A Computer Vision-Based Yoga Pose Grading Approach Using Contrastive Skeleton Feature Representations

Contrastive learning has demonstrated promising results in unsupervised abstractive summarization.However, existing methods rely on manually crafted negative examples, demanding substantial human effort and domain knowledge.Moreover, these human-generated negative examples may be poor in quality and lack adaptability during model training.To address these issues, we propose a novel approach that learns trainable negative examples for contrastive learning in unsupervised abstractive summarization, which eliminates the need for manual negative example design.Our framework introduces an adversarial optimization process between a negative example network and a representation network (including the summarizer and encoders).The negative example network is trained to synthesize hard negative examples that are close to the positive examples, driving the representation network to improve the quality of the generated summaries.We evaluate our method on two benchmark datasets for unsupervised abstractive summarization and observe significant performance improvements compared to strong baseline models.

Obstacle detection (OD) techniques for collision avoidance applications in automobiles and unmanned aerial vehicles have been of research interest in the recent years. This technical idea is transferred to develop a local navigation system for the visually impaired individuals which is the focus of this paper. Several types of equipment and approaches namely sensors, Computer Vision (CV) methods, micro-controllers etc. are used for detecting obstacles in the path. To enable an utmost collision free mobility of the users we intend to use Ultrasonic (US) sensors and CV systems on a smartphone to detect and alarm the user of any obstacle in the path. In this paper, the accuracy of detection of obstacles and estimating their distance from the user using US sensors is determined. Further, the quality of object detection techniques using two OD algorithms is tested on a PC for further application of the best performing algorithm in combination with the US sensors for our major goal of developing a local obstacle-free navigation system for the visually impaired using mobile devices like smartphones.

Real-time monitoring networks are increasingly prevalent in supporting the management of environmental systems as the technology for live data collection becomes more accessible. Additionally, ecosystem and water resource pressures have persisted and intensified under climate pressures and an expanding anthropogenic footprint. The way in which models and data are fused in the day-to-day management of water resources operations, as well as for long-term planning and investment, has been a critical field of research. An adaptive real-time monitoring-integrated learning modelling approach was developed and applied to improve the understanding of the mixing dynamics in a water supply reservoir in Queensland, Australia. This was accomplished through the combination of sequentially linked catchment and reservoir models with in situ real-time measurements of temperature and flow along with meteorological forecasts from an Australian numerical weather model, to produce short-term water quality forecasts. An adaptive learning catchment model was developed and linked for each inflow arm of the reservoir using the Australian Water Balance Model. This framework enabled automated online communication to researchers and managers around the current performance of the inflow predictions and the confidence expected in the current forecasts. Moreover, this live learning catchment model was coupled with a real-time adaptive three-dimensional hydrodynamic model of the reservoir iteratively training using data from the deployed real-time temperature monitoring system. A prototype internet-connected remotely operable autonomous surface vessel was deployed with a winching system for conducting dynamic water quality profiling operations under the guidance of waypoints guidance generated from the real-time adaptive modelling forecasts. Data collected by ASV was subsequently provided back to the modelling system in real-time. The complete system facilitated the online adaptive forecasting of mixing dynamics in the reservoir and the automated identification of features of interest for water quality profiling, as well as dynamically monitoring the areas potentially most valuable for model learning development to improve system-wide understanding and forecast certainty through addition into the live dataset for ongoing training and evaluation. Evidence was found in support of a rolling iterative calibration procedure for increasing model skill sensitivity to different processes occurring over temporal and spatial scales across both catchment and receiving water models. Dynamically guided spatial monitoring generated from maximum predicted areas of variation and parameter sensitivity in the real-time adaptive receiving water model demonstrated that monitoring of the receiving water inflow arms during inflow events was necessary during inflow events to train the model on the strongest signal of the driving force of changes in the receiving water environment. Overall, the uncertainty in rainfall events from both forecasted and observed sources cascading with the uncertainty in catchment simulations with only static indirect monitoring of flow (ungauged at any of the inflow arms to the reservoir) was found to be the most significant hindrance to the utility of the applied real-time adaptive modelling framework. The application of an adaptive computer vision-based stream gauging approach was then trialled on one of the ungauged inflow arms in order to supplement this gap.

