Floating photovoltaic (FPV) panel characterization is highly dependent to the ambient conditions as it receives the cooling effect between the temperature variants of vapor and water body. The current limitation in the existing measurement tool does not cover all range of weather and energy monitoring. To support the FPV study, a remote weather and energy monitoring system is developed based on IEC 61724-1:2021 standard to acquire online monitoring of data. The monitoring parameters includes the cell temperature, current and voltage from PV panel, and the meteorological parameters such as solar radiation, ambient temperature, water temperature, wind speed, wind direction, humidity and rainfall. The system uses a novel relay switch strategy to measure the raw current and voltage of the FPV under no-load condition and uses an algorithm to trace the I-V curve maximum power point and transmits the data online to a remote site. An empirical study with an FPV system was carried out to test the performance of the proposed remote weather and energy data acquisition system. The result showed reliable data measurement with a success rate of 91.20% data reception with a common discrepancy of signal loss and surrounding interference. The four-day FPV panel characterization test finds a significant cooling effect in the maximum power point based on ambient conditions with an average temperature reduction of 0.23°C corresponding to a relative yield difference of of 9.82% in favor of floating photovoltaics compared to a traditional ground-mounted system.

Abstract The outfit testing using virtual try-on technology is increasingly in demand for online commercialization due to its flexibility and adaptability. With the advanced development in computer vision and three-dimensional (3D) modelling, it can measure the best fit of garment outfits for a digital human. The capture of physical humans and garments using scanning tools can be processed in point cloud, voxel, meshes and RGB-D format. Garment reconstruction is then processed by extracting clothing from existing 3D garment datasets or employing 2D-to-3D garment lifting techniques. The garment is applied to a human avatar, which is generally reconstructed using implicit function-based human reconstruction, neural radiance fields and 3D Gaussian Splatting techniques. By implementing generative adversarial networks or diffusion models, the garments can be transferred onto a new human body pose. This paper delves into three core components in virtual try-on technology, i.e. 3D garment generation, 3D human avatar reconstruction and body-garment transformation. A detailed quantitative comparison evaluates the effectiveness of digital avatar reconstruction and garment transformation methods based on outputs, datasets and performance metrics. Furthermore, this study explores practical application, identifies existing challenges, and outlines promising directions for future research. By providing an in-depth analysis of the current state of 3D VTO, this work serves as a valuable resource for advancing innovation in digital fashion technologies.

Defect detection is an essential step to ensure the quality of manufactured metal products. Industrially, detecting metallic surface defects can be challenging due to a resourceconstrained environment in terms of hardware computational availability. In this study, we explore a lightweight defect detection model, specifically the single-shot multi-box detector variant called SSDLite. In this study, the SSDLite is paired with varying backbone base networks, MobileViT, MobileNetv2 and MobileNetv3. Transfer learning is employed to enhance the SSDLite learning by enabling a relation between previous tasks and the targeted task, which is a domain-specific task. The same implementation details are applied across all the SSDLite models (MobileViT, MobileNetv2 and MobileNetv3) being trained on the PASCAL VOC dataset, and then the prior knowledge in the form of pre-trained weights is used to fine-tune the model on a domainspecific metallic surface defect dataset. The metallic surface defect in this study is based on a hot-rolled steel strip surface defect dataset called the NEU-DET dataset. This study lays the groundwork for future investigations into alternative backbone architectures with SSD for enhanced detection accuracy and efficiency in real-world manufacturing applications.

With the introduction of 5G technologies that promise lower latency, more reliable, and higher network capacity, many researchers have proposed to integrate various 5G technologies such as millimeter waves, multi-radio access technology, and reconfigurable intelligent surfaces (RIS) with vehicle-to-everything (V2X). RIS is first introduced as an arranged array of reflectors that can be independently and dynamically reconfigured to adjust the signal propagation direction in a favorable way. With RIS, the overall performance of the wireless communication system can be improved. However, it is unclear whether RIS can be directly applied in areas of high mobility, specifically within vehicular traffic. This paper presents an RIS-aided network model that can be utilized to enhance the V2X communication networks with multiple vehicles (V2V) and infrastructure (V2I) links. A Soft-Actor Critic Deep Reinforcement Learning (SACDRL) approach is proposed to optimize the network model to achieve the stringent quality-of-service (QoS) requirements. Subsequently, numerical simulations were performed to validate the performance of the proposed SAC-DRL method.

