Video Anomaly Detection (VAD) is the process of identifying anomalous events by analyzing spatiotemporal patterns in video. Furthermore, VAD is a complex task due to difficulties in obtaining frame-level annotations, data imbalance issues, and the equivocal and context-dependent nature of video anomalies. To address these issues, this article presents a weakly supervised learning-based Multi-head Global-Local Attention-enabled Dual-Stream Network (MGLA-DSNet) that effectively utilizes spatial (appearance) and temporal (motion) features, with an emphasis on context dependency. The proposed model uses two streams to extract RGB and optical flow features corresponding to appearance (spatial) and motion (temporal) properties, respectively. Subsequently, multi-head global and location attention with adaptive gating and head-wise specialization is applied to the concatenated RGB and Flow features to efficiently model global and local contexts, respectively, using multiple instance learning Finally, the proposed MGLA-DSNet model outperforms state-of-the-art methods across three benchmark datasets, including CUHK Avenue, ShanghaiTech Campus, and UCF-Crime.

Predictive maintenance (PdM) in the context of Industry 4.0 faces a dual challenge: accurately detecting rare equipment failures (extreme class imbalance), while simultaneously avoiding excessive false alarms that can cause operator fatigue and reduced system trust. Existing machine learning approaches often rely on heuristic decision thresholds or unconstrained cost minimization, which lack explicit control over False Alarm Rate (FAR) and may compromise safety compliance during deployment. A novel framework, Cost-Sensitive Thresholding-based Conformal Light Gradient Boosting Machine (CST-Conformal-LGBM), which integrates algorithm-level cost-sensitive LightGBM training with distribution-free conformal calibration and FAR-constrained thresholding, is proposed to address these challenges. The approach introduces a risk-controlled decision rule that minimizes economic cost when the decision threshold is within a statistically defined safe region. SHAP-based feature attribution and manifold-learning-based diagnostics are integrated to enhance interpretability. Comprehensive experimental evaluation demonstrates that the proposed framework achieves approximately 1.50% FAR on the test set, within the predefined 2% safety constraint, while maintaining competitive operational costs and stable decision performance.

Predictive maintenance (PdM) is an integral part of any modern industry. It aims to predict failure and maintenance schedules for key components across the industry using historical and current sensor data, leveraging Artificial Intelligence (AI). However, the costs associated with industrial maintenance exhibit extreme asymmetry, with the economic penalty for a False Negative (missed failure) significantly higher than that for a False Positive (nuisance alarm). Furthermore, existing cost-sensitive boosting methods often struggle with mixed objectives, where misclassification costs are directly injected into the training phase (e.g., through gradient weighting). Subsequently, these models generate noisy alarm profiles that are difficult to tune to asset-specific risk tolerances. Therefore, there is a critical need for a framework that preserves ranking integrity while also allowing precise, decoupled cost optimization at deployment time. Hence, this paper proposes a Cost-Sensitive Thresholding-based Light Gradient Boosting Machine (Decoupled), abbreviated as (CST-LGBM-D) model, which separates cost-sensitive decision optimization from model training for reliable PdM. The proposed CST-LGBM-D model offers accuracy of 94.00%. The comprehensive experimental analysis on the benchmarked dataset supported by explainable AI validates the superiority of the proposed CST-LGBM-D model.

The proposed work detects moving vehicles using unsupervised methods and estimates their speed as well as distance using surveillance cameras mounted in road infrastructure for collision avoidance at sharp corners. The goal is to develop IoT-based computer vision-assisted vehicle-to-infrastructure (V2I) communication for autonomous vehicles. Information like the availability of vehicles in the blind zone, speed, and distance of the upcoming vehicle can be shared with the drivers beforehand for safety purposes. Computer vision-based lightweight algorithms using simple morphological operations have been proposed to detect the incoming vehicle and estimate the associated speed and distance. Further, an IoT-edge paradigm-enabled computing platform is developed to facilitate efficient computation for latency-sensitive real-time applications. An auto-generated audio-visual alarm guides the driver by a fixed roadside unit near the turning point when the approaching vehicle crosses the predefined threshold zone decided for a particular turning point. Hence, it enables the vehicle to prevent a collision.

Generally, various handheld weapons, such as guns, swords, knives, etc., are used in criminal activities. Further, real-time detection of these weapons using intelligent video surveillance systems can act as a deterrent and legal evidence in smart city applications. Hence, this paper proposes a compact and efficient weapon detector based on the You Only Look Once (YOLOv5)-GhostNet model. A new weapon dataset, “Weapon7,” comprises seven weapon classes such as Axe, Bow and Arrow, Gun, Kinfe, Lathi, Pistol, and Sword, with proper annotation files have been developed. Experimental analysis shows that the proposed model performs better than the equivalent reported works in terms of online performance metrics such as precision, recall, mAP, FPS, and GFLOPS.

