Jamuna S Murthy

I am an Assistant Professor at Ramaiah Institute of Technology, Bengaluru, India. My research focuses on multimodal video understanding, anomaly detection, and generative video models, with an emphasis on real-time intelligent systems and cross-modal learning.

I completed my Master’s degree at Ramaiah Institute of Technology, where my thesis titled Distributed Framework for Real-Time Twitter Sentiment Analysis and Visualization Framework(TwitSenti) received the Best Master’s Thesis Award in 2017. In 2024, I was honored with the STEM Young Researcher Award.

My work develops practical video intelligence systems for complex real-world scenarios such as FewShot-SPT, ObjectDetect, SarcasNet-99, DHF, TTM, and TwitSenti.

Email  /  CV  /  Scholar  /  Twitter  /  GitHub

Research

My research focuses on multimodal representation learning, video understanding, real-time anomaly detection, and generative models. I study how motion, behavior, semantics, and cross-modal signals from visual, textual, and audio data can be modeled for intelligent real-world systems. My broader goal is to build reliable AI systems for understanding and generating video content in safety-critical and dynamic environments. Some papers are highlighted.

FewShot-SPT is a novel transformer-based framework for few-shot anomaly detection in surveillance systems. It introduces Event-Guided Keyframe Extraction (EGKE), Adaptive Modality Gating (AMG) for multimodal fusion, and adaptive prototypical learning for improved generalization to unseen anomalies. The model achieves 91.6% AUC on UCF-Crime and 84.2% accuracy on XD-Violence, outperforming state-of-the-art methods.

Deployed on NVIDIA Jetson TX2, the system enables real-time park surveillance under edge constraints, effectively detecting incidents such as falls, intrusions, and weapon-related activities in real-world environments.

Introduces a Deep Hybrid Fusion (DHF) algorithm for multimodal video analytics by integrating text, image, and audio features. The framework uses BiLSTM-based modality representation and attention-based fusion to achieve superior performance (95.84% accuracy) across benchmark datasets for humor detection and multimodal understanding.

Proposes a multimodal sarcasm detection framework integrating text (Enhanced-BERT), image, and audio features. A novel adaptive fusion strategy combined with SarcasNet-99 achieves 97.3% accuracy on large-scale video datasets, outperforming state-of-the-art models in sarcasm detection.

A real-time multimodal object detection framework for Advanced Driver Assistance Systems (ADAS), integrating vision and LiDAR-based perception to enable robust environmental understanding. The system leverages YOLOv5 for high-speed and high-accuracy detection, achieving ~95% accuracy while significantly improving inference speed compared to YOLOv3 and YOLOv4. Designed with principles from remote sensing and spatial perception, the framework supports reliable object detection and tracking under real-world conditions, making it suitable for autonomous driving and large-scale inspection scenarios.

Early work on real-time sentiment modeling in streaming environments, integrating NLP pipelines with scalable distributed systems (Apache Storm, Redis). Focused on low-latency inference and interactive visualization for social media analytics.

A real-time Twitter trend analysis framework that integrates streaming data processing with semantic analysis and interactive visualization. The system identifies emerging trends, tracks their momentum, and provides insights into public discourse dynamics on social media platforms.