Software Optimization for Edge Computing Lab is involved in the following projects.

Real-Time Automatic Object Recognition (AOR) With Edge Capability

The project aims to develop a real-time computer vision system for automatic object recognition and tracking, leveraging machine learning techniques. Using the Jetson Orin edge computing device, the system offers significant advantages, including automatic object identification to reduce human error and reaction time, real-time tracking for continuous awareness of target movement, and edge computing capabilities to ensure low latency and data privacy through on-device processing. This approach enhances operational efficiency and reliability, making it ideal for various applications requiring precise and immediate target recognition and tracking.

---

Real-time Video Transcription and Sentiment Analysis for Unprecedented Monitoring in Media

An AI-Powered, Fast, and Cost-Effective Solution for Converting Video Content into Summarized Text with Sentiment Analysis. Leveraging Cutting-Edge Speech-to-Text and Text Summarization Technologies, it Offers a Valuable Resource for Swift and Affordable Content Analysis in Various Fields, including Media.

---

AI-Powered Web Link Analyzer

AI-Powered Web Link Analyzer focuses on extracting and analyzing web content through web scraping, encompassing diverse website structures or various website formats. The system adeptly cleans and preprocesses raw data using NLP, providing comprehensive website summaries and insightful sentiment analyses.

---

Visual Attendance System

The demand for secure and efficient identification systems has driven advancements in facial recognition technology. Traditional methods often struggle with image quality, spoofing attacks, alignment issues, and computational inefficiencies. The "Visual Attendance System" addresses these challenges by integrating advanced AI technologies for robust face detection and recognition, optimized for edge deployment on customized hardware with secure enclosures. Utilizing FaceBoxes for real-time detection, Laplacian variance for blur detection, and facial landmarks for orientation, the system ensures precise and efficient performance. An anti-spoofing mechanism using depth information and mathematical modeling enhances security. FaceNet512 generates high-dimensional embeddings for accurate recognition, with comparisons made using Cosine Similarity. Specialized hardware boost performance, making this AI-powered solution ideal for secure, scalable, and real-time face authentication applications, setting a new standard for identification systems.

---

Sketch Based Image Retrieval System

Sketch-Based Image Retrieval (SBIR) revolutionizes visual interactions by seamlessly connecting hand-drawn sketches and digital images through AI, enabling users to input sketches for efficient retrieval of visually similar images from extensive databases.

---

Suspicious Person Detection using AI-Powered Body Pose Analysis

The primary objective is the early detection of potential threats, specifically focusing on suspicious persons detection.

• Public Safety and Technology: In the pursuit of public safety, the integration of technology is paramount.
• AI-Powered Azure Kinect and LSTM Unleashed: This presentation delves deep into the extraordinary fusion of Azure Kinect body Tracking and LSTM (Long Short-Term Memory) networks, crafting an AI-powered solution that redefines threat detection dynamics.
• Methodological Pillars: The approach involves advanced feature extraction, LSTM model training, and real-time classification systems. Discover how these pillars form the foundation of innovation in threat detection.
• Innovation in Threat Detection: By exploring the nuances of body pose analysis, this research pioneers a sophisticated method for threat detection.

---

Automatic Car Parking

This project focuses on developing a comprehensive system for autonomous vehicle navigation in parking environments. It begins with the creation of a digital model of the parking area and the labeling of data for training. The input data undergoes preprocessing through cleaning, normalizing, and augmenting. A house detection model using deep learning techniques like CNNs is then developed to identify key elements in the environment. The system also includes target locking for selecting and tracking destinations, such as parking spaces or garage entrances, and path planning to determine the optimal route for efficient and obstacle-free navigation.

---

Deployment of Unmanned Aerial Vehicles (UAV’s)/UGV’s Trajectory Prediction in GPS-Denied Mode

Deploying operations in GPS-denied environments presents significant challenges, potentially leading to mission aborts. To overcome this hurdle, a novel sensor fusion approach is implemented, integrating an AI-based Long Short-Term Memory (LSTM) prediction model with an Extended Kalman Filter (EKF). This fusion enhances path prediction and mitigates mission termination risks. The system showcases autonomous navigation with waypoint missions, demonstrating a rover's capacity for independent navigation. Amid GPS denial, collaboration between the EKF and an Inertial Measurement Unit (IMU) ensures continuous path prediction and seamless mission resumption upon GPS recovery. This technology is pivotal for trajectory prediction in both Unmanned Aerial Vehicles (UAVs) and Unmanned Ground Vehicles (UGVs) operating in GPS-denied modes on edge platforms.