Abstract Fault detection and diagnosis (FDD) plays a vital role in abnormal situation management of chemical industrial processes. Current FDD technologies mostly rely on data-driven solutions by making full use of abundant process data collected by the state-of-the-art distributed process instruments and sensors. Deep learning algorithms were widely used among all the data-driven algorithms. Industrial process time series data could be processed with ease by deep learning algorithms, particularly transformer-based models because of their multi-head attention mechanism. Different lengths of snippets of sequence (or subsequence) would have a multitude of perspectives viewed by the deep learning algorithms, subsequently impacting their FDD performance. This study, therefore, aims to investigate the effects of varying subsequence lengths on the FDD performance of common deep learning algorithms, consisting of a multilayer perceptron, convolutional neural network, long short-term memory, transformer, industrial process optimisation—vision transformer (IPO-ViT) using two benchmark case studies, namely continuous stirred tank reactor (CSTR) and Tennessee Eastman Process (TEP). Additionally, faulty data are rare to occur, and the fault labelling process is generally tedious and expensive to perform. The effects of labelled training data sizes were also studied on the FDD performance. The findings clearly indicate that the IPO-ViT, a variant of transformer-based models, exhibited the best FDD performance under 10% and 50% subsequence length of data on CSTR and TEP case studies, respectively, for optimal feature extraction, even with 10% of fully labelled input data.

Data compression is highly required to reduce the massive size of data while achieving lossless information over the transmission. In this paper, a novel multi-channel time series codec framework (LSCodec) is proposed to decouple long-term trend and short-term fluctuation using manually guided preprocessing data. The proposed LSCodec contains an encoder-decoder architecture network integrated with a group residual vector quantizer. The input data is decoupled through two paths layer by layer. In one path, a LSTM model is introduced for long-term trend feature learning. Another path produces fluctuation signal by subtracting between the real-time series and the trend signal to learn its hidden representation. The output of both path are then quantized using two individual residual vector quantization. A joint reconstruction loss combining its trend and fluctuation loss is used to support the training process. A balancer is used to stabilize training gradient of reconstruction loss to avoid local optimal solution or unstable state. Our experimental results show that the compression rate can vary to different bite rate according to strides setting. For multi-channel time series, it can compress data into average 10% with an acceptable reconstruction result. By reducing part of coding index, it is able to reconstruct part of curve with its main distribution. LSCodec can achieve a relatively good result in downstream task for a dataset that contains high ratio of anomaly. The proposed method can restore data distribution without losing its abnormal part. Several comparative studies are performed on with or without manually guided. The result shows the effectiveness of data guiding strategy. Code and models are available at https://github.com/HaiweiZuo/LSCodec.

Abstract Remote photoplethysmography (rPPG) gains recent great interest due to its potential in contactless heart rate measurement using consumer-level cameras. This paper presents a detailed review of rPPG measurement using computer vision and deep learning techniques for heart rate estimation. Several common gaps and difficulties of rPPG development are highlighted for the feasibility study in real-world applications. Numerous computer vision and deep learning methods are reviewed to mitigate crucial issues such as motion artifact and illumination variation. In comparison, deep learning approaches are proven more accurate than conventional computer vision methods due to their adaptive pattern learning and generalization characteristics. An increasing trend of applying deep learning techniques in rPPG can improve effective heart rate estimation and artifact removal. To consider more realistic disturbances into account, additional vital signs and large training datasets are crucial to improve the accuracy of heart rate estimations. By taking the benefit of contactless and accurate estimation, the application of rPPG can be greatly adopted in real-world activities, especially in precision sports.

