Skip to main content

Academic Vision & Focus

Research Directions

Developing principled, scalable sequence models and representation learning techniques for continuous spatiotemporal data, mobility analytics, and complex real-world dynamic systems.

Primary Research Direction

Efficient representation learning and sequence modelling for complex spatiotemporal data, particularly vehicle trajectories.

My research investigates methods for learning meaningful, compact representations from complex sequential and spatiotemporal data streams. I focus on developing computationally efficient sequence models that capture structural, spatial, and temporal dynamics in mobility systems, intelligent transportation, and multimodal observation environments.

Research Statement

Modern mobility and intelligent transportation systems generate immense volumes of continuous spatiotemporal data. Traditional sequence modelling architectures often suffer from high computational complexity, memory overheads, or inadequate capture of irregular sampling and topological constraints inherent in real-world trajectory streams.

My work focuses on addressing these bottlenecks through principled representation learning—specifically designing gated linear and delta-rule sequence encoders that preserve long-range dependencies while maintaining sub-quadratic inference and training efficiency. By pairing sequence models with spatial topology, these approaches enable scalable trajectory classification, route prediction, and travel time estimation.

Beyond vehicle and pedestrian trajectory learning, my broader research interests span multimodal deep learning, medical imaging analytics, and predictive data science. I aim to create robust, interpretable representation models capable of bridging raw sensory observations with real-world decision making in complex dynamic environments.

Broad Research Domains

Categorized research topics spanning foundational machine learning to intelligent transportation

Spatiotemporal & Trajectory Computing

  • Vehicle Trajectory Representation Learning
  • Pedestrian Trajectory Prediction
  • Spatiotemporal Data Mining
  • Intelligent Transportation Systems (ITS)
  • Streaming Data Processing

Sequence Modelling & Deep Learning

  • Efficient Sequence Modelling
  • Gated Delta-Rule & Linear RNN Encoders
  • Time Series Analysis & Forecasting
  • Deep Representation Learning
  • Attention & Clique Mechanisms

Artificial Intelligence & Applied Analytics

  • Computer Vision & Multimodal Learning
  • Transfer Learning & Domain Adaptation
  • Medical Image Analysis
  • Data Science & Predictive Modelling

Active Research Initiatives

Concrete implementations and experimental frameworks derived from our research inquiries

Trajectory Representation LearningUnder Review

DeltaPath: Gated Delta-Rule Encoding for Efficient Vehicle Trajectory Representation Learning

A gated delta-rule encoding framework designed for efficient, long-sequence trajectory representation learning in intelligent transportation systems.

Challenge: Traditional recurrent models struggle with long-range dependencies, while full-attention Transformers incur quadratic computational complexity when processing dense vehicle trajectory sequences.
Methodology: Formulates a gated delta-rule sequence encoder that dynamically updates memory representations while maintaining linear computational efficiency with respect to trajectory sequence length.
Impact / Findings: Achieved superior representation quality on trajectory similarity and downstream routing tasks while substantially cutting training and inference overhead.
PythonPyTorchSequence ModellingSpatiotemporal Data Mining
Trajectory Prediction & Attention ModelsAccepted

TSCA-Net: Temporal-Spatial Clique Attention for Multimodal Pedestrian Trajectory Prediction

Temporal-spatial clique attention network modeling multi-agent interactions and environmental contexts for robust multimodal trajectory forecasting.

Challenge: Accurate trajectory forecasting in crowded environments requires joint modeling of subtle individual intentions, complex social conventions, and multi-agent spatial cliques.
Methodology: Introduces a temporal-spatial clique attention mechanism that groups interacting pedestrians into topological cliques and extracts coordinated motion features over time.
Impact / Findings: Demonstrated competitive displacement error reductions across benchmark pedestrian trajectory datasets (accepted at MSN 2026).
PythonPyTorchClique AttentionPedestrian Dynamics
Medical Computer VisionPublished

Transfer Learning-Based COVID-19 Detection Using Chest X-Ray Images

A deep transfer learning pipeline utilizing pretrained convolutional representations coupled with a tailored classifier for diagnostic radiography screening.

Challenge: Limited labeled medical imaging datasets during critical health events impede training deep neural networks from scratch without severe overfitting.
Methodology: Leveraged pretrained convolutional feature extractors fine-tuned with customized classification layers and rigorous data augmentation on chest X-ray scans.
Impact / Findings: Achieved robust diagnostic classification performance on multi-class radiography benchmarks (published in IEEE QPAIN 2025).
PythonPyTorchTransfer LearningComputer Vision