---

Swarm Communication & Control

DeployinSwarm communication and control are essential components for coordinating multiple rovers efficiently. Inter-rover communication is established through ROS2/MQTT over a Wi-Fi router, enabling seamless data exchange among various sensors, including GPS and IMU. For controlling and guidance, a manifest menu has been developed for MAVLink-based rover control. This menu facilitates the guidance of multiple rovers from a central point, enhancing coordination and enabling efficient deployment of swarm robotics systems

Autonomous AI and Decision Support (AAI & DS) lab is actively involved in the following projects.

Simulation-to-Real (Sim2Real) Deployment using Reinforcement Learning Description

For sim2real, it was necessary to understand and learn the software in the pipeline mainly Robot Operating System (ROS2) and its integration with OpenAI Gymnasium library. Pipeline was being explored first using wheeled robot. We worked on wheeled robot first before moving to quadruped because of the complexity of the pipeline. Digital twin of differential drive wheeled robot has been created with camera and LIDAR sensors. Simulations were being done for training it for obstacle avoidance using reinforcement learning.

ML/ DL lab is involved in the following projects.

AI-based Wheat Crop Yield Estimation

The Project Utilizes AI-based Advanced Techniques to Estimate the Wheat Crop Yield Based on Wheat Tassel Counting. It helps to provide Timely and Reliable Crop Yield Estimation for Efficiently Controlling the Imports and Exports to Ensure Sustainable Demand Supply Chain. We have utilized state-of-the-art algorithms to produce results on public dataset. We have also processed and annotated another dataset which was provided by one of the collaborators.

---

AI-based Wheat Disease Detection and Classification

This project utilizes AI-based Advanced Techniques to Detect and Classify the 5 Disease Classes in Wheat Crops (including leaf rust, powdery mildew, stem rust, stripe rust, and smut). We have produced initial results on a self-collected dataset. Also, published a conference paper where we compared two of the most widely used algorithms i.e. CNN and ViT to determine which one works the best for the disease classification problem.

---

AI-based HTML Forms Parser for Air-Gap Networks

This project utilizes the Advanced AI Techniques to extract the useful Information within the Air Gap Network without any Physical Connection. We have successfully demonstrated the proof of concept.

---

Human Face Emotion Recognition System

This project aims to develop a system capable of accurately detecting and categorizing human emotions from facial expressions in real-time, leveraging computer vision and machine learning techniques.

---

Machine Learning Based Resource Utilization for Energy Harvesting Internet of Things (IoTs)

This project aims to develop an AI-based critical infrastructure security system using IoT. It is a heterogeneous network with multiple sensors including cameras, temperature, motion, acoustic sensors etc. Multi-hopping is obtained using ZigBee as a communication protocol.

---

Physical Defect Detection in ICs using AI

This project aims to develop an AI-based system that should detect three different types of defects in ICs i.e. improper texture, corrosion, and black topping. Initial work is started including model selection.

Following R & D projects are executed at predictive analytics and inferencing lab.

Aging Management Analysis and Remaining Life Assessment

This project leverages advanced machine learning algorithms and data analytics to monitor, predict, and manage the aging of industrial assets and infrastructure. By analyzing historical and real-time data, AI can identify patterns of wear and degradation, predict future performance, and estimate the remaining useful life of equipment. This approach enhances the accuracy and efficiency of maintenance schedules, reduces downtime, and extends the lifespan of assets by enabling proactive and data-driven decision-making, ultimately leading to improved safety, reliability, and cost-effectiveness in asset management.

---

Condition Assessment of Rotating Equipment

It employs sophisticated machine learning models and data analysis techniques to monitor and evaluate the health and performance of rotating machinery such as pumps, motors, and turbines etc. By analyzing vibration data, temperature readings, acoustic signals, and other operational parameters in real-time, AI can detect anomalies, predict failures, and provide insights into the underlying causes of equipment issues. This proactive approach allows for timely maintenance and repairs, minimizing unexpected breakdowns, extending the operational life of the equipment, and optimizing overall plant efficiency and reliability.