Cloud load forecasting predicts future demand on computing resources such as CPU, memory, and network band-width to facilitate proactive resource allocation, cost-effectiveness, and service-level conformance in cloud environments. This paper proposes an attention-driven spatiotemporal deep hybrid model for precise multivariate cloud load forecasting. The proposed architecture incorporates Convolutional Neural Networks (CNN) to extract short-term spatial and temporal trends, Long Short-Term Memory (LSTM) networks to capture long-range dependencies, and a soft attention mechanism to selectively highlight relevant timesteps in past observations. Hence, the proposed model, i.e., the Attention-enabled CNN-LSTM deep hybrid spatiotemporal model for Cloud load forecasting (ACL-Cloud), is trained on a publicly available cloud data center workload forecast dataset in an end-to-end pipeline. Input sequences are built utilizing a sliding window strategy over five normalized resource measures: CPU usage, memory consumption, memory capacity provisioned, and network transmitted and received throughput. Comprehensive experimental results and analysis show that the proposed ACL-Cloud model exhibits better accuracy and generalization in cloud load forecasting tasks, enabling its potential deployment in an intelligent and scalable cloud infrastructure management framework.

Intrusion Detection Systems (IDS) are essential in securing modern networks from advanced cyber threats or intrusions. With increasing network complexity and traffic volume, flow-based intrusion detection has been in the spotlight for scalability and performance. This paper introduces a unified benchmark to evaluate the effectiveness of the boosting and tree-based artificial intelligence models for flow-based network intrusion detection. Six boosting and tree-based machine learning models, such as Decision Tree (DT), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Histogram-Based Gradient Boosting (HistGBM), Categorical Boosting (CatBoost), and Natural Gradient Boosting (NGBoost), are trained and tested on one of the most recent flow-based network intrusion datasets, the HIKARI-2021 dataset. The model development focuses on emulating real-world traffic and feature reduction to mimic realistic deployment conditions. The experimental results demonstrate that ensemble boosting models outperform typical trees in both accuracy and reliability across the board, with some models still displaying more than 98% accuracy after dimensionality reduction. This research offers a holistic guide for choosing proper models to design flow-based IDS and works toward standardizing testing practice in this area.

In recent years, road accidents have become a growing threat to public safety, primarily due to the high stress levels and fatigued mental state of drivers. Monitoring driver stress and predicting when levels may become abnormal could play a vital role in reducing hazards on the road. This study explores machine learning (ML) techniques to process and analyze physiological signals such as heart rate (HR), galvanic skin response (GSR), electromyogram (EMG), electrocardiogram (ECG), and respiratory (RESP) rate, with the aim to train and develop a model to recognize stress patterns as well as predict aspects of stress. Various ML models (i.e., random forest (RF), decision tree (DT), K-nearest neighbor (KNN), support vector machine (SVM), AdaBoost) are used to classify the stress level. The proposed work performs various pre-processing steps to extract important features such as average value, resting value, and standard deviation for analyzing the variability of the signals, considering segmented windows. This work evaluates the performance of various ML models on the benchmark dataset "drivedb"and the RF technique shows improvement compared to existing models with an accuracy of 98.97%. The model analyzes physiological signals and provides higher performance for detecting driver stress compared to state-of-the-art techniques, which will be able to contribute to safer roads.

Data-driven forecasting models are used to understand environmental climatology data better but often overlook missing values and noise, leading to ineffective temporal modeling and inadequate correlations between weather parameters. Consequently, these limitations adversely impact the accuracy of predictions. In response to these issues, a novel deep hybrid model, i.e., a channel attention-enabled bidirectional gated recurrent unit with neural basis expansion analysis for time series (ChAT-BiGRU-NBEATS), is proposed. In this case, a bidirectional gated recurrent unit (Bi-GRU) network is augmented with a channel-attention mechanism and an NBEATS model that facilitates the extraction of complex data features and the prediction of long data sequences. The efficacy of the proposed model is assessed using a comparative analysis against several state-of-the-art deep learning models, utilizing error metrics such as mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and the R2 score. It is evident from the results that the proposed hybrid model surpasses other models in terms of its heightened accuracy.

Video anomaly detection is the process of localizing abnormal video patterns spatiotemporally. The extraction of spatiotemporal features incorporating both local and global spatiotemporal dependency from the video is one of the essential and challenging tasks for video anomaly detection. In order to solve this issue, this article proposes a novel deep reconstruction model-based Self-Attention-enabled Spatio-Temporal Network (SASTNet) model that uses Self-Attention-enabled Convolutional Bidirectional Long-Short-Term-Memory-based Auto-Encoder (SAConvBiLSTMAE) architecture. A novel Self-Attention-enabled Convolutional Bidirectional Long-Short-Term-Memory (SAConvBiLSTM) block is designed to learn the global spatiotemporal dependencies during the frame reconstruction in an end-to-end autoencoder framework for detecting video anomalies. The foundation of the proposed SASTNet model is supported by a comprehensive mathematical problem formulation using a deep reconstruction approach for detecting video anomalies in a single-scene scenario. The selection of an appropriate threshold is one of the crucial factors in video anomaly detection problems. Hence, an efficient strategy has been implemented to select the optimal threshold by utilizing the Receiver Operating Characteristics (ROC) curve for detecting video anomalies. Finally, extensive experimental analysis and comparison with state-of-the-art approaches using three bench-marked single-scene video anomaly datasets, such as UCSD Ped1, UCSD Ped2, and CUHK Avenue, indicate the superiority of the proposed framework in video anomaly detection.