There is growing interest among smartphone users in the ability to determine their precise location in their environment for a variety of applications related to wayfinding, travel and shopping. While GPS provides valuable self-localization estimates, its accuracy is limited to approximately 10 meters in most urban locations. This paper focuses on the self-localization needs of blind or visually impaired travelers, who are faced with the challenge of negotiating street intersections. These travelers need more precise self-localization to help them align themselves properly to crosswalks, signal lights and other features such as walk light pushbuttons. We demonstrate a novel computer vision-based localization approach that is tailored to the street intersection domain. Unlike most work on computer vision-based localization techniques, which typically assume the presence of detailed, high-quality 3D models of urban environments, our technique harnesses the availability of simple, ubiquitous satellite imagery (e.g., Google Maps) to create simple maps of each intersection. Not only does this technique scale naturally to the great majority of street intersections in urban areas, but it has the added advantage of incorporating the specific metric information that blind or visually impaired travelers need, namely, the locations of intersection features such as crosswalks. Key to our approach is the integration of IMU (inertial measurement unit) information with geometric information obtained from image panorama stitchings. Finally, we evaluate the localization performance of our algorithm on a dataset of intersection panoramas, demonstrating the feasibility of our approach.

How do drivers perceive risk? Understanding what situations and factors cause drivers to perceive situations as critical can improve our understanding of road user behavior and inform automated driving technology. To investigate the factors that shape drivers’ risk perception, we conducted an eye-tracking study with 27 participants who watched dashcam videos and continuously rated the perceived risk of various driving situations. Using the resulting dataset, we developed a computer vision-based machine learning approach that generates explainable predictions of perceived risk from video and eye-tracking data. Our SHAP analysis reveals that the proximity of objects and number of cars in a scene are the most significant contributors to perceived risk. Most interestingly, while people tend to sample similar objects in critical situations, their risk perception remains highly personal making visual sampling behavior a weak predictor of perceived risk. Overall, our explanations reveal non-linear insights beyond previous work, suggesting that risk perception is not only shaped by visual input, but primarily by cognitive processes which is in line with theoretical models of situation awareness. The dataset, source code, and a comprehensive usage guide are publicly available.

Soil classification is one of the major affairs and emanating topics in a large number of countries. The population of the world is rising at a majorly rapid pace and along with the increase in population, the demand for food surges actively. Typical techniques employed by the farmers are not adequate enough to fulfill the increasing requirements and therefore they have to hinder the cultivating soil. For proper crop yield, farmers should be aware of the correct soil type for a particular crop, which affects the increased demand for food. There are various laboratory and field methods to classify soil, but these have limitations like time and labor-consuming. There is a requirement of computer-based soil classification techniques which will help farmers in the field and won’t take a lot of time. This paper talks about different computer-based soil classification practices divided into two streams. First is image processing and computer vision-based soil classification approaches which include the conventional image processing algorithms and methods to classify soil using different features like texture, color, and particle size. Second is deep learning and machine learning-based soil classification approaches, such as CNN, which yields state-of-the-art results. Deep learning applications mostly diminish the dependency on spatial-form designs and preprocessing techniques by facilitating the end-to-end process. This paper also presents some databases created by the researchers according to the objective of the study. Databases are created under different environmental and illumination conditions, using different appliances such as digital cameras, digital camcorder, and a smartphone camera. Also, evaluation metrics are briefly discussed to layout some graded measures for differentiation. This review serves as a brief guide to new researchers in the field of soil classification, it provides fundamental understanding and general knowledge of the modern state-of-the-art researches, in addition to skillful researchers considering some dynamic trends for future work.

Human Eye is one of the most dedicated organs. Eyes help the human beings to see the world around them. But by using the power of computer vision and machine learning, researchers are working on the way to detect human eyes and find a way to control a computer system. The authors' main idea is to work on an algorithm that will enable human beings to access any computing platform just by their eyes. The main challenge is to use eye movements to track and access a computing platform. The foremost thing is that the algorithm will work on complex images or video feed regarding any constraints on the background or any color pigment of human skin complexion or tone.