---

Prediction of Equipment Dynamic Behavior and Accident Avoidance

It involves utilizing advanced machine learning algorithms and predictive analytics to anticipate the dynamic responses and potential failures of industrial equipment. By continuously analyzing operational data such as vibrations, pressures, and thermal conditions etc, AI models can forecast the future behavior of machinery under various operating conditions. This predictive capability allows for the early identification of potential issues that could lead to accidents, enabling preemptive corrective actions. Consequently, this approach not only enhances the safety and reliability of industrial operations but also reduces downtime and maintenance costs by preventing accidents before they occur.

---

Industrial Test Rig Design and Development for data generation, fault simulation and classification

Small scale test rig design & developed, and sensors configured. Full scale design completed and manufacturing under process.

Quantum Machine Learning and Cognitive Computing Lab is involved in the following projects.

Semi prime key breaking through Shor’s Algorithm

The initial implementation of semi-prime key breaking through Shor's Algorithm has been successfully completed using quantum simulators and IBM's online quantum hardware. This breakthrough demonstrates the potential of quantum computing to solve complex cryptographic problems that are intractable for classical computers. The current focus is on error mitigation analysis to improve the accuracy and reliability of the results, addressing the inherent noise and decoherence issues in quantum hardware to refine the algorithm's performance and bring it closer to practical.

---

Grover’s Search Algorithm

Grover’s Search Algorithm has been initially implemented on simulators and online IBM hardware, showcasing the feasibility and functionality of the quantum search algorithm in practical applications. A graphical user interface (GUI) for analyzing 10 qubits has been successfully deployed, facilitating user interaction and experimentation with the algorithm. Currently, efforts are focused on enhancing the algorithm's capabilities for pattern matching in complex data, aiming to leverage quantum computing's potential for significant improvements in search efficiency and performance.

---

Error Mitigation Techniques

Error mitigation techniques in quantum computing are essential due to qubits' susceptibility to errors, which vary by type, such as superconducting, trapped ions, or photonic qubits. Strategies include error correction codes like Surface Code and Shor Code, which protect quantum information by encoding it across multiple qubits, and error suppression methods like dynamical decoupling, which reduces decoherence through periodic pulses. Additionally, techniques like zero-noise extrapolation and probabilistic error cancellation mitigate errors without full error correction. The effectiveness of these methods varies across different qubit platforms, emphasizing the need for tailored error mitigation strategies.

---

Qubit to Qubit Data Transfer

An initial quantum circuit for single qubit data transfer has been developed and deployed on a quantum simulator, marking a significant step in quantum computing research. This foundational work enables the simulation of qubit-to-qubit data transfer, essential for future quantum communication and computation tasks. Alongside the circuit development, a graphical user interface (GUI) has been created to facilitate interaction with and visualization of the quantum processes involved. Ongoing research aims to enhance the circuit's efficiency and reliability, paving the way for more complex quantum data transfer systems.

Computer vision lab has been undertaking the following R&D projects from commissions and as self-initiatives.

Detection of Chemicals Using Nano-material Based Sensor Measurements

Chemical (gas) detector based on a single sensor are being developed indigenously using nano-material. This detector has a higher false positive rate than the desired value. It is desired to reduce the false positive rate using AI techniques and include more gases in the repository for detection.

---

Urdu Hand-written Words Recognition using Deep Learning Techniques

In computer vision and pattern recognition domains, Urdu handwritten words recognition is one of the most complicated and challenging tasks such as accessibility of various handwriting styles, the shape comparability of distinct characters, and the cursive nature of the text. Computer Vision Lab has undertaken this self-initiative project to implement DL techniques for Urdu hand-written recognition words, because Urdu language is broadly spoken and written in the territory of South-East Asia such as India, Pakistan, Bangladesh, and Afghanistan. This has many applications such as document digitization, human-computer interaction, paper examination, online accessibility of old work, automating official duties, data entry forms, signature authentication, postal address interpretation, bank receipts, automated transcription, preservation of cultural heritage, and historical text analysis. The project will initial focus on around 100 Urdu handwritten words. Literature survey, data collection, annotation & labeling, implementation has been completed.

---

Road Pavement Condition Monitoring

Main objective of this self-initiative project is to devise an AI-based pavement condition assessment solution for repair and resources management through accurate detection and efficient analysis, thus ensuring safe and smooth flow of traffic. For this purpose, we will develop a deep learning based approach to figure out the abnormalities in road surfaces. This project is composed of Object detection, Localization (locates objects in image/frame), Classification (predicts category/class of object), Object Counting Via tracking and Report generation.

