IEEE i-COSTE 2025 – Page 8 – NPS Australia Submission System

VISION BASED TRAFFIC CONTROL SYSTEM FOR PEDESTRIAN SAFETY

This research presents a vision-based dynamic traffic signal control system that leverages computer vision and deep learning (YOLOv8) to improve pedestrian safety and traffic efficiency. Unlike traditional fixed-time signals, the system dynamically adjusts pedestrian crossing times based on real-time pedestrian density, ensuring safe crossings while minimizing vehicle delays. It further incorporates automated violation detection, including red-light running, jaywalking, and stop-line violations, supported by license plate recognition for enforcement. With high detection accuracy (95.4%) and near real-time response, the proposed system offers a scalable and intelligent solution for modern urban mobility. Its contribution lies in combining adaptive traffic light optimization with automated violation monitoring, laying the groundwork for integration into future smart city infrastructure.

Decentralized Access Control Using Blockchain and Smart Contracts for Enhanced Cybersecurity

This paper identifies and outlines the challenges regarding cybersecurity in centralized access control architectures.
The paper proposes a decentralized access control architecture using smart contracts deployed on a blockchain. This system eliminates centralized points of failure and provides automated, tamper-resistant access decision-making using role-based policies.

Clegora – Legal ChatBot for Indian Consumer and Civil Rights

The current paper brings a design, development and evaluation of an AI-based legal assistant named Clegora to suit the requirements of the Indian legal system. However, based on a state-of- the-art large language model, Groq Llama 3.2, deployed in a Retrieval-Augmented Generation (RAG) framework, Clegora delivers contextual and high-quality responses through the dynamic retrieval of relevant legal documents in a specialized vector database. The system uses a new token management approach that scales input lengths based on query complexity with the aim of being efficient in resource utilization without compromising on the quality of responses. Clegora has privacy and ethical standards, as it is built with secure authentication of the user by Firebase and strong content safety measures. Performance tests show short response times of less than five seconds to simultaneous users and citation efficiency of more than 95%. The responses of the users point out how the assistant could be used to simplify complex legal cases, multi-turn conversations, and provide users with confidence when handling legal cases. Though the existing constraints are the processing of language diversity in the region and very narrow domain queries, the modular design and continual data refreshing makes Clegora a scalable solution that will democratize the access to legal information, enabling informed decision-making in diverse Indian populations.

Design and Modelling of a Small-Scale Automated Biogas Digester with Monitoring and Control Capabilities

This research presents the integration of IoT and Mechatronic Systems in a compact, automated biogas digester designed for Small-Scale household use. The system enhances biogas yield efficiency through control of key parameters and real-time monitoring. The system offers a scalable solution for decentralized renewable energy and sustainable waste management in Small Island Developing States.

EfficientNetV2S with End Ensemble for Robust Bangla Handwritten Character Recognition

Recognition of handwritten characters is an important task in the field of image processing. Bangla handwritten characters exhibit greater complexity and variation in shape, along with high inter-class similarity, making their recognition particularly challenging. To address this, we propose the “EfficientNetV2S” model, designed to recognize complex and visually similar characters in the Bangla language. Although Bangla is one of the most spoken languages in the world, research on Bangla character recognition is still comparatively limited. In this study, we apply a deep learning model trained and tested on a custom dataset of Bangla handwritten characters. The model is utilized both for feature extraction and for its proven efficiency in image classification tasks, enabling it to learn image features effectively and recognize characters with remarkable accuracy. Our approach achieves an impressive 96.63% accuracy, highlighting the strength and reliability of the proposed method. To broaden the range of classes, we combined the “BanglaLekha-Isolated” and “Matrivasa-raw (Ekush)” datasets. The results show that the End Ensemble technique effectively tackles recognition challenges and offers strong potential for real-world applications like automation and education. This research advances Bangla character recognition and encourages further exploration in the field.

IGNN: An Attentional Graph-Based Approach for Vision Transformer-Empowered Plant Disease Detection

This study’s primary contribution is the development of a novel ViT-GNN hybrid deep learning model callled IGNN for plant disease classification. By combining the feature-learning prowess of a Vision Transformer with the relational-modeling capability of a Graph Neural Network, the model achieves superior performance and interpretability over traditional methods. This work demonstrates a new paradigm for leveraging both rich visual features and structural dataset relationships, advancing the field of automated plant disease diagnosis.

Comprehensive Analysis of Machine Learning Models on Physical Layer Anomaly in Smart Grids

It contributes a lot to provide information about selection of best machine learning models for anomaly detection.

Techno-Economic Evaluation of Solid-State and Sodium-Ion Batteries in E-Mobility Using a MATLAB Tool

The selection of suitable battery technologies is a key lever in shaping the global transition to sustainable mobility. This study investigates the techno-economic potential of two emerging battery types – solid-state batteries and sodium-ion batteries. While solid-state batteries promise higher energy densities, sodium-ion batteries offer advantages in material availability and potential cost stability due to the abundance of sodium. A MATLAB-based evaluation is used to conduct a holistic analysis focusing on their application within the Tesla Model Y and the VW ID.3. The assessment considers key indicators such as energy density, specific costs, CO2 emissions, and achievable driving range. The results highlight technological trade-offs and demonstrate how different battery choices impact the ecological and economic performance of electric vehicles. Moreover, the flexibility of the modeling approach enables its application across a range of current and future vehicle configurations, supporting strategic decision making in battery selection.

Advanced Machine Learning Models for Prediction, and Performance Optimization in Renewable Energy

The research by Mohammed Alghassab, conducted at Shaqra University, significantly advances renewable energy systems through the application of advanced machine learning (ML) models—Random Forest, Support Vector Regressor, Gradient Boosting, and CatBoost. Key contributions include achieving high prediction accuracy (Gradient Boosting: 94.2% classification accuracy, 1.86 MW RMSE) and perfect scalability prediction (CatBoost: 1.0 accuracy) for solar, wind, hydro, and geothermal systems. These models enhance energy output forecasting, resource classification, and performance optimization by leveraging feature engineering and hyperparameter tuning. The study demonstrates a 12% accuracy improvement and 10% error reduction over baselines, supporting grid stability, cost-efficiency, and CO2 reduction. By addressing intermittency, scalability, and dependability challenges, the research aligns with global sustainability goals, fostering innovation in smart grids and policy-driven energy planning for a low-carbon future.

Data Resilience in Cyber-Physical Systems: Overcoming MQTT Data Loss for Reliable AI Subsystems

The paper introduces and experimentally validates an end-to-end data-resilience layer for MQTT-based CPS that maintains AI subsystem reliability under adverse network conditions. Concretely, it contributes practical mechanisms (loss-aware buffering/retransmission, QoS/retention strategies, deduplication/back-pressure, and gap-tolerant ingestion for AI) and shows, through controlled tests with packet loss and instability, that operational continuity and model performance can be sustained despite communication faults.

Conference Site: IEEE i-COSTE 2025