We present a novel computer vision-based deep learning approach for metadata extraction as both a central component of and an ancillary aid to structured information extraction from scientific literature which has various formats. The number of scientific publications is growing rapidly, but existing methods cannot combine the techniques of layout extraction and text recognition efficiently because of the various formats used by scientific literature publishers. In this paper, we introduce an end-to-end trainable neural network for segmenting and labeling the main regions of scientific documents, while simultaneously recognizing text from the detected regions. The proposed framework combines object detection techniques based on Recurrent Convolutional Neural Network (RCNN) for scientific document layout detection with Convolutional Recurrent Neural Network (CRNN) for text recognition. We also contribute a novel data set of main region annotations for scientific literature metadata information extraction to complement the limited availability of high-quality data set. The final outputs of the network are the text content (payload) and the corresponding labels of the major regions. Our results show that our model outperforms state-of-the-field baselines.

Colorimetric detection, quantification and extraction of Fe(III) in water by acrylic polymers with pendant Kojic acid motifs

Lane detection systems play a critical role in ensuring safe and secure driving by alerting the driver of lane departures. Lane detection may also save passengers’ lives if they go off the road owing to driver distraction. The article presents a three-step approach for detecting lanes on high-speed video pictures in real-time and invariant lighting. The first phase involves doing appropriate prepossessing, such as noise reduction, RGB to grey-scale conversion, and binarizing the input picture. Then, a polygon area in front of the vehicle is picked as the zone of interest to accelerate processing. Finally, the edge detection technique is used to acquire the image’s edges in the area of interest, and the Hough transform is used to identify lanes on both sides of the vehicle. The suggested approach was implemented using the IROADS database as a data source. The recommended method is effective in various daylight circumstances, including sunny, snowy, and rainy days, as well as inside tunnels. The proposed approach processes frame on average in 28 ms and have a detection accuracy of 96.78%, as shown by implementation results. This article aims to provide a simple technique for identifying road lines on high-speed video pictures utilizing the edge feature.

Falls are one of the most serious dangers for elderly people who live alone at home. It has become a widespread issue all across the world. Reliable fall detection systems can help to mitigate the negative consequences of accidentally falling. Several techniques for automatically falling detection machines have been suggested in the past. The existing technologies are classified into three types of fall detectors: wearable sensor-based, ambient device-based, and computer vision-based approaches. This paper focuses on a dataset comprising signals from wearable sensors, ambient sensors, and vision devices. We propose a novel feature subset selection to reduce the number of effective input attributes based on a hybridized metaheuristic - an Adaptive Particle Swarm and Grey Wolf Optimization (APGWO). Classification results use various machine learning classifiers such as Logistic Regression (LR), K-Nearest Neighbor, Support Vector Machine (SVM), Decision Tree (DT), Naive Bayes (NB), Random Forest (RF), and Multilayer Perceptron (MLP), show that the proposed approach is highly effective. Classification accuracy and F1 score can reach as high as 99% and 96%, respectively.

Commercial Unmanned aerial vehicle (UAV) industry, which is publicly known as drone, has seen a tremendous increase in last few years, making these devices highly accessible to public. This phenomenon has immediately raised security concerns due to fact that these devices can intentionally or unintentionally cause serious hazards. In order to protect critical locations, the academia and industry have proposed several solutions in recent years. Computer vision is extensively used to detect drones autonomously compared to other proposed solutions such as RADAR, acoustics and RF signal analysis thanks to its robustness. Among these computer vision-based approaches, we see the preference of deep learning algorithms thanks to their effectiveness. In this paper, we are presenting an autonomous drone detection and tracking system which uses a static wide-angle camera and a lower-angle camera mounted on a rotating turret. In order to use memory and time efficiently, we propose a combined multi-frame deep learning detection technique, where the frame coming from the zoomed camera on the turret is overlaid on the wide-angle static camera’s frame. With this approach, we are able to build an efficient pipeline where the initial detection of small sized aerial intruders on the main image plane and their detection on the zoomed image plane is performed simultaneously, minimizing the cost of resource exhaustive detection algorithm. In addition to this, we present the integral system including tracking algorithms, deep learning classification architectures and the protocols.

