DR-NTU (Digital Repository of NTU)
Not a member yet
116018 research outputs found
Sort by
Object detection using spiking neural network on event data
No full textIn the era of artificial intelligence, various sensors and intelligent algorithms enable smart systems to perceive and interpret the real world. Object detection is a fundamental task in computer vision that identifies the category and location of objects, which aids smart systems understand their surroundings and make informed decisions. Consequently, it plays a crucial role in numerous applications, including autonomous driving, robotic obstacle avoidance, and surveillance.
Currently, the dominant object detection paradigm commonly utilizes artificial neural networks (ANNs) to process images captured by conventional frame-based RGB cameras. However, this approach faces significant challenges: firstly, traditional RGB cameras struggle in low-light conditions and are prone to motion blur when capturing fast-moving objects, leading to degraded image quality and unreliable inputs; secondly, the high computational complexity and energy demands of ANNs limit their deployment in real-time and resource-constrained applications, restricting their broader use in object detection. To tackle these challenges, this work investigates dynamic vision sensors (DVSs) and spiking neural networks (SNNs) to achieve accurate and efficient object detection. DVSs asynchronously generate events when pixel-wise brightness changes exceed a threshold, enabling high temporal resolution, low latency, and low power consumption. SNNs process information using binary spikes, offering energy-efficient computation. However, effectively handling the sparse and asynchronous nature of event data with SNNs remains a challenge.
This project proposes a Multi-scale Spiking Transformer for Event-Based Object Detection (MSTED), leveraging Spiking Neural Networks (SNNs) to efficiently process event data while preserving spatio-temporal information. The model outlines a Data Conversion module, a Multi-scale Feature Extraction module, and a YOLOX-based detection head. Trained and evaluated on the Gen1 Automotive Detection Dataset, MSTED achieves a mean Average Precision (mAP) of 0.392, demonstrating competitive performance while addressing the limitations of existing methods. The project highlights the potential of SNNs for real-world event-based object detection and provides insights for future improvements in temporal modelling and energy efficiency.Bachelor's degre
Ommatokoita: a graphic novel
No full textOmmatokoita is a graphic narrative covering themes related to familial trauma. It is produced digitally using Clip Studio Paint software in black, white and grey tones. Techniques including sketching, inking, hatching, shading and hand-crafted lettering are used in the development and production of this graphic novel. The story follows Benji, a 30-year-old, whose father has just passed away. As he deals with his father’s post-life arrangements, he reflects on the incidents in his life following a train stabbing incident that renders his arm paralysed, and how his relationship with his father and himself eventually took a turn for the worst. This report explains in detail the structure and production of Ommatokoita, and the inspirations surrounding it. By breaking down these aspects, I aim to explore how introspection as a storytelling device can be effectively incorporated in a graphic novel.Bachelor's degre
AI for sustainability
No full textWith climate change playing a growing role in public policy and driving stricter regulations in corporate sustainability reporting, it is increasingly important for companies to evaluate the carbon footprint of their products. In particular, food production contributes significantly to global greenhouse gas emissions, accounting for 26% of global emissions, of which meat products such as beef have particularly high carbon footprints. Life Cycle Assessment (LCA) offers a comprehensive approach to evaluating the environmental impacts of products, including carbon footprint. However, Life Cycle Assessment is a resource-intensive and complex process, limiting its accessibility. Thus, this project will explore the feasibility of Artificial Intelligence (AI) and Machine Learning (ML) in streamlining the traditional LCA approach by developing an AI-driven model to predict the carbon emissions of meat products. This project includes comparison of different data preprocessing methods utilising ML and AI, alongside testing of a range of regression models to evaluate prediction accuracy. Key findings conclude the use of AI-based clustering in data preprocessing combined with ensemble models yielded the most accurate carbon footprint predictions. The report demonstrates the potential of ML applications in LCA, to provide quick and accessible carbon emission predictions for stakeholders who may lack the resources to conduct their own.Bachelor's degre
The fabrication and mechanical properties of textured alumina-HEO composites
