Abstract

Road safety remains a critical global issue, particularly in rapidly developing regions like India, which reports over 461,000 road traffic accidents annually. This thesis explores the integration of AIdriven Advanced Driver Assistance Systems (ADAS) and Collision Avoidance Systems (CAS) in public transport to enhance road safety. Increasing urbanization and mixed traffic conditions in Indian cities make such interventions vital for reducing accidents and improving driver behavior.This research investigates how AI-enabled safety systems influence driver behavior, infrastructure planning, and policy frameworks. Using a mixed-methods approach—including field observations, driver interviews, and policy analysis—it assesses the effectiveness of AI interventions, driver adaptation, and socio-technical challenges, and offers policy recommendations. It also examines AI’s potential role in proactive accident prevention. Findings indicate that ADAS significantly improves safety by enhancing driver awareness and response time through real-time alerts and behavioral feedback. However, effective deployment depends on addressing human factors such as training, perceptions of surveillance, and infrastructural readiness. The study also notes concerns about technology acceptance, particularly job security and driver autonomy. The thesis further examines how ADAS can inform policy and urban transport planning. AI-generated insights help identify high-risk zones, optimize infrastructure, and support targeted safety programs. Nonetheless, barriers such as low technological literacy and resistance to behavioral change must be addressed to realize AI’s potential fully. A human-centered AI approach is essential, ensuring technology augments rather than replaces human judgment. This includes incorporating user feedback, adaptive learning, and socio-cultural factors into design and deployment. Engagement with drivers, policymakers, and transport authorities through continuous training is key to fostering acceptance and long-term sustainability. By integrating technological and anthropological perspectives, this thesis contributes to the growing literature on AI and road safety. It offers actionable insights for policymakers, transport planners, and developers, advocating a data-driven yet human-centric approach to safer mobility in India and beyond. It also highlights the ethical and governance challenges of AI, underscoring the need for inclusive policy and accountability frameworks. Ultimately, this study proposes a roadmap for integrating AI into road safety efforts, balancing innovation with human needs to support an equitable and sustainable transportation ecosystem.

Abstract

Whole slide images (WSIs) contain rich information for computational pathology, yet systematic evaluations of cross-organ and multi-cohort generalization remain limited. This thesis addresses these challenges through two complementary studies. In the first study, patch-level convolutional neural networks (CNNs) were trained on 9,792 slides from The Cancer Genome Atlas (TCGA), spanning 11 cancer subtypes across seven organs, to distinguish cancerous from normal tissue. Both within-organ and cross-organ inference were evaluated, revealing that cancers such as breast, colorectal, and liver can be reliably detected by models trained on other organs. Strong transferability was also observed between subtypes within an organ, such as kidney and lung. To investigate these patterns, feature similarity, overlap in high-attention regions, and nuclear geometry were analyzed, all of which showed positive correlations with cross-organ transfer performance. The second study focuses on lung cancer in the Indian population through the introduction of IPD-Lung, a curated dataset of adenocarcinoma and squamous carcinoma cases. Benchmark evaluations were established using multiple instance learning (MIL) models, with additional experiments exploring model transferability from publicly available TCGA-Lung dataset. Stain normalization methods, particularly Macenko, reduced domain discrepancies to some extent but did not fully bridge performance gaps, especially for squamous carcinoma. A domain adaptation strategy based on a gradient reversal layer (GRL) similarly yielded limited improvements. In contrast, models trained directly on IPD-Lung achieved substantial performance gains, with further enhancements obtained by training on expert-annotated regions of interest (RoIs). Together, these studies demonstrate that deep learning models can uncover meaningful cross-organ similarities in cancer histopathology. At the same time, they highlight the difficulty of generalizing across cohorts, even for the same cancer subtype, underscoring the influence of dataset-specific factors such as scanners and staining protocols. Finally, the findings show that expert annotations can substantially improve performance in small and imbalanced datasets, providing a practical pathway to more robust and clinically relevant computational pathology models