Weapon detection is the process of identifying handheld weapons such as guns, knives, etc., and creating a bounding box around them to highlight the spatial locations. Weapon detection is one of the key building blocks of the intelligent video surveillance system for security applications in smart cities. However, detecting handheld weapons from surveillance videos is quite challenging due to small object size, occlusion, illumination variation, model complexity, and latency. Hence, an efficient, novel, robust, and lightweight YOLOv8-based weapon detector with GhostNet backbone and C3 module with transformer block (C3TR) neck (YOLO-GTWDNet model) is proposed for detecting the weapons either from stored images or from the live video streams. The proposed model is trained using a weapon dataset named “Weapon7,” which is developed by collecting various weapon classes, such as Axe, Bow and arrow, Gun, Knife, Lathi, Pistol, and Sword, from various publicly available datasets, Internet, and own camera capture. Extensive experimental analysis is carried out to demonstrate the effectiveness of the proposed YOLO-GTWDNet model. The proposed model outperforms the state-of-the-art models when compared using both quantitative and qualitative performance metrics. The deployment of the proposed model is expected to bolster public safety significantly, providing city authorities with a powerful tool to mitigate risks and swiftly address potential threats.

Video anomaly detection and localization is one of the key components of the intelligent video surveillance system. Video anomaly detection refers to the process of spatiotemporal localization of the abnormal or anomalous pattern present in the video. The performance of the deep learning-based video anomaly detector depends on the quality and quantity of the video anomaly datasets used for training. However, there is a scarcity of effective video anomaly datasets due to inherent natures such as rareness, context-dependency, and equivocal nature. Further, state-of-the-art lacks a review that presents a comprehensive study of video anomaly datasets, including issues associated with the existing datasets, comparative analysis of the available datasets, potential solutions using both model-centric and data-centric approaches. Hence, a comprehensive review of the publicly available video anomaly datasets for video anomaly detection and localization is presented in this article. Further, a comparative study of the existing video anomaly datasets at qualitative and quantitative levels is presented to decide the right strategies for the desired application. Subsequently, model-centric and data-centric approaches required to solve various problems associated with the video anomaly datasets are presented. Finally, current research trends, research challenges, potential applications, and future research directions are outlined.

The process of automatically finding and localizing the available anomalies (or defects) in the images of the products is known as Image Anomaly Detection (IAD). In this paper, an attention-enabled convolutional autoencoder has been proposed to detect industrial defects using images of the products. The proposed model uses image-wise defect detection. The model classifies each test image as either defective or defect-free based on the magnitude of its reconstruction error. The Structural Similarity Index Measure (SSIM) is employed to assess image quality by quantifying the reconstruction error. SSIM goes beyond comparing individual pixel values and analyzes the inter-relationships between local image regions. This incorporates luminance, contrast, and structural information, providing a more comprehensive evaluation aligned with human visual perception. Comparative result analysis and ablation study validate the superiority of the proposed model.

The process of automatically finding and localizing the available anomalies (or defects) in the images of the products is known as Image Anomaly Detection and Localization (IADL). The IADL improves the efficiency of industrial quality inspection and ensures the desired quality level of the final products. Further, most of the supervised techniques are unsuitable for the IADL due to inherent data imbalance and ambiguity associated with the anomalies. Hence, this paper investigates key deep learning-based unsupervised IADL methods, such as Patch Distribution Modeling (PaDiM), Student-Teacher Feature Pyramid Matching (STFPM), Conditional Normalizing Flow (CFlow), Probabilistic Modeling of Deep Features for Out-of-Distribution and Adversarial Detection (DFM), and Deep Feature Kernel Density Estimation (DFKDE), for three publicly available bench-marked industrial defect detection datasets: MVTec AD, Visa and BTAD. Finally, a comparative analysis using both quantitative and qualitative performance metrics at the image as well as pixel levels is performed to draw some insightful conclusions.

Anomalous event classification automatically identifies anomalous events using the videos in an intelligent video surveillance system. However, anomalous event classification is challenging due to inherent research challenges such as the requirement of high-end computational infrastructure, data imbalances, and data scarcity. Typically, a combination of Convolution Neural Networks (CNNs) and Long-Short-Term-Memory (LSTM) are used to model the spatiotemporal dynamics of the videos for video classification. However, these models have no attention mechanism to boost the relevant spatiotemporal features and discard the irrelevant features. Hence, a Space-Time Attention Model (STAM)-based anomalous event classifier is proposed. The model is trained and validated on the "Anomalous Event Classification 22,"i.e., the "AEC22 dataset"comprising twenty-two anomalous event classes such as abuse, arrest, arson, assault, etc. The STAM is a combined spatial and temporal transformer that takes a series of frames extracted from the input video and predicts corresponding video-level classification as the output. Subsequently, the proposed model provides 92.84% classification accuracy, which is compared with the two state-of-the-art video classification methods to validate its superiority. The proposed model has huge potential for classifying anomalous events in smart city applications.