---

Generating Multi-domain Deep Fake Images using Latent Transformations of Explainable AI

The main objective of this self-initiative research project is to generate CFs explanation using the latent transformations technique. Latent transformation utilize multi-domain mapping to transform image feature and enhance generated image quality using different architecture variations in latent transformations. Major applications of this research include Synthetic data generation, Image-to-image translation, Art work and security (deep fake detection).

---

Facial Recognition based Entry/Exit Logger / System

The main objective of this project is to develop a touch-free biometric attendance system based on facial recognition with a high accuracy, for AITeC building in NCP.

---

Long Range Object of Interest Detection in Visible/IR Videos

This project focus to use an unmanned aerial vehicle (UAV) with a payload of color / greyscale and IR (thermal) cameras for aerial video(s). The payload cameras can operate for video acquisition at a slant range of up to 15 Km under different zoom levels. Currently, a human operator has to sit for hours to manually detect humans from long range visible and IR videos collected from payload cameras aboard a UAV. The project aims to develop an automated solution using AI techniques for this purpose to enhances the system capability.

---

Conversion of RGB Images to Thermal Images

This project has an objective to generate a Thermal image as output of an AI model from in RGB input image to the model.

Data Science Lab has been actively engaged in the following research and development activities assigned from commissions and self-initiated tasks.

Reactor Core Surveillance of NPPs using Artificial Intelligence

The integration of machine learning techniques in reactor core monitoring enables system enhances safety and efficiency in nuclear power plants. By analyzing vast amounts of sensor data, Artificial Intelligence algorithms can predict anomalies, optimize reactor performance, and facilitate proactive maintenance. This approach not only improves operational reliability but also reduces the risk of unforeseen failures. The application of AI in this domain marks a significant advancement in nuclear technology.

---

Model Development of Deep Learning-Based Change Detection from Satellite Imagery

Change detection in satellite imagery is a fundamental process for monitoring changes in the Earth's surface over time. By analyzing images acquired from satellites at different intervals, change detection algorithms identify and highlight areas where significant alterations have occurred, ranging from urban expansion and deforestation to natural disasters and environmental changes. Through preprocessing, image registration, and algorithm implementation, these techniques provide valuable insights for urban planning, land management, environmental monitoring, and disaster assessment, aiding decision-makers in understanding trends and implementing strategies for sustainable development and mitigation efforts.

---

Model Development of Deep Learning Land Cover Classification from Satellite Imagery

Land Cover data are an important input for ecological, hydrological, and agricultural models. That Application is developed by Data Science Lab AITeC from Landsat imagery. However, these data have traditionally had large temporal gaps (~5 years) as they are computationally intensive to create. More temporally granular land cover data are needed for a studying a rapidly changing environment. We used Semantic Segmentation techniques where every pixel in an image is given a label of a corresponding class for Sensing and identifying different classes of land use from satellite imagery to solve the problem.

---

Model Development of Deep Leaning Based Semantic Segmentation of Satellite Imagery Data

Semantic segmentation analyzes satellite images to classify land types, aiding applications like urban planning and environmental tracking. This task faces challenges like low resolution and complex object variations. We explored existing deep learning methods and developed a custom architecture optimized for object detection in satellite imagery.
Status: Proof of concept completed the phase 1 of the project and developed the GUI interface.

---

Smart Capture: Big Data Intelligence & Engineering for Video Storage Efficiency

Surveillance systems generate vast amounts of video data, placing a significant burden on storage capacity. Traditionally, surveillance footage is stored on hard disk drives, and due to limited storage capacities, it often necessitates periodic deletion. To tackle this issue, we have introduced an innovative method named Smart Capture Big Data Intelligence & Engineering for enhancing Video Storage Efficiency.

---

Brain Tumor Classification based on MRI Images

Deep learning is transforming brain tumor classification by acting as a powerful tool for analyzing MRI images. Convolutional neural networks, a type of deep learning algorithm, are particularly adept at this task. By sifting through vast amounts of labeled MRI scans, these algorithms can learn intricate patterns that distinguish healthy brain tissue from various tumor types. Deep learning models offer the promise of improved diagnostic accuracy, surpassing human ability in some cases. Additionally, these models can significantly reduce the time doctors spend analyzing complex scans, freeing up valuable time for patient care. Perhaps most significantly, deep learning's ability to detect subtle abnormalities may lead to earlier identification of tumors, allowing for swifter intervention and potentially improving patient outcomes.