Abstract

Mangoes hold immense economic and cultural significance in India and globally, making accurate yield prediction essential for optimizing domestic consumption, enhancing international trade, and supporting farmer decision-making. Traditional estimation methods, such as manual counting, are laborintensive, error-prone, and impractical for large-scale orchards, highlighting the need for automated solutions. In this work, we present a novel time-series vision dataset for mango yield prediction, capturing high-resolution images of 32 trees across eight spatial orientations over two growing seasons to analyze fruit growth and flowering dynamics. Weather data from the nearest weather station was also collected. To reduce manual effort, we developed a semi-automatic annotation pipeline by integrating YOLO for object detection and SAM for precise segmentation. Benchmarking experiments were conducted for both segmentation and detection tasks. Building on these results, we propose a baseline framework for early-stage yield prediction using flowering cues, combining SegFormer-based segmentation with a regression pipeline and incorporating contextual features such as weather and scale. Our multimodal analysis shows that weather factors complement visual features, and their integration further improves prediction performance. The single-image model, based on the SegFormer-B1 encoder, achieved a mean absolute error (MAE) of 7.21, an R2 of 0.80, and a mean squared error (MSE) of 96.83. These results highlight the potential of vision-based models for yield estimation from early-stage flowering cues. To the best of our knowledge, this is the first work to address mango yield prediction using images from the flowering stage in combination with weather data.

Abstract

Groundwater forms a vital component of freshwater resources, particularly in developing countries like India, where borewells are extensively used for agricultural and domestic water needs. However, unsustainable extraction and limited monitoring have led to significant depletion of groundwater levels, exacerbating water scarcity. Traditional methods for measuring borewell water levels are often manual, error-prone, and lack temporal resolution. This thesis presents the design and deployment of an IoT-enabled borewell monitoring system based on a novel string-tension mechanism that automates the measurement of water depth with minimal human intervention, while providing real-time, automated borewell monitoring. The proposed system leverages low-cost sensors and microcontrollers to periodically record water levels, which are transmitted to cloud platforms for storage and analysis. The device is designed to accommodate the geometrical constraints of borewells and withstand practical deployment challenges, including those involving submersible pump infrastructure. Multiple iterations of sensor integration, including FSRs, encoders, and limit switches, are discussed, with emphasis on reliability, calibration, and mechanical design. To ensure data quality, the raw measurements undergo a multi-stage preprocessing pipeline that includes outlier removal, median filtering, and dynamic window-based smoothing. Realworld deployments demonstrate the system’s ability to produce reliable, high-resolution timeseries data across different locations and elevations. The collected data not only enables real-time water level monitoring but also supports longterm groundwater studies. By enabling scalable and cost-effective data collection, the system facilitates greater awareness of groundwater dynamics, promotes rainwater harvesting practices, and contributes empirical inputs for data-driven policymaking and water resource governance.

Abstract

The Internet of Things (IoT) has the potential to completely transform traditional air quality monitoring and management by utilizing low-cost sensors. In urban areas of a developing country like India, air pollution is a critical environmental and public health concern, with Particulate Matter (PM) being among the most hazardous pollutants due to their ability to penetrate deep into the respiratory system. Urban environments face vulnerability due to their multiple emission sources, as well as episodic events like festivals and daily traffic that generate pollution spikes, necessitating effective event detection capabilities. Traditional monitoring stations, while accurate, require substantial investment, resulting in sparse spatial coverage and creating significant gaps in understanding the spatial and temporal variations of pollution across urban areas. This thesis addresses limitations by using two IoT-based monitoring approaches. The first stationary-based approach utilizes data from a dense IoT-based air pollution monitoring network developed and managed by IIIT-Hyderabad across the Gachibowli region. Analysis of over 10 million data points collected during 2021 and 2022 enabled detailed event-specific analysis, particularly examining the impact of Diwali fireworks and vehicular traffic on PM concentrations. Spatial interpolation techniques identified pollution hotspots, while temporal analysis revealed significant seasonal variations and daily patterns. Results demonstrated that PM10 levels during Diwali increased 2-3 times above normal ranges, while peak vehicular traffic hours consistently elevated pollution levels, with winter showing the highest concentrations compared to summer and monsoon seasons. The second mobile-based approach presents an IoT-based mobile air pollution monitoring system that was developed and deployed on four college buses covering different urban routes across Hyderabad. Over six consecutive months, more than 500,000 route-specific, multi-season measurements were collected across 25 distinct urban regions. Spatial visualization and temporal analysis revealed substantial variations in PM levels based on location, traffic density, construction activities, season, and time of day. Morning measurements consistently showed higher PM10 concentrations (often exceeding 350 µg/m³) compared to evening measurements, attributed to rush hour traffic and atmospheric conditions. Multiple pollution hotspots were identified across the city, with regions experiencing high traffic and ongoing construction showing PM10 levels frequently in the ’poor’ to ’very poor’ ranges. The combined findings from both stationary and mobile IoT monitoring approaches demonstrate scalable and cost-effective IoT-based methodologies for fine-grained air pollution monitoring in rapidly urbanizing regions, highlighting the potential of low-cost IoT sensors to complement traditional monitoring stations and enable a comprehensive understanding of urban air pollution.