Spatio-temporal localization of the abnormal patterns in the video is known as video anomaly detection. Video anomaly detection is the most essential building block of any advanced video surveillance-based security application system. Annotation of the normal and anomalous videos at frame level is a tedious, time-consuming, and erroneous task. Hence, recently, weakly supervised video anomaly detection (WSVAD) methods, which use weakly labeled trained videos (or video-level annotations), have been proposed. This paper attempts to investigate and implement eight key state-of-the-art (SOTA) WSVAD methods on two publicly available benchmarked video anomaly datasets such as UCF crime and ShanghaiTech. The eight SOTA WSVAD implemented methods are Multiple Instance Learning (MIL), Robust Temporal Feature Magnitude (RTFM), Anomaly Regression Net (AR-Net), Bi-directional Encoder Representations from Transformers (BERT), Magnitude-Contrastive Glance-and-Focus Network (MGFN), Temporal Self-Attention (TSA), Weakly Supervised Anomaly Localization (WSAL), Prompt-Enhanced Learning (PEL) and Temporal Context Aggregation (TCA). Subsequently, a comparative analysis of these implemented WSVAD methods is carried out to draw some insightful conclusions.

The global energy demand has been increasing exponentially due to population growth, modern lifestyle, and advancement of consumer technology. Energy technologies are currently moving toward renewable energy sources to reduce the impact of global warming. Solar energy is one of the prominent energy sources widely used due to its high-power density and ubiquitous characteristics. It is adopted in a range of versatile applications, among which smart grids, Internet of Things, consumer electronics, and smart agriculture are some major applications. However, dependability of the performance of solar panels on weather conditions is still considered a major drawback in this domain. Several techniques, including machine learning and deep learning, have been implemented to predict the solar energy in the long term and in the short term in the recent past. However, implementing these frameworks in the field requires sophisticated hardware and a significant amount of power. In this chapter, the performance of several multivariate statistical models, such as vector autoregression, vector autoregressive moving average, vector error correction model, mean variance regularization, Bayesian linear regression, and light gradient boosting machine (LGBM), have been investigated to predict the output power of the solar panel. The models have been trained and tested with a publicly available dataset. Principal component analysis has been implemented as feature selection technique for selecting important features from the dataset. LGBM outperforms all other statistical models by achieving a maximum R2 score of 0.84 and a minimum mean square error of 0.15. Subsequently, artificially missing data maximum of up to 15% has been created, which are later imputed using several interpolation techniques, such as linear, cubic spline, pad, and nearest. Attempts have been made to analyze the performance of the models with missing or corrupted data to evaluate the robustness of the models to handle them in the dataset.

Video anomaly detection and localization is the process of spatiotemporally localizing the anomalous video segment corresponding to the abnormal event or activities. It is challenging due to the inherent ambiguity of anomalies, diverse environmental factors, the intricate nature of human activities, and the absence of adequate datasets. Further, the spatial localization of the video anomalies (video anomaly localization) after the temporal localization of the video anomalies (video anomaly detection) is also a complex task. Video anomaly localization is essential for pinpointing the anomalous event or object in the spatial domain. Hence, the intelligent video surveillance system must have video anomaly detection and localization as key functionalities. However, the state-of-the-art lacks a dedicated survey of video anomaly localization. Hence, this article comprehensively surveys the cutting-edge approaches for video anomaly localization, associated threshold selection strategies, publicly available datasets, performance evaluation criteria, and open trending research challenges with potential solution strategies.

Water plays an important role in the livelihood of mankind. Hence, water that is used for agriculture, marine culture, human consumption, etc., should be in good condition to minimize the hazardous effect of water pollution on human health. Rapid unsustainable industrialization, improper huge waste disposal, excess amount fertilizer usage, etc., are responsible for the rapid deterioration of the water quality in rivers and other freshwater bodies. Manual continuous water quality measurement is risky, expensive, and time-consuming. Hence, it is essential to forecast the water quality using statistical time series models. In this paper, three widely used statistical multivariate techniques such as Vector Moving Average (VMA), Vector Auto Regression (VAR), and Vector Auto Regression Moving Average (VARMA), are investigated to forecast water quality parameters like Fecal Coliform (FC), Total Coliform (TC), Biological Oxygen Demand (BOD), Dissolved Oxygen (DO), and the associated Water Quality Index (WQI) of the Ganga River. Most of the previous methods worked on forecasting the future values based on past values of individual parameters without considering the interdependency among the water quality parameters. Here, correlation among each parameter is estimated. Subsequently, the future values of a parameter are estimated based on its previous values and the previous values of its correlated parameters. The proposed research work can help properly manage the water quality of the river Ganga by utilizing the forecasted results for the planning of the pollution control strategies. Finally, it helps improve the quality of human beings by minimizing the health issues caused by water pollution.