---

Image Steganography: Hiding Text Messages in Images

Steganography, the art of hiding data in plain sight, can be achieved through image steganography. Here, secret messages are tucked away within digital images by slightly modifying pixels in a way undetectable to the naked eye. This allows the image to appear normal while containing a hidden message, potentially useful for covert communication or secure data storage.

---

Development of Prototype of AutoML Tool

The AutoML Web App simplifies the development of AI models through an intuitive graphical interface. It streamlines the entire process, from data analysis both with or without AI assistance to machine learning model creation. Users can easily upload their data and let the app handle the complex tasks, eliminating the need for extensive coding knowledge.

---

GPU based AITEC Cluster System

GPU-based AITEC Cluster System Data Science Lab is a high-performance computing environment designed for data science and artificial intelligence workloads. It utilizes multiple GPU-enabled systems in a distributed architecture to accelerate complex computations by parallelizing training and inference tasks across multiple independent GPU system.

---

Development of ETL Pipeline

The development of an Extract, Transform, Load (ETL) pipeline is critical for managing and processing large volumes of data from various sources. This pipeline automates data extraction, transformation into a suitable format, and loading into a data warehouse. By ensuring data consistency, quality, and accessibility, an efficient ETL pipeline supports robust data analysis and decision-making processes. Its implementation is fundamental in achieving streamlined data workflows and comprehensive business intelligence.

---

Development of Real Time Object Detection using Aerial Imagery

Real-time object detection on aerial imagery involves leveraging advanced computer vision techniques to identify and classify objects from high-altitude images. This technology is pivotal in applications such as surveillance, disaster management, and environmental monitoring. By processing aerial images in real-time, the system provides immediate insights and actionable data, enhancing situational awareness and operational response. The development of such systems represents a leap forward in remote sensing capabilities.

---

Development of Real Time Semantic Segmentation of Objects from Live Camera Feed

The real-time semantic segmentation of objects from live camera feeds employs deep learning to accurately classify and delineate objects within a video stream. This technology is essential in fields such as autonomous driving, robotics, and augmented reality. By enabling precise object identification and context-aware analysis, it enhances the interaction between machines and their environments. The development of this system ensures that live video data can be effectively utilized for real-time decision-making.

Development of Data Mining Tool on Time Series Data

Creating a data mining tool for time series data involves designing algorithms that can uncover patterns, trends, and anomalies in sequential datasets. This tool is instrumental in various domains, including finance, healthcare, and manufacturing, where understanding temporal relationships is crucial. By providing insights into historical data and forecasting future events, the tool aids in predictive analytics and strategic planning. Its development underscores the importance of temporal data analysis in modern data science.

Using ETL Tool to Monitor Real Time System Performance and visualize it on Dashboard

Utilizing an ETL tool to monitor real-time system performance involves continuous data extraction, transformation, and loading into a centralized system for real-time analysis. This setup allows for the visualization of system metrics on an interactive dashboard, facilitating instant performance monitoring and decision-making. Such a tool helps in identifying bottlenecks, optimizing resource utilization, and ensuring system reliability. The combination of ETL processes and dashboard visualization enhances operational transparency and efficiency.

---

Development of Big Data Analytics Ecosystem

The development of a big data analytics ecosystem encompasses the integration of various tools and technologies to handle, process, and analyze large-scale data. This ecosystem includes data storage solutions, processing frameworks, and analytical tools designed to extract meaningful insights from vast datasets. By enabling comprehensive data analysis, it supports complex decision-making and drives innovation across industries. The creation of such an ecosystem is pivotal in harnessing the full potential of big data.

---

Development of Standalone Object Detection on Edge System

Developing a standalone object detection system on edge devices involves implementing efficient machine learning models that can run locally without reliance on cloud resources. This approach is crucial for applications requiring low latency and real-time processing, such as surveillance, smart home devices, and industrial automation. By enabling on-device computation, it reduces dependency on network connectivity and enhances data privacy. The advancement of edge-based object detection represents a significant step towards autonomous and decentralized intelligent systems.