No full textTextured alumina ceramics have gained considerable attention for their improved mechanical properties, particularly fracture toughness and strength, owing to crystallographic alignment. However, the lack of effective interfacial design continues to limit their damage tolerance and overall structural performance. In this study, spinel-structured high-entropy oxides (HEOs) are introduced as novel interfacial phases to enhance the mechanical behavior of textured alumina. Four HEOs with increasing cationic complexity, Al₂(ZnMgNi)O₄, Al₂(ZnMgNiCu)O₄, (AlFe)₂(ZnMgNiCu)O₄, and (AlCrFe)₂(ZnMgNiCu)O₄, were synthesized via direct casting, each forming a uniform single-phase spinel structure. Among them, Al₂(ZnMgNiCu)O₄ exhibited best intrinsic mechanical properties, including a flexural strength of 234.8 ± 26.3 MPa, an elastic modulus of 102.85 ± 19.07 GPa, and a crack initiation toughness of 3.1 ± 0.47 MPa·m⁰·⁵. These HEOs were then employed as interfacial components in the fabrication of horizontally aligned textured alumina via magnetic-assisted slip casting (MASC), where they were found to be evenly distributed along alumina grain boundaries without phase separation. The alumina grain thickness and relative density increased from 4 to 6 constituent cations and slightly decreased at 7, while grain aspect ratio increased and then stabilized. The composite incorporating (AlFe)₂(ZnMgNiCu)O₄ achieved the highest performance, with a relative density of 93.5 ± 1.3%, grain aspect ratio of 4.35 ± 0.99, flexural strength of 452.7 ± 39.3 MPa, elastic modulus of 116.25 ± 19.01 GPa, and crack initiation toughness of 4.16 ± 0.49 MPa·m⁰·⁵. These enhancements are linked to improved interfacial bonding, promoted by liquid-phase sintering, and favorable grain morphology. Overall, this work highlights the potential of spinel HEOs as effective interfacial design elements in textured ceramics, offering a versatile pathway toward the development of structurally robust, damage-tolerant ceramic systems via high-entropy oxide engineering.Bachelor's degre
Optimal liquidity and risk management: the use of CAT bonds
No full textThis article develops a model to study the utilization of catastrophe (CAT) bonds in liquidity and risk management for an insurance company. We consider an insurer who can manage its risk exposure through reinsurance and CAT bonds with index triggers. In addition, the insurer determines a dividend policy to maximize shareholder value; that is, the expected discounted dividends until ruin. We solve the mixed regular-singular stochastic control problem and derive the optimal strategies. We find that, unless significant basis risk is present, the use of CAT bonds generally advances dividend payments, expands insurers’ capacity, and adds substantial value to shareholders. Moreover, the insurer’s ability to expand capacity is crucial for leveraging the benefits of CAT bonds.Nanyang Technological UniversityMinistry of Education (MOE)Submitted/Accepted versionYL is grateful for financial support from the National Natural Science Foundation of China (No. 71932002), the Natural Science Foundation of Beijing Municipality (No. 9242004) and Nanyang Technological University for the hospitality received during his visit. PW acknowledges financial support through a startup grant at Nanyang Technological University and the Singapore Ministry of Education Academic Research Fund Tier 1 Grant (RS12/21). JZ acknowledges the research funding support from the Singapore Ministry of Education Academic Research Fund Tier 1 Grant (RG49/24)
Deep generative learning models for scalable and adaptive route planning in multi-agent fleet management
No full textFleet management integrates strategies and technologies to solve fleet-related challenges efficiently while balancing the objectives of diverse stakeholders such as customers, fleet owners, service providers, and city governors. This thesis focuses on Vehicle Routing Problems (VRPs), which involve planning routes for capacitated vehicles to fulfill customer demands optimally under various constraints. Traditional methods, including exact and metaheuristic approaches, face inefficiencies and scalability issues with increasing problem complexity. Learning-based methods offer promising alternatives, providing high-quality solutions quickly once trained. However, challenges remain in managing extensive solution spaces in multi-objective settings, developing proactive depot generation strategies, and achieving scale generalization with a single training session.
This research addresses these challenges through three studies. The first study introduces a Distance-Based Attention Mechanism (DBAM) that incorporates 2-D spatial positional information to enhance model performance in multi-objective VRPs, demonstrating superior results over traditional and existing learning-based methods. The second study presents a generative DRL framework for proactive depot generation in multi-depot logistics scenarios, showing significant improvements in routing efficiency. The third study proposes a Mix-scale learning framework enabling single-session training for multiple problem scales, achieving comparable or superior performance to scale-specific trained models.
Collectively, these studies advance learning-based fleet management by enhancing reliability, efficiency, and scalability, offering practical, scalable solutions for real-world fleet logistics challenges.Doctor of Philosoph
Software fairness dilemma: is bias mitigation a zero-sum game?