Natural calamities are frequent nowadays due to global warming caused by the adverse impact created by unsustainable development and associated environmental pollution. Atmospheric weather is highly influenced by global warming. Hence, the present work predicts five important meteorological parameters responsible for weather conditions, such as temperature, humidity, pressure, wind speed, and wind direction of Bengaluru City, from the respective historical data available from January 2009 to January 2020, using statistical time series forecasting models. The comparative analysis of these statistical models shows that the vector auto-regressive moving average model outperforms other models in predicting all the above mentioned parameters.

Agriculture is inextricably linked to the environment. Climate change has an effect directly on agricultural activities. India gets severely impacted if there is a loss of yields, which affects human lives. Hence, monitoring climate and its impact on the agricultural field is essential for a country like India. This chapter proposes a novel deep learning-based hybrid vector autoregressive-gated recurrent unit model (VAR-GRU model) for weather forecasting involving the four important weather parameters such as temperature, pressure, humidity, and wind speed for the cities of Bengaluru and temperature, pressure, dew point, and wind speed for the cities of Dongsi. The effectiveness of the proposed VAR-GRU model is proven by comparing its performance metrics (MAE, MSE, RMSE, and R2 Score) with that of other baseline models such as LSTM, VAR, GRU, and another hybrid VAR-LSTM model. The outcomes of this research work can help in increasing crop yields by utilizing the weather forecasting results in smart farming applications.

The food security of any country may be jeopardized due to improper management of agricultural insect pests. Accurate pest detection and efficient pest control strategies must be employed in time to grow healthy crops for achieving food security in a country and worldwide. Hence, developing efficient and robust techniques to detect agricultural insect pests using computer vision approaches is one of the essential steps for timely managing insect pests. This chapter presents a short survey of deep learning-based object detection techniques focusing on insect pest detection and associated insect pest image datasets. Subsequently, a transfer learning-based custom You Only Look Once (YOLOv5) model is developed using the publicly available dataset IP102 for detecting agricultural insect pests with the help of computer vision approaches. The hyperparameters of the proposed insect pest detector are optimized using the genetic algorithm-based hyperparameter evolution method. The performance metrics of the proposed insect pest detector are found to be promising.

With the rise in Internet of Things (IoT) technology in recent years the amount of resource requirement of various IoT applications substantially increases. The data generated by various geo-distributed IoT devices is growing continuously. The existing IoT–Cloud paradigm limits the number of various emergent IoT applications as per the requirement. To utilize various IoT applications effectively and efficiently resource decentralization mechanism is a time of need. Fog computing is a novel computing approach for various IoT applications, specifically time-sensitive. Though, the resource-constrained nature of Fog devices in the fog computing model certainly fails to cater to multiple services for these IoT applications. The heterogeneity and dynamicity of the application request from the IoT devices need a quick decision regarding the placement of the application in the fog layer. Hence, optimal allocation of resources is quite essential for providing uninterrupted services to the end-users. This article comprehensively analyzes the different types of possible IoT application models and strategies for allocating resources to these applications. Since the resource allocation problem has already been proven to be a computationally NP-hard problem, finding a non-deterministic algorithm for allocating resources is our target. In this article, we have taken the IoT application placement problem in fog computing (APFC) as a single and multiple objective optimization problems. We have conducted a survey based on single and multiple objectives to address other possible issues in APFC. Finally, the challenges and promising directions for further research are presented.

The process of automatically detecting abnormal video patterns in the intelligent surveillance framework is known as video anomaly detection. However, video anomaly detection is challenging due to inherent research challenges such as equivocal nature, data imbalances, data scarcity, the complex nature of the entities involved in the anomaly, etc. Hence, a self-attention-enabled convolutional spatiotemporal autoencoder is proposed to detect video anomalies efficiently. The proposed Self-Attention-enabled Convolutional Long-Short-Term-Memory Auto-Encoder (SA-ConvLSTM2D- AE)-based video anomaly detector is comprised of three sequential stages: spatial encoder to learn spatial (appearance) features of individual frames, temporal encode-decoder to learn temporal (motion) features of encoded spatial features, and spatial decoder to decode the encoded spatial features for reconstructing the individual frames. Here, the self-attention mechanism is embedded into the convolutional Long Short Term Memory block present in the temporal encoder-decoder section to generate the Spatial-Attention-enabled ConvLSTM block for learning better spatiotemporal features. An efficient threshold selection criteria based on the finding of the optimized Geometric mean value of the sensitivity and specificity from the Receiver Operating Characteristics curve is implemented. The model is trained on only the video frame sequences corresponding to the normal incidents. However, the model poorly reconstructed test frame sequences with video anomalies, as anomalous samples are never exposed during training. Hence, when the anomaly score of individual frames exceeds the selected optimum threshold level, then an anomaly is said to be detected.

Water pollution is a global problem. In developing countries like India, water pollution is growing exponentially due to faster unsustainable industrial developments and poor waste-water management. Hence, it is essential to predict the future levels of pollutants from the historical water quality data of the reservoir with the help of appropriate water quality modeling and forecasting. Subsequently, these forecasting results can be utilized to plan and execute the water quality management steps in advance. This chapter presents a comprehensive review of time series forecasting of the water quality parameters using classical statistical and artificial intelligence-based techniques. Here, important methods used to calculate the water quality index are discussed briefly. Further, a problem formulation for the modeling of water quality parameters, the performance metrics suitable for evaluating the time-series methods, comparative analysis, and important research challenges of the water quality time-series modeling and forecasting are presented.