Abstract The increasing number of diabetes individuals in the globe has alarmed the medical sector to seek alternatives to improve their medical technologies. Machine learning and deep learning approaches are active research in developing intelligent and efficient diabetes detection systems. This study profoundly investigates and discusses the impacts of the latest machine learning and deep learning approaches in diabetes identification/classifications. It is observed that diabetes data are limited in availability. Available databases comprise lab-based and invasive test measurements. Investigating anthropometric measurements and non-invasive tests must be performed to create a cost-effective yet high-performance solution. Several findings showed the possibility of reconstructing the detection models based on anthropometric measurements and non-invasive medical indicators. This study investigated the consequences of oversampling techniques and data dimensionality reduction through feature selection approaches. The future direction is highlighted in the research of feature selection approaches to improve the accuracy and reliability of diabetes identifications.

Real time power quality monitoring is important to ensure stable functioning of the electrical appliances especially for the manufacturing sector. Deep- WT-ConvT is proposed to to better characterise and differentiate the minor differences between different types of power quality disturbances. However, the use of deep networks requires longer training time, and poses the risk of getting internal covariant shift issues due to distribution change in layer's input during training phase. This issue can be prevented by proper parameter initialisation and with lower learning rate, which slows down the training process. Batch normalisation (BN) layers are proposed to improve the classification performance of the PQD classifier network WT-ConvT. Results shows significant improvement on Deep- WT-ConvT model with accuracy improvement from 92.95% without BN layers to 94.44% with BN layers on 20dB SNR AWGN noise test.

Inertial Measurement Unit (IMU) has been widely recognized to be the practical alternative to capture and analyze human gait. However, due to its inherent characteristics, it can only measure the basic kinematics of the body segment it attached to. With the help of the machine learning, IMU can be used to determine the dynamic behavior of the major lower extremity muscle. This paper explores the use of feature-extracted IMU data and a neural network to estimate muscle activity during walking. IMU and Electromyogram (EMG) data were collected from fifty-eight healthy participants. Principal Component Analysis (PCA) and Tsfresh (Time Series FeatuRe Extraction on basis of Scalable Hypothesis tests) were applied to extract the relevant features from the data. These features were then used to train the Feedforward Neural Network (FNN). A combination of Tsfresh and FNN yielded the best results with correlation coefficient$(r)$of 95.73% and Root Mean Square Error (RMSE) of 11.20%. This research can potentially help reduce the number of sensors needed in gait analysis, allow for portable motion capture, and improve the accuracy and efficiency of the FNN model in estimating muscle activity.

Weightlifting is a demanding sport requiring power, flexibility, and the correct technique. The snatch and the clean and jerk involve the fast lifting of weight to an overhead position. Incorrect technique or posture may lead to inefficient lifts or even injury. This paper presents a new framework for biomechanics analysis and barbell trajectory tracking in weightlifting by leveraging the capabilities of Keypoint-RCNN and YOLOv7 deep learning models. The proposed framework extracts skeletal information from weightlifting video sequences using a pre-trained Keypoint-RCNN model for human pose estimation and a custom YOLOv7 model to detect and track barbell trajectories. The Keypoint-RCNN model estimates human pose without manual annotation or specialised apparatus, while the YOLOv7 model provides real-time, non-intrusive barbell tracking. The efficacy of barbell trajectory tracking with YOLOv7 on a public weightlifting dataset of 973 images (70–30 train-test ratio) was evaluated, obtaining high precision (0.9214), recall (0.9678), and [email protected] of 0.9792 and [email protected]:0.95 of 0.7765, indicating the applicability of this model to weight training applications. The proposed framework presents a cost-effective, user-friendly, and easily accessible alternative to conventional motion capture and analysis systems, making it accessible for lifters of all skill levels and training environments.