Diabetic retinopathy (DR), a frequent complication of diabetes mellitus, is the prime reason for vision loss worldwide. The current method for identifying diabetic retinopathy is a time-consuming and manual process, requiring trained medical professionals to examine digital color fundus photographs of the retina and look for the presence of lesions associated with the disease’s vascular abnormalities. Despite technological advancements, accurately identifying the early stages of diabetic retinopathy remains a challenge. To address this issue, different computer vision-based approaches have been introduced to detect DR and classify various stages from retina images. However, these techniques have limited success, particularly in accurately identifying the early stages of the disease. In this study, we employed a publicly available Kaggle Eyepac dataset to train our model for the detection of DR which is known to have five distinct stages, namely, proliferate DR, severe DR, moderate DR, mild DR, and no DR. The overall accuracy, precision, recall, Fl-score, and AUC score obtained by our proposed model were 88.31± 0.23,89.15± 0.19,90.25± 0.41,89.69± 0.25, and 87.38± 1.45 respectively. The performance of our model is improved by employing a pre-processing technique known as Contrast Limited Adaptive Histogram Equalization (CLAHE) followed by a state-of-the-art deep learning model known as ConvNeXt to achieve promising results.

Sign Language is the elementary communication media for Deaf & Mute (D&M) people. On the other hand, it seems too tenacious for the general people to understand this language. In order to tear out this communication barrier, a real-time automated translator is essential. Through this research, a computer vision-based approach has been developed for the recognition of Bangla Sign Language (BdSL) characters. In this work, a deep learning-based recognition model has been developed. Adaptive thresholding has been integrated with 2D Convolutional Neural Network (CNN) to construct this model. Proposed model has been trained to build this real-time automated translator through our own created dataset (dataset containing 3600 different images for 36 distinct characters). The proposed model has been trained and tested with 2880 (80%) training images and 720 (20%) testing images respectively. Thirty-six unique characters of Bangla Sign Language can be recognized through this model with significant accuracy. The model delivers validation accuracy of 99.72% and validation loss of 0.73%. A significant result has been achieved for the recognition and translation of Bangla Sign Language characters with this dataset over other existing Bangla Sign Language Recognition model.

Computer vision-based approaches are very useful for dynamic displacement measurement, damage detection, and structural health monitoring. However, for the application using a large number of existing cameras in buildings, the computational cost of videos from dozens of cameras using a centralized computer becomes a huge burden. Moreover, when a manual process is required for processing the videos, prompt safety assessment of tens of thousands of buildings after a catastrophic earthquake striking a megacity becomes very challenging. Therefore, a decentralized and fully automatic computer vision-based approach for prompt building safety assessment and decision-making is desired for practical applications. In this study, a prototype of a novel stand-alone smart camera system for measuring interstory drifts was developed. The proposed system is composed of a single camera, a single-board computer, and two accelerometers with a microcontroller unit. The system is capable of compensating for rotational effects of the camera during earthquake excitations. Furthermore, by fusing the camera-based interstory drifts with the accelerometer-based ones, the interstory drifts can be measured accurately even when residual interstory drifts exist. Algorithms used to compensate for the camera’s rotational effects, algorithms used to track the movement of three targets within three regions of interest, artificial neural networks used to convert the interstory drifts to engineering units, and some necessary signal processing algorithms, including interpolation, cross-correlation, and filtering algorithms, were embedded in the smart camera system. As a result, online processing of the video data and acceleration data using decentralized computational resources is achieved in each individual smart camera system to obtain interstory drifts. Using the maximum interstory drifts measured during an earthquake, the safety of a building can be assessed right after the earthquake excitation. We validated the feasibility of the prototype of the proposed smart camera system through the use of large-scale shaking table tests of a steel building. The results show that the proposed smart camera system had very promising results in terms of assessing the safety of steel building specimens after earthquake excitations.