Video anomaly detection in the surveillance video is one of the essential components of the intelligent video surveillance system. However, anomaly detection remains an ill-defined problem, despite the diverse applications due to its rareness and equivocal nature. A Long Short Term Memory-Variational Autoencoder (LSTM-VAE) model is proposed to detect video anomalies. The model consists of a spatial encoder comprised of convolutional layers, a temporal encoder as well as a decoder comprised of Convolutional LSTM (ConvLSTM), and a spatial decoder consisting of transposed convolution layers. The generative model is trained only on normal video clips with the objective of minimizing the reconstruction error. Then, the trained model is tested on the test video sequences comprised of both normal and abnormal incidents. The reconstruction error corresponding to the test frame sequences having video anomalies will be very high as the model is not trained to reconstruct them. Subsequently, the corresponding frames will have a low regularity score. An appropriate threshold regularity score is set to segregate the anomaly frames from the normal ones. Frames having a regularity score less than the set threshold value are considered as anomalous frames. The model is developed by using one of the publicly available bench-marked video anomaly datasets, i.e., UCSD Ped2. The performance metrics of the proposed model are promising.

Pollution of water is a global environmental crisis faced by many countries. Particularly, India being a developing country, is facing this water pollution across various water bodies such as rivers, ponds, lakes, and groundwater. Further, the river Ganga, the largest river in India, has been highly polluted due to unsustainable industrialization and urbanization across the river belts. Manual water quality measurement and forecasting techniques are time-consuming, tedious, erroneous, and risky. Hence, it is necessary to predict the future quality of water using data-driven approaches for devising strategies to control the pollution levels of the river Ganga. In this paper, a multivariate hybrid model, i.e., Vector Auto Regression - Long Short Term Memory (VAR-LSTM), is proposed for predicting the pollution levels of the river Ganga. The VAR-LSTM is developed by cascading one statistical model, Vector Auto Regression (VAR), with another deep learning model, Long Short Term Memory (LSTM). Here, VAR is used to model the interdependency of various water pollutants using multivariate time series analysis. Subsequently, the fitted values of the VAR model are fed into the LSTM model to explore the temporal feature of the time series water quality data for predicting the water quality. Finally, the proposed hybrid VAR-LSTM model predicts the four water pollutants like Total Coliform (TC), Dissolved Oxygen (DO), Fecal Coliform (FC) and Biological Oxygen Demand (BOD) of river Ganga along with their associated Water Quality Index (WQI). The prediction accuracy of the proposed model is found to be promising.

Video anomaly detection (VAD) is a subject in computer vision that has a variety of applications such as civil protection, surveillance systems, intrusion detection, etc. Anomaly detection remains an ill-defined problem, despite its diverse applications. The proposed model for VAD is a Dual-Stream Variational Auto-Encoder (DSVAE), which consists of two stacked Variational Auto-Encoders (VAE) models. One model is a shallow generative model, i.e., Fully connected VAE (FCVAE) and the other is Skip connected VAE (SCVAE). The FCVAE model tries to learn the overall features of the model and rejects some of the unwanted features. The SCVAE attempts to extract the spatial and temporal features of the image frames in detail. SCVAE also uses the skip connection to connect the features from the encoder and decoder to minimize the information loss. The model is trained only on the normal video clips, and it tries to minimize the reconstruction error. Then the trained model is tested on the test videos. The reconstruction error is very high for testing video clips, resulting in a low regularity scores and a high anomaly. A properly set threshold regularity score segregates the anomaly frame from the normal one. Anomaly frames usually have low regularity score than the threshold value. The proposed DSVAE model is trained and tested on one of the widely used publicly available datasets, i.e., UCSD ped2, and the obtained performance results are found to be promising.

Meteorological variables such as temperature, humidity, and pressure significantly impact living things. Because of the ambiguity and rapid climatic change in the environment, weather prediction with higher accuracy is essential. With the help of deep learning models, the prediction of weather parameters becomes easier and more accurate as compared to traditional methods. This paper investigates various deep learning models such as Long Short Term Memory (LSTM), Gated Recurrent Unit (GRU), Bidirectional Long Short Term Memory (BiLSTM), Bidirectional Gated Recurrent Unit (BiGRU), and Neural Basis Expansion Analysis for Time Series (NBEATS) for the prediction of the temperature of the city of Bhubaneswar. The comparative analysis of these developed models in terms of various performance metrics, such as MAE, MSE, RMSE, and R2 score, concludes that the prediction of the BiGRU model is more accurate as compared to the other implemented models.