Archery is a precision sport that requires high consistency and accuracy. The score's consistencies are highly correlated with respect to the archer's shooting pattern. With the current advancement of technology, archery scoring remains manual using scope observation in the field. By leveraging computer vision, recent efforts are made to automate the scoring process with the use of cameras. Previously studied computer vision approaches primarily explored processes such as homography estimation, image subtraction, and circle detection. However, these approaches do not accommodate the real target board faces, highly skewed camera angle, and estimated circles that may deviate from actual target rings. In this paper, a two-stage automatic archery scoreboard scoring system is proposed to detect arrows using camera. Firstly, the initiation stage extracts the scoring areas of the target ring by incorporating curve interpolation. Subsequently, the operation stage detects and localizes an arrow in consecutive frames. The proposed method demonstrates accurate arrow extraction from the target ring with less susceptibility to noise. It is able to operate consistently in challenging conditions due to an isolated detection stage.

Abstract Cracks are one of the critical structural defects in building assessment to determine the integrity of civil structure. Structural surveying process using computer vision is required to automatically identify cracks. The application of Convolutional Neural Networks (CNNs) is limited by its fixed geometric kernels to extract the irregular shape of cracks. In this paper, a mask Region-based Denoised Deformable Convolutional Network (R-DDCN) is proposed to detect cracks for accurate instance segmentation and image classification. Denoised deformable convolution is introduced to improve the modeling capability of convolution layer. It adopts the existing deformable convolution, with non-local means as a denoising mechanism to optimize the augmentation of spatial sampling locations with filtered offsets. Experimental results show that the proposed mask R-DDCN has lower validation loss and improved mean accuracy precision of mAP75 from 66.7% to 76.7% as compared to the mask R-CNN. Mask R-DDCN can perform better modeling capability in cracks identification.

Abstract Postural assessment is crucial in the sports screening system to reduce the risk of severe injury. The capture of the athlete’s posture using computer vision attracts huge attention in the sports community due to its markerless motion capture and less interference in the physical training. In this paper, a novel markerless gait estimation and tracking algorithm is proposed to locate human key-points in spatial-temporal sequences for gait analysis. First, human pose estimation using OpenPose network to detect 14 core key-points from the human body. The ratio of body joints is normalized with neck-to-pelvis distance to obtain camera invariant key-points. These key-points are subsequently used to generate a spatial-temporal sequences and it is fed into Long-Short-Term-Memory network for gait recognition. An indexed person is tracked for quick local pose estimation and postural analysis. This proposed algorithm can automate the capture of human joints for postural assessment to analyze the human motion. The proposed system is implemented on Intel Up Squared Board and it can achieve up to 9 frames-per-second with 95% accuracy of gait recognition.

Abstract Conventional training mechanisms often encounter limited classification performance due to the need of large training samples. To counter such an issue, the field of meta-learning has shown great potential in fine tuning and generalizing to new tasks using mini dataset. As a variant derived from the concept of Model Agnostic Meta-Learning (MAML), an one-step MAML incorporated with the two-phase switching optimization strategy is proposed in this paper to improve performance using less iterations. One-step MAML uses two loops to conduct the training, known as the inner and the outer loop. During the inner loop, gradient update is performed only once per task. At the outer loop, gradient is updated based on losses accumulated by the evaluation set during each inner loop. Several experiments using the BERT-Tiny model are conducted to analyze and compare the performance of the one-step MAML with five benchmark datasets. The performance of evaluation shows that the best loss and accuracy can be achieved using one-step MAML that is coupled with the two-phase switching optimizer. It is also observed that this combination reaches its peak accuracy with the fewest number of steps.

Optimization in a deep neural network is always challenging due to the vanishing gradient problem and intensive fine-tuning of network hyperparameters. Inspired by multistage decision control systems, the stochastic diagonal approximate greatest descent (SDAGD) algorithm is proposed in this article to seek for optimal learning weights using a two-phase switching optimization strategy. The proposed optimizer controls the relative step length derived based on the long-term optimal trajectory and adopts the diagonal approximated Hessian for efficient weight update. In Phase-I, it computes the greatest step length at the boundary of each local spherical search region and, subsequently, descends rapidly toward the direction of an optimal solution. In Phase-II, it switches to an approximate Newton method automatically once it is closer to the optimal solution to achieve fast convergence. The experiments show that SDAGD produces steeper learning curves and achieves lower misclassification rates compared with other optimization techniques. Implementation of the proposed optimizer to deeper networks is also investigated in this article to study the vanishing gradient problem.

