constituent trees(CDTC) for aspect-based sentiment analysis. Specifically, collaborative computation of dependency type graph convolutional networks and syntax encoders enhances task efficiency. Utilizing dependency types helps distinguish important relationships, while the phrase segmentation and hierarchical structure of constituent trees provide richer syntactic information. Extensive experiments conducted on three datasets demonstrate that our proposed CDTC model achieves state-of-the-art performance.

Traffic sign recognition is a challenging task for unmanned systems, especially the problem of detecting small targets. During traffic sign recognition (TSR), traffic sign targets are small, represented by a small number of pixels in the image, and lack sufficient detail, making it difficult to detect them using a detector. To solve these problems a new traffic sign target detection algorithm is proposed, which introduces a new convolutional module SPD-Conv to capture the feature information of small target traffic signs more efficiently and replaces the C3 module in the backbone network with a C2f module, which greatly reduces the parameters and computation of the original network and achieves a lighter weight to make the model easier to deploy. A new weighted bi-directional feature pyramid network is used to replace the original feature pyramid network. In addition, a coordinate lightweight attention module is introduced for the small target feature loss problem, and finally, a new loss function is proposed to improve the convergence speed of the model while improving the small target detection accuracy. Extensive experimental results on CCTSDB and tt-100k datasets show that our method is more versatile and superior for traffic sign small target detection compared to several state-of-the-art methods.

Strengthening the supervision of bidding business is an important measure to optimise the business environment of bidding and ensure fair competition among enterprises. The existing blockchain-based bidding platform has already achieved the uploading of simple business, but with the development of the bidding industry, single-modal data storage is not enough to support the credible traceability of business processes. Therefore, to address the problems of multimodal data uploading, storage space limitation and performance in the bidding field, and combined with the node complexity of the bidding scenario, we design and propose a bidding management platform for cloud-chain convergence(BMPC3), in which ChainMaker serves as the implementation platform. In this alliance chain, we meet different business logic requirements by designing specific smart contracts and storage structures. According to different business data types, the ledger storage method is optimised to meet the on-chain storage requirements of multimodal data such as images, videos, PDFs, etc. Meanwhile, we explore the blockchain architecture and system configuration, and propose an optimisation scheme, which effectively improves the system transaction throughput. The final experiment shows that the transaction throughput of the system can reach 3300TPS, which can meet the application requirements.

Predicting the sentiment polarity of aspect terms in sentences is the goal of Aspect-Based Sentiment Analysis(ABSA) task. Graph Convolutional Network(GCN) is used in majority of the ABSA task due to its ability to effectively capture the dependencies among words or entities within sentences. However, it may not work as expected when some sentences have no obvious syntactic structure. To alleviate this issue, we propose a Context-guided and Syntactic Augmented Dual Graph Convolutional Network(CSADGCN) model for the ABSA task. Specifically, we propose a context-guided attention mechanism that captures both global and local information by combining self-attention and aspect-level attention, even though some sentences have no obvious syntactic structure. In addition,we augment the GCN with multiple linguistic features and utilize a biaffine attention module to capture the relationship between words. On three datasets, extensive experimental modifications reveal that our CSADGCN model performs better than the most recent baseline approach.

Due to the complexity of industrial environments, such as construction sites and production workshops, the objects to be detected are easily occluded and perceived as small objects, which poses certain challenges for object detection. To ensure a safe industrial environment, this study adopts YOLOv5 as the basic framework and integrates the depth-to-space convolution module to improve the model’s ability to extract feature information of small targets. Second, the global attention mechanism is incorporated into the network to enhance the global interaction information, reducing the feature information loss, and improving the model performance. Finally, to alleviate the contradiction between classification and regression tasks in object detection, the YOLOv5 head is replaced with a decoupled head to achieve better classification and accelerate model convergence. To improve data diversity and enhance model robustness, we augmented the open-source safety helmet wearing dataset (SHWD) and smoking behavior detection dataset (SBDD). We test the performance of the proposed model (SGD-YOLOv5) on the large and small object detection dataset (SODA-D) and VisDrone2021-DET datasets. Furthermore, its ability to detect small objects was also evaluated. Experiments show that our model outperforms all baseline models on SHWD and SBDD. Compared to the TPH-YOLOv5 on the SODA-D dataset, AP and Recall achieve improvements of 18.3% to 19.2% and 11.9% to 13.3% respectively. On the Visdrone 2021-DET dataset, [email protected] achieved an improvement from 35.45% to 35.70% compared to the state-of-the-art model YOLO-Drone.

In certain specific industrial scenarios, smoking and cellphone usage are strictly prohibited behaviors. In these scenarios, it is crucial to rapidly and accurately detect smoking and cell phone usage, and promptly issue warnings, to ensure industrial safety. The detection of prohibited behaviors using computer vision has gained attention from researchers with the development of artificial intelligence. However, detecting small objects like cigarettes and cell phones in complex backgrounds and varying angles poses a challenge due to their changing shapes. In this paper, we propose a detection model called CB-YOLO, which utilizes YOLOv7 as the baseline model, for detecting smoking and mobile phone usage behavior. Firstly, we propose a new channel space pyramid network called CSPPF that pools features at different scales and introduces scale focusing on improving the perceptual ability of the network to better handle targets at various scales and locations. Secondly, we propose an enhanced feature pyramid called AWBFPN. This introduces additional learnable weight parameters to improve the model's ability to fuse multi-scale features effectively. Finally, we have also proposed a new loss function called Size-IoU. The experimental results demonstrate that our algorithm outperforms the baseline model. Specifically, it achieves a 3.9% improvement in Precision, a 2.1% improvement in Recall, a 6.4% improvement in [email protected], and a 2% improvement in [email protected]:0.95 on the Phone_check dataset. Similarly, on the Smoking dataset, our algorithm achieves a 3.3% improvement in Precision, a 1.9% improvement in Recall, a 1.7% improvement in [email protected], and a 0.3% improvement in [email protected]:0.95.