Video surveillance systems are popular and used in public places such as market places, shopping malls, hospitals, banks, streets, education institutions, city administrative offices, and smart cities to enhance the safety of public lives and assets. Most of the time, the timely and accurate detection of video anomalies is the main objective of security applications. The video anomalies such as anomalous activities and anomalous entities are defined as the abnormal or irregular patterns present in the video that do not conform to the normal trained patterns. Anomalous activities such as fighting, riots, traffic rule violations, and stampede as well as anomalous entities such as weapons at the sensitive place and abandoned luggage should be detected automatically in time. However, the detection of video anomalies is challenging due to the ambiguous nature of the anomaly, various environmental conditions, the complex nature of human behaviors, and the lack of proper datasets. There are only a few dedicated surveys related to deep learning-based video anomaly detection as the research domain is in its early stages. However, state of the art lacks a review that provides a comprehensive study covering all the aspects such as definitions, classifications, modelings, performance evaluation methodologies, open and trending research challenges of video anomaly detection. Hence, in this survey, we present a comprehensive study of the deep learning-based methods reported in state of the art to detect the video anomalies. Further, we discuss the comparative analysis of the state of the art methods in terms of datasets, computational infrastructure, and performance metrics for both quantitative and qualitative analyses. Finally, we outline the challenges and promising directions for further research.

Water pollution is a serious issue faced not only in India but also in the overall globe. Almost all major Indian rivers are polluted due to the rapid increase in industrialization and poor water quality management. Particularly, the pollution level in the river Ganga has increased significantly. Hence, it is necessary to monitor and manage the pollution levels of the river Ganga using efficient data-driven methods. In this paper, a deep learning-based Convolutional Neural Network - Bidirectional Long Short Term Memory - Support Vector Regression (CNN-BiLSTM-SVR) hybrid model is proposed to forecast the water pollution levels of river Ganga. Four different deep learning models, such as LSTM, BiLSTM, CNN-LSTM, and CNN-BiLSTM, have been developed as a baseline to compare the performance with that of the proposed model. These models are implemented using water quality data of river Ganga collected from the Uttar Pradesh Pollution Control Board's official website. Here, only two parameters, i.e., Dissolved Oxygen (DO) and Biochemical Oxygen Demand (BOD), are used in modeling. The proposed CNN-BiLSTM-SVR model provides better forecasting results for two water pollutants, such as DO, BOD, and the associated Water Quality Index (WQI).

Water pollution is a global problem. In developing countries like India, water pollution is growing exponentially due to faster unsustainable industrial developments. Recently, the river Ganga has been polluted faster and caused lots of diseases among humans and aqua-animals. Hence, continuous water quality monitoring with appropriate water quality management plans is required to maintain sustainable growth. The manual methods of water quality analysis are not suitable in order to get the proper results due to the involvement of life risk and high time consumption. Therefore, it is essential to move towards some advanced data collection, processing, and monitoring approaches that are easy, less costly, and fast. This can be achieved by using data-driven approaches like deep learning techniques due to their strong decision-making ability and automatically learning capabilities from their experience. Hence, a deep hybrid model using Convolutional Neural Networks - Gated Recurrent Units - Support Vector Regression (CNN-GRU-SVR) is proposed to forecast the water quality of the river Ganga using historical data. Here, only two crucial available water pollutants, such as dissolved oxygen and biochemical oxygen demand, collected from Uttar Pradesh Pollution Control Board's official website, are considered for forecasting. The effectiveness of the proposed model is experimentally established by comparing the results with that of the five different deep learning models that have been developed as baseline models.

Water problem is one of the important issues faced across globe, particularly developing countries like India. Hence, there is a need for continuous monitoring and forecasting of water quality with the most advanced techniques having low implementation cost, less time consumption as well as high accuracy. This will help the concerned authorities and governments to plan and implement necessary steps to improve the quality of the water, particularly freshwater available in the rivers. Specifically, the water quality of the river Ganga has been deteriorated to a great extent and requires continuous monitoring as well as forecasting of water pollutants to help in water quality management. Hence, in this article, three widely used time series-based models such as Auto-Regressive Integrated Moving Average (ARIMA), Seasonal ARIMA (SARIMA), and Prophet have been implemented to predict the water quality of the river Ganga. Here, the models are developed on the Uttar Pradesh Pollution Control Board's official data for the river Ganga corresponding to nine water quality monitoring stations situated in Uttar Pradesh. Further, only two important water parameters such as dissolved oxygen and biochemical oxygen demand, are considered for prediction and subsequently for the forecasting of the water quality. The experimental analysis concludes that SARIMA and Prophet model predict the water quality parameters as well as Water Quality Index (WQI) more accurately.

A convolutional spatiotemporal autoencoder is used for video anomaly detection. The proposed model architecture comprises of three major sections, such as spatial encoder, temporal encoder-decoder, and spatial decoder. The spatial encoder is implemented using three layers of the convolutional layers. Then, the temporal encoder-decoder is realized with the help of Convolutional Long Short Term Memory (ConvLSTM), gated with the tanh and sigmoid activation functions. Finally, the spatial decoder is implemented using three layers of deconvolutional layers. The proposed model is trained only on the dataset comprises the normal classes by minimizing the reconstruction error. Later, when the trained model is tested using the test dataset susceptible to contain anomalous activities, then high reconstruction error has resulted. Subsequently, a high anomaly score and low regularity score has resulted. When the regularity score of the frames falls below the set threshold level, then the corresponding frames are treated as anomalous ones. The proposed model is trained and tested on UCSD Ped1 and Ped2 dataset successfully. The results of the performance evaluation are found to be promising.

