ProjectsAugust 1, 2024

Video Action Recognition

Overview

Video Action Recognition is a deep learning project developed for Samsung Prism that classifies human actions in video clips—such as walking, running, and gesturing—using a CNN-based MovieNet architecture. The model processes frame sequences from compressed video datasets and predicts the action label for each clip. Achieving 97% training accuracy, the project demonstrates practical video understanding techniques applicable to surveillance, fitness tracking, and human-computer interaction scenarios.

Tech Stack

The model is implemented in TensorFlow using a MovieNet-inspired CNN architecture designed for spatiotemporal feature extraction from video frames. OpenCV handles video decoding, frame extraction, and preprocessing including resizing and normalization. Python scripts manage dataset loading, training loops, evaluation metrics, and model checkpointing on compressed video inputs.

Highlights

Achieved 97% training accuracy on a compressed video action recognition dataset.
Implemented MovieNet-style CNN architecture for spatiotemporal video classification.
Built a preprocessing pipeline with OpenCV for frame extraction and normalization.
Developed as part of the Samsung Prism program with production-oriented ML practices.

More projects

Python

Deep Researcher RAG System

Multi-agent research pipeline with FAISS vector store, Streamlit UI, and FastAPI backend. Generates comprehensive research reports from a single query using RAG architecture.

Case study View project

Next.js

IdeaClinic Forum

Full-featured forum platform built for DEI, SRMIST using Next.js and Supabase for real-time data and authentication.

Case study View project

React.js

MingleMate

Real-time chat app supporting 100+ concurrent users with Tailwind CSS, DaisyUI, secure JWT auth, and optimized MongoDB queries.

Case study View project