No full textFairness is a critical requirement for Machine Learning (ML) software, driving the development of numerous bias mitigation methods. Previous research has identified a leveling-down effect in bias mitigation for computer vision and natural language processing tasks, where fairness is achieved by lowering performance for all groups without benefiting the unprivileged group. However, it remains unclear whether this effect applies to bias mitigation for tabular data tasks, a key area in fairness research with significant real-world applications. This study evaluates eight bias mitigation methods for tabular data, including both widely used and cutting-edge approaches, across 44 tasks using five real-world datasets and four common ML models. Contrary to earlier findings, our results show that these methods operate in a zero-sum fashion, where improvements for unprivileged groups are related to reduced benefits for traditionally privileged groups. However, previous research indicates that the perception of a zero-sum trade-off might complicate the broader adoption of fairness policies. To explore alternatives, we investigate an approach that applies the state-of-the-art bias mitigation method solely to unprivileged groups, showing potential to enhance benefits of unprivileged groups without negatively affecting privileged groups or overall ML performance. Our study highlights potential pathways for achieving fairness improvements without zero-sum trade-offs, which could help advance the adoption of bias mitigation methods.National Research Foundation (NRF)Cyber Security AgencyPublished versionThis research is supported by the National Research Foundation Singapore and DSO National Laboratories under the AI Singapore Programme (AISG Award No. AISG2-RP-2020-019); by the National Research Foundation Singapore and the Cyber Security Agency of Singapore under the National Cybersecurity R&D Programme (NCRP25-P04-TAICeN); and by the National Research Foundation, Prime Minister’s Office, Singapore under the Campus for Research Excellence and Technological Enterprise (CREATE) programme. Any opinions, findings conclusions, or recommendations expressed in this paper are those of the authors and do not reflect the views of the National Research Foundation Singapore or the Cyber Security Agency of Singapore
OccluTrack: rethinking awareness of occlusion for enhancing multiple pedestrian tracking
No full textMultiple pedestrian tracking is crucial for enhancing safety and efficiency in intelligent transport and autonomous driving systems by predicting movements and enabling adaptive decision-making in dynamic environments. It optimizes traffic flow, facilitates human interaction, and ensures compliance with regulations. However, it faces the challenge of tracking pedestrians in the presence of occlusion. Existing methods overlook effects caused by abnormal detections during partial occlusion. Subsequently, these abnormal detections can lead to inaccurate motion estimation, unreliable appearance features, and unfair association. To address these issues, we propose an adaptive occlusion-aware multiple pedestrian tracker, OccluTrack, to mitigate the effects caused by partial occlusion. Specifically, we first introduce a plug-and-play abnormal motion suppression mechanism into the Kalman Filter to adaptively detect and suppress outlier motions caused by partial occlusion. Second, we develop a pose-guided re-identification (Re-ID) module to extract discriminative part features for partially occluded pedestrians. Last, we develop a new occlusion-aware association method towards fair Intersection over Union (IoU) and appearance embedding distance measurement for occluded pedestrians. Extensive evaluation results demonstrate that our method outperforms state-of-the-art methods on MOTChallenge and DanceTrack datasets. Particularly, the performance improvements on IDF1 and ID Switches, as well as visualized results, demonstrate the effectiveness of our method in multiple pedestrian tracking.Agency for Science, Technology and Research (A*STAR)Nanyang Technological UniversitySubmitted/Accepted versionThis work was supported in part by the Agency for Science, Technology and Research (A*STAR) under its IAF-ICP Program under Grant I2001E0067 and in part by the Schaeffler Hub for Advanced Research at Nanyang Technological University
Assessing the mechanism of elastotaxis of Myxococcus xanthus in response to compressive forces
No full textMyxococcus xanthus (M. xanthus) is a soil-dwelling bacterium known for its complex social
behaviour and coordinated motility. Recent studies suggest thatM. xanthus is capable of mechanosensing
and responding to physical forces in its environment. This research aims to propose a
improved methodology to investigate how compressive forces influence colony morphology by
applying controlled compression to agar substrates and observing resulting changes in colony
shape. Through observational analysis of the deformation patterns and directional growth
responses, this study seeks to better understand its collective behaviour and spatial organisation
Tensile-strained GeSn/Ge rolled-up nanomembrane with enhanced photoluminescence
No full textWe present the development of a rolled-up GeSn/Ge nanomembrane designed to enhance its photoluminescence (PL). The fabrication process involves growing a GeSn/Ge/Si heterostructure on a silicon-on-insulator (SOI) substrate, followed by selective etching to release the nanomembrane. This process results in a rolled-up configuration, which not only relaxes the compressive strain but also achieves tensile strain on the GeSn layer, confirmed by Raman measurements. PL measurements exhibit a redshift in emission peak from 2361 to 2719 nm, indicating a reduction in bandgap energy to 0.456 eV. Additionally, PL intensity increases by 160% compared to the as-grown sample, highlighting the enhanced light emission efficiency owing to enhanced directness of bandgap by tensile strain.Agency for Science, Technology and Research (A*STAR)Submitted/Accepted versionA*STAR, Singapore, Advanced Manufacturing and Engineering (AME) Individual Research Grant (IRG) (M23M6c0099); National Science and Technology Council (NSTC 112-2636-E-194-001)