Project Synopsis
Title: Human Activity Recognition from CCTV
Footage
1. Introduction
Human
Activity Recognition (HAR) plays a crucial role in intelligent surveillance
systems, smart cities, and public safety. With the rapid deployment of CCTV
cameras in public and private spaces, there is a growing demand for automated
systems that can monitor, analyze, and recognize human activities in real time.
Such systems can be used for crime detection, anomaly detection, crowd
monitoring, and workplace safety.
This
project focuses on building a machine learning and deep learning-based HAR
framework capable of recognizing different human activities (e.g., walking,
running, sitting, fighting, loitering) from CCTV video footage.
2. Problem Statement
Traditional
CCTV surveillance relies on human operators to monitor video streams, which is
time-consuming, error-prone, and inefficient. Manual monitoring:
- Leads
to missed incidents due to operator fatigue.
- Cannot
provide real-time alerts.
- Struggles