Abstract

Deep learning models are increasingly being used at particle colliders for various tasks like jet classification, anomaly detection, detector unfolding, etc. In this thesis, I investigate the efficacy of three machine learning models using various different representations of collider data as per the models’ strengths. I compare boosted decision trees, deep neural networks, and graph neural networks for the event classification task of identifying a typical Beyond Standard Model (BSM) signal coming from 𝑇𝑒𝑉-scale particles. The signal of interest to us here is that coming from a pair-produced heavy bottom quark partner, denoted by 𝐵, which decays through a new pseudoscalar/ scalar boson, denoted 𝜙, similar to the Higgs boson in the Standard Model. Such signatures have been gaining interest recently; Vector Like Quarks (VLQs) themselves are being searched for by the prominent collider experiments ATLAS and CMS at the LHC, but their signatures elude us so far. We perform the analyses for a significantly challenging final state where the subsequent decays of BSM particles are fully hadronic, that is, consisting of only objects arising from quark and gluon jet hadronization; no objects like electrons, photons, or muons exist to give a good handle on the signal. We see that basic graph neural networks significantly outperform other classical ML models. We further improve our results by using methods that can only be deployed in a graph neural network. Therefore, we conclude we can enhance future collider searches for 𝑇𝑒𝑉-scale particle arising from new physics phenomena. The research methodology leverages state-of-the-art simulation tools, including MadGraph5, Pythia8, and Delphes3. We scan over a wide range of parameter configuration to perform robust evaluations of the ML/DL models and propose novel training strategies, opening new avenues for sophisticated particle detection techniques in high-energy collider experiments.

Abstract

Earthquake detection systems require robust discrimination between seismic signals and noise to mitigate risks and enable timely responses. Traditional methods, such as the Short-Term Average/Long-Term Average (STA/LTA) algorithms and deterministic machine learning models, often struggle with noisy data and fail to quantify uncertainty in their output—a critical gap for high-stakes decision-making. This thesis proposes a Bayesian Convolutional Neural Network (BCNN) framework that advances seismic signal classification, specifically, the binary discrimination of earthquake signals from noise waveforms by integrating uncertainty quantification. The proposed framework models both epistemic and aleatoric uncertainties. Epistemic uncertainty stems from limited or unrepresentative training data and can be reduced with additional data. Conversely, aleatoric uncertainty captures the inherent variability in data, such as low signal-to-noise ratios (SNR), overlapping features between earthquake and noise waveforms, and complex wave propagation effects; and is irreducible. By quantifying these uncertainties, the BCNN framework identifies low-confidence predictions, enabling more reliable decision-making in earthquake early warning and monitoring systems. A preliminary comparative study evaluated three architectures, namely, Fully Connected Neural Networks (FCNN), Recurrent Neural Networks (RNN), and Convolutional Neural Networks (CNN) on the Southern California Earthquake Data Center (SCEDC) dataset, achieving validation accuracies of 94.7%, 96.0%, and 99.3%, respectively. Given CNNs superiority in capturing spatiotemporal patterns, it was selected as the foundation for the Bayesian framework. Leveraging Flipout layers and variational inference, the BCNN models uncertainty, providing confidence estimates alongside classifications. Trained on the Southern California Earthquake Data Center (SCEDC) dataset, the model achieves 99.1% accuracy, while identifying ambiguous cases through entropy-based metrics. Comparative analyses reveal its superiority over deterministic CNNs, particularly for low SNR and complex waveform variability. The framework is evaluated across diverse tectonic settings, including Himalayan earthquakes and near-/far-fault waveforms. While it maintains strong performance for data resembling its training distribution (e.g., near-fault signals), its accuracy declines for geologically distinct regions (63.3% for Himalayan events, 68.4% for near-fault and 86.7% for far-fault), highlighting the need of region-specific training. Uncertainty estimates correlate with classification errors, enabling proactive flagging of low-confidence signals for human review. This work offers a scalable, interpretable tool for disaster risk reduction. The BCNN’s probabilistic outputs not only enhance reliability but also align with ethical AI principles, ensuring transparency in life-critical applications.