There is a huge demand of video surveillance based intelligent security systems which can automatically detect the unauthorized entry or mal-intentional intrusion to the unattended sensitive areas and notify to the concerned authorities in real-time. A novel video-based Intrusion Detection System (IDS) using deep learning is proposed. Here, You Only Look Once (YOLO) algorithm is used for object detection and intrusion is decided using our proposed algorithm based on the shifted center of mass of the detected object. Further, Simple Online and Real-time Tracking (SORT) algorithm is used for the tracking of the intruder in real-time. The developed system is also implemented and tested for live video stream using NVIDIA Jetson TX2 development platform with an accuracy of 97% and average fps of 30. Here, the proposed IDS is a generic one where the user can select the region of interest (the area to be intrusion free) of any size and shape from the reference (starting) frame and potential intruders such as a person, vehicle, etc. from the list of trained object classes. Hence, it can have a wide range of smart city applications such as person intrusion free zone, no vehicle entry zone, no parking zone, smart home security, etc.

A deep-learning-based Loitering Detection System (LDS) with re-identification (ReID) capability over a multicamera network is proposed. The proposed LDS is mainly comprising of object detection and tracking, loitering detection, feature extraction, camera switching, and re-identification of the loiterer. The person is detected using You Only Look Once (YOLOv3) and tracked using Simple Online Real-time Tracking with a deep association matrix (DeepSORT). From the trajectory analysis, once the time and displacements thresholds are satisfied, the person is treated as a loiterer. When the loiterer moves one camera to another, then the algorithm is switched to the appropriate camera feed as per the proposed camera switching algorithm to minimize the computational cost. Subsequently, the loiterer is reidentified in the switched camera feed by comparing the features of the loiterer extracted by the MobileNets with that of the other detected persons based on the triplet loss criteria. The proposed system provides an enhanced accuracy of 96 % on average fps of 33 (without ReID) and 81.5 % at average fps of 30 (with ReID).

Ground penetrating radar (GPR) is a special type of radar, which uses electromagnetic (EM) wave to detect the buried objects in the shallow subsurface. The overall efficiency and performance of the GPR system significantly depend on the quality of the acquired data and hence on the antenna and transceiver electronics sensor used in GPR surveying. Bow-tie antennas are widely used for GPR applications, as they possess a set of stringent antenna performances such as a low frequency of operation, ultra-wideband performance, minimal ringing, compact, planner, and lightweight. In this article, a comprehensive review of the bow-tie antennas used for GPR applications is presented. A brief review of the theoretical analysis of the bow-tie antennas indicates the advantageous features of this type of antennas with elliptical structures over the traditional triangular shapes. A comparative analysis of popularly used feed networks confirms that there is a requirement of a balun (balanced to unbalanced) which is used to convert the unbalanced current flow of the coaxial cable to the balanced current flow of two symmetrical lines used for exciting a balanced antenna. A comparative analysis of design techniques used to improve the traditional bow-tie antennas to overcome the various limitations, such as end-fire reflections, narrow bandwidth, dispersion characteristics, low efficiency, and gain, is presented. There is ample scope to propose a novel bow-tie antenna or to improve the existing Bow-tie antennas to achieve compactness, lightweight, reduced end-fire reflections, dispersionless characteristics, better gain and directivity, high radiation efficiency, and so on.

The Substrate Integrated Waveguide (SIW) filter is widely used in various RF communications to receive only the desired frequency with least delay. However, the design of SIW filters having high efficiency, compact size and low cost is still a design challenge. In this research work, a compact SIW bandpass filter with tapered via transition and multiple 'U' shaped slots is proposed for the Ku band applications. The proposed filter uses a tapered via transition to provide a smooth transition from microstrip line (planar structure) to SIW (waveguide structure) with minimal reflection which results in better S11 performance and wider bandwidth. The stopband performance of the proposed filter is improved significantly by introducing four numbers of 'U' shaped slots in the SIW structure which introduces transmission zeros in the upper stopband. The simulated results obtained from HFSS v.14 shows that the proposed filter has improved performance parameters such as low reflection, high isolation, minimal group delay, etc which make it suitable for Ku band applications.

The efficiency of the ground penetrating radar (GPR) system significantly depends on the antenna performance as signal has to propagate through lossy and inhomogeneous media. In this research work a resistively loaded compact Bow-tie antenna which can operate through a wide bandwidth of 4.1 GHz is proposed. The sharp corners of the slot antenna are rounded so as to minimize the end-fire reflections. The proposed antenna employs a resistive loading technique through a thin sheet of graphite to attain the ultra-wide bandwidth. The simulated results obtained from CST Microwave Studio v14 and HFSS v14 show a good amount of agreement for the antenna performance parameters. The proposed antenna has potential to apply for the GPR applications as it provides improved radiation efficiency, enhanced bandwidth, gain, directivity and reduced end-fire reflections.