In the digital society era, the generation frequency of multimodal data is rapidly increasing. Blockchain, recognized as a trusted distributed database technology, provides a new solution for the trustworthy storage and efficient management of multimodal data. However, blockchain systems support only query that use transaction hash values as keywords and cannot directly leverage the content features of multimodal data, leading to generally low query efficiency. To address this issue, this paper introduces an efficient multi-layer indexing method on blockchain for multimodal data querying. It establishes an effective mapping between on-chain and off-chain data through a verifiable on-chain and off-chain collaborative storage architecture. The paper also proposes Multi-layer Bitmap Block Index (MBBI) and Cuckoo Merkel Tree (C-MT) to optimize the querying process. Experimental results demonstrate that this method not only ensures the consistency and integrity of metadata across on-chain and off-chain but also significantly enhances the efficiency of multimodal data query. This offers a feasible solution for the storage and query needs of large-scale multimodal data.

In order to ensure food safety, it is required for personnel to wear hats and masks during food handling processes. To accurately detect the wearing status of kitchen staff, the ARP-YOLO model is proposed. Firstly, images are obtained from multiple kitchens and angles to construct a dataset reflecting the wearing status of hats and masks. To simulate more complex kitchen environments, Gaussian noise is added to the data and lighting conditions are adjusted for data augmentation. Lighting conditions in the kitchen can affect detection, causing the same target to exhibit different shapes and features under different lighting conditions, leading to missed detections. To address the above issues, we propose ARCA (Adaptive-Residual-Coordinate-Attention), which uses residual connections to strengthen attention to important features while preserving original features, and employs adaptive convolution reduction to reduce module parameters. To improve target localization accuracy, P2 detection layers are added in the Neck to obtain more accurate target position information. The ARP-YOLO model demonstrates significant improvements over the baseline model, with a 13.6% increase in Recall, allowing for more effective target capture and reduced missed detection risk. Additionally, [email protected] has increased by 10%, enhancing target localization accuracy. The F1-Score has also increased by 7.4%, better balancing the relationship between Precision and Recall. To validate the effectiveness of the model, comparative experiments with other models are conducted, showing that ARP-YOLO model's Recall and Average Precision(AP) are higher than those of other models.

The detection and evaluation of gait kinematics is vital for patient diagnosis and rehabilitation. Aiming at limitations of commonly used optical capture and wearable sensing systems in clinical applications, this paper proposes a thigh attitude angle prediction method based on the hip error tolerance from the kinematic data of two shank-mounted IMUs (Inertial Measurement Unit). The novelties of the proposed method are summarized as follows: i) It develops a parallel approach to regress variation of hip error tolerance for different subjects. This parallel approach, by simultaneously deconstructing the shank kinematics data via different regression algorithm including support vector machine, boosting tree, and stepwise linearity regression, is able to well accommodate the characteristics of both health subjects and patients. ii) It develops some evaluation indices based on gait symmetry, consistency, and activity to fully evaluate the human lower limbs motion performance in gait by only two IMUs. The effectiveness of the proposed method is verified by the experimental results among 8 healthy subjects and 16 cerebral infarction patients. For the healthy subjects, the estimated error of thigh prediction compared with Xsens and Vicon are 3.3 ± 0.3° and 3.5 ± 0.7°, respectively. For the patients, the estimated error compared with Xsens-measured angle is 4.8 ± 1.7°. Its broad significance in actual intelligent healthcare and robotics-assisted rehabilitation is three-fold: First, it is a recursive real-time method as it only needs data from the previous gait cycle to predict the thigh angle. Second, its accuracy meets the actual clinical needs. Third, it is a low-cost method that only needs two IMUs and has high potentials of clinical applications. Note to Practitioners—Gait is of great significance to quantify the degree of movement disorders in clinical practice, and the existing equipment is rarely able to meet the needs of full dimension, low cost, and low place constraints in clinical practice. This paper presents an innovative gait kinematic prediction method, which only uses two IMUs attached to the shanks to predict the orientation of thigh. Compared with related works, the proposed methods have achieved high-precision, real-time and low-cost acquisition. This paper is inspired by the problems of large number, high motion interference, and high cost of wearable gait measurement devices in clinical practice. The proposed method have potential to be integrated into exoskeletons and medical walking AIDS, which could greatly improve the comfort and control precision of the wearable system.

In the wireless sensor networks with multiple mobile sinks, the movement of sinks or failure of sensor nodes may leads to the breakage of existing routes. In order to repair broken path with lower communication overhead in terms of both energy and delay, we propose an efficient routing recovery protocol with endocrine cooperative particle swarm optimization algorithm to establish and optimize the alternative path. With this method, the alternative path from source nodes to the sink with the optimal QoS parameters can be selected. Simulation results demonstrate that ECPSOA can adapt to rapid topological changes with multiple mobile sinks, while decreasing communication overhead and efficiently reducing the energy consumption.