Abstract

Control software plays a pivotal role in orchestrating the interactions of industrial systems, including sensors, actuators, robots, and machinery, to ensure real-time coordination and reliability. Developing such software, however, is a complex and resource-intensive process, requiring significant time, cost, and specialized domain expertise. For example, the control software development for the Square Kilometre Array (SKA) radio telescope, which manages thousands of devices such as antennas and sensors, required over 100 engineers and spanned over five years. Furthermore, hardware upgrades, maintenance, and evolving operational requirements necessitate frequent reconfiguration of control software in industrial systems. Due to production demands or machine replacements, such reconfigurations may occur multiple times per month. Manual development and reconfiguration of control software are inherently time-consuming, resource-intensive, and heavily reliant on domain expertise. Automating these processes is essential for scalable, cost-effective, and rapid development. A significant challenge in automation lies in capturing and reusing domain-specific knowledge concerning device behaviors and interactions. Current generative AI (GenAI) technologies lack domain-specific knowledge, explainability, and robust reasoning capabilities to address the precision and constraints in reconfiguration scenarios. This thesis addresses these challenges by automating control software development and reconfiguration using domain knowledge. The knowledge of control components is systematically captured through a Capability Ontology. Domain-specific languages (DSLs) were developed to leverage this knowledge, providing a syntactic interface to describe abstract requirements, device capabilities, and control designs. A synthesis algorithm automatically generates control designs, translated into executable software via a code-generation engine. These components are integrated into the Knowledge-aided Integrated Development Environment (K-IDE). K-IDE facilitates the creation of new device capabilities and workflows, enabling the generation of control software from scratch. Additionally, it supports reusing existing knowledge to reconfigure systems and produce updated software. By combining the synthesis algorithm with domain knowledge, K-IDE automates the development and reconfiguration of control software.To bridge domain knowledge with executable artifacts, the thesis also introduces a domainspecific Knowledge Representation Language (KRL) that translates semantic models into multiple target representations such as code and visualization formats. K-IDE was validated across three industrial setups: (i) a warehouse robotics system with over 200 components, (ii) an automated vehicle entry system comprising 150 devices, and (iii) a smart meeting room with 50 interconnected devices. Validation results demonstrated significant efficiency improvements, reducing development time by 60% and reconfiguration effort by 50% compared to manual development approaches. In addition to technical validation, the usability of K-IDE was evaluated through user interviews, questionnaires, and practical deployment, indicating substantial gains in efficiency, flexibility, error reduction, and ease of adoption across diverse engineering roles.