Stereo-vision technology has shown its advantages to overcome the occlusion and realistic information. However, marker-less human motion detection and tracking in the unconstrained environment were led to the difficulty of features extraction. In this paper, we proposed a hybrid technique of Gaussian and median filter to improve the shadow and sudden change of the illumination problems. The skeleton model of the detected human was constructed using the sequential mathematical morphology. Based on the results, the skeleton model produced was not affected by the shadow and the illumination issue. Proposed approach and the normalized filter approach produces up to 86% and 71% of the average accuracy tracking respectively in the real-time tracking. Hence, the proposed approach could improve the performance of the human detection in the unconstrained environment.

Optimization is important in neural networks to iteratively update weights for pattern classification. Existing optimization techniques suffer from suboptimal local minima and slow convergence rate. In this paper, stochastic diagonal Approximate Greatest Descent (SDAGD) algorithm is proposed to optimize neural network weights using multi-stage backpropagation manner. SDAGD is derived from the operation of a multi-stage decision control system. It uses the concept of control system consisting of: (1) when the local search region does not contain a minimum point, the iteration shall be defined at the boundary of the local search region, (2) when the local region contains a minimum point, the Newton method is used to search for the optimum solution. The implementation of SDAGD on Multilayer perceptron (MLP) is investigated with the goal of improving the learning ability and structural simplicity. Simulation results showed that two layer MLP with SDAGD achieved a misclassification rate of 4.7% on MNIST dataset.

Biofeedbacks delivery during rehabilitation have been known to improve postural control and shorten rehabilitation periods. A biofeedback system communicates with the human central nervous system (CNS) through a variety of feedback modalities. Among the many available modalities vibrotactile feedback devices are gaining much attention. This is due to their desirable characteristics and simplistic manner of presenting information to the CNS. An intelligent biofeedback system integrated with wireless sensors for monitoring postural control during rehabilitation was hypothesized to shorten rehabilitation periods. This work presents the design of a postural control measuring device integrated with real-time intelligent biofeedback for postural correction. The system integrates three modules: (a) inertial measurement units (IMUs), (b) fuzzy knowledge base, and (c) feedback driver circuit. Human posture is measured using Euler angular measurements from the IMUs. A fuzzy inference system (FIS) was used to determine quality of postural control, based on measurements from the IMUs. Forewarning of poor postural control is given by vibrotactile actuators (biofeedback). Experiments were conducted to test viability of the system in achieving accurate real-time measurements and interventions. The results observed improvements in postural control when biofeedback intervention was present.

Autonomous vehicles are equipped with optical sensors and micro-processing units to perform intelligent visual analysis of its surroundings. Due to the high speed of moving vehicle, the captured information has to be processed in a short duration to avoid possible collision. In this paper, a ground-image plane mapping technique is proposed to quickly locate detected object if the object's position is known in the real world. A three dimensional (3D) world coordinate is mathematically derived to an image plane using pinhole camera model. Several 3D perspective parameters such as vehicle's steering angle and its velocity, sensor's height and tilting angle are encompassed in the ground plane measurement. The optical sensor's intrinsic parameters such as focal length, principal point, pixel's height and width are also inserted for the mathematical model derivation. The importance of this ground to image plane mapping enables a rapid search of an object in a moving scene to achieve fast object identification during sensor acquisition. Experimental results have been carried on the application of lane marks detection with 93.82% correct mapping, using approximately 20% less processing time.

A dynamic counterpropagation network based on the forward only counterpropagation network (CPN) is applied as the classifier for face detection. The network, called the dynamic supervised forward-propagation network (DSFPN) trains using a supervised algorithm that grows dynamically during training allowing subclasses in the training data to be learnt. The network is trained using a reduced dimensionality categorized wavelet coefficients of the image data. Experimental results obtained show that a 94% correct detection rate can be achieved with less than 6% false positives.