Md Hasibur Rahman

Research

Multimodal intelligence for perception and health

I develop learning systems that combine visual, temporal, spatial, and biomedical evidence, with an emphasis on reliability when observations are sparse, noisy, or incomplete.

Biomedical AI direction

My planned biomedical AI research will examine practical systems that interpret multimodal health data in real-world settings. I am particularly interested in combining imaging, behavioral, and sensor-derived evidence to characterize clinically meaningful variation, support personalized assessment, and remain dependable across patients, devices, and acquisition conditions. The goal is to develop models whose evidence can be validated, interpreted, and translated into useful decision support.

This direction builds on my prior work in radiogenomic imaging and genomic language modeling.

Selected research

STARD-Net architecture for spatiotemporal airborne object detection
STARD-Net: spatial and temporal reasoning for tiny airborne targets.

Computer vision · Detection

STARD-Net

Abstract. Detecting small airborne objects from a moving UAV is difficult because targets occupy few pixels and are easily confused with clutter, camera motion, camouflage, or partial occlusion. STARD-Net addresses this setting with a spatiotemporal architecture that refines weak visual evidence using attention, residual and dilated feature extraction, and temporal context. The work focuses on preserving target evidence across frames rather than treating each image independently, providing a structured approach to reliable small-object detection in dynamic aerial scenes.

Project detailsPaperCode
KRAfT formation-aware multi-object tracking overview
KRAfT: formation-aware association and conservative track recovery.

Computer vision · Tracking

KRAfT

Abstract. Multi-UAV tracking becomes fragile when detections are missing, motion estimates are noisy, and nearby objects create ambiguous associations. KRAfT introduces formation-aware reasoning to preserve identity through these failures. It combines motion prediction with Kalman-residual refinement and conservative recovery, using relationships among neighboring tracks as additional evidence when a single trajectory is uncertain. The resulting framework is designed for coordinated aerial motion, where group geometry can constrain association without replacing object-level dynamics.

Project detailsCode
MultiLiDAR UAV sensing methodology with heterogeneous sensor streams
MultiLiDAR sensing: aligning sparse, asynchronous evidence over time.

Multimodal AI · Sensor fusion

MultiLiDAR UAV Sensing

Abstract. Heterogeneous LiDAR systems produce sparse point clouds with different sampling patterns, timing, range, and reliability. This ongoing research studies how temporal accumulation, candidate formation, and reliability-aware fusion can transform those asynchronous measurements into coherent UAV tracks. The central question is how to preserve useful evidence without allowing stale or low-confidence observations to dominate. The work provides a testbed for multimodal learning under sensor disagreement and missing data; quantitative evaluation is in progress.

Project details
Radiogenomic brain tumor classification problem illustrated with MRI sequences
Radiogenomic classification from complementary MRI sequences.

Biomedical AI · Medical imaging

Radiogenomic MRI Classification

Abstract. This work investigates whether complementary brain MRI sequences can support non-invasive prediction of tumor molecular status. The approach treats each sequence as a distinct source of anatomical evidence and studies their integration under limited, heterogeneous clinical data. Beyond classification, the project motivates a broader biomedical AI agenda: models should remain robust to acquisition variation, expose the evidence behind a decision, and be evaluated for clinical relevance rather than benchmark performance alone. The work was presented as a student research poster and is not claimed as a peer-reviewed publication.

Project detailsCode

Related work in genomic AI, remote sensing, and applied machine learning is listed on the projects page.