Abstract

Spoken grammar refers to the grammatical structure in the spoken language. It is essential for clear and effective oral communication. Thus, it is an important skill in language acquisition. Spoken Grammar Error Detection (SGED) involves identifying grammatical errors in spoken utterances, and is a necessary component of computer-assisted language learning (CALL) systems that aim to provide learners with automatic feedback. Traditionally, SGED has been implemented through a cascaded pipeline where Automatic Speech Recognition (ASR) converts speech to text, followed by a text-based grammatical error detector (GED). However, this approach is limited by ASR error propagation, where recognition errors and language-model biases can overwrite or mask the learner’s actual grammatical mistakes, leading to sub-optimal performance. In this work, we systematically study statistical and end-to-end ASR systems for SGED under three speaking practices, which are commonly employed in CALL systems: memorised, semi-spontaneous, and spontaneous speaking practices. In memorised and semi-spontaneous speaking conditions, the intended grammatically correct text is known in advance, allowing us to study how different ASR-derived cues behave when learners produce incorrect forms. In the statistical ASR setting, we analyse pronunciation and alignment-based features such as Goodness of Pronunciation (GoP) scores and Force Alignment (FA) likelihoods, using the known grammatically correct text reference to examine how these cues vary with grammatical deviations. In the end-to-end ASR case, we similarly use Connectionist Temporal Classification (CTC) aligner probabilities computed against the ground truth grammatically correct text to understand how alignment behaviour changes when the spoken utterance deviates from the expected grammatical form. For spontaneous speaking condition, where no reference text is available, most of the existing works rely on the traditional cascaded system consisting of ASR and GED approach relying on the ASR decoded text; however, this design is constrained by ASR error propagation and language-model bias. To overcome the inherent limitations of cascaded SGED, we propose a method that directly transforms raw speech into a textual embedding space, termed unified embeddings, for SGED in two settings: (1) an end-to-end approach that predicts grammaticality directly from speech, and (2) an end-to-end fusionbased approach that combines unified embeddings with ASR transcript embeddings to leverage both modalities jointly. The unified embeddings are obtained by distilling knowledge from a pre-trained text encoder into a speech encoder via speech-text unification. To the best of our knowledge, this is the first work to explore both end-to-end and fusion-based approaches for the SGED task. Experiments show that the proposed unification-based fusion approach outperforms the cascaded baseline and endto-end SGED, highlighting that unified embeddings act as an effective complementary information for mitigating ASR errors in SGED

Abstract

Human communication is inherently multimodal, asynchronous, and goal-driven, relying on a fluid interplay of speech, gesture, and sketching to establish shared understanding. While Artificial Intelligence has made substantial progress in static vision and turn-based dialogue, the challenge of modeling continuous, real-time collaboration remains largely unaddressed. This thesis introduces SKETCHTOPIA, a comprehensive framework designed to benchmark and investigate these complex dynamics. First, we present the SKETCHTOPIA dataset, a large-scale corpus of over 20,000 collaborative Pictionary-style sessions. Unlike prior resources, this dataset captures the complete temporal evolution of interaction, including 263,000 vector strokes, 56,000 timestamped guesses, and 19,400 iconic feedback events, thereby enabling the study of communication as a continuous process rather than a discrete exchange. To quantify this process, we introduce a novel suite of evaluation metrics, including the Feedback Responsiveness Score (FRS) and Multimodal Action Timing Similarity (MATS), which measure the “collaborative rhythm” and responsiveness of an agent, moving beyond simple task success rates. Second, we propose and implement foundational AI agents capable of operating within this asynchronous environment. We introduce a novel hierarchical architecture centered on the ACTIONDECIDER module, which decouples low-latency state monitoring from high-level action execution. This allows agents to break free from rigid turn-taking and engage in proactive, event-driven interaction. We develop DRAWBOT, a generative agent using conditioned diffusion models for incremental sketching, and GUESSBOT, a retrieval-based agent that emulates human interpretation strategies. Finally, we conduct an empirical evaluation comparing these agents against human baselines and state-of-the-art Vision-Language Models (VLMs). The results highlight role-specific differences in agent performance: while GUESSBOT achieves win rates close to those observed in human–human sessions (81.7%), DRAWBOT exhibits difficulty in communicating abstract concepts, achieving a 30% success rate on adjectives. We further observe that evaluated general-purpose VLMs face practical limitations in this setting due to their turn-taking nature, inference latency, and verbosity. Overall, this thesis establishes an experimental foundation, dataset, and analytical tools for studying dynamic, feedback-rich collaboration, to enable more systematic investigation of asynchronous multimodal interaction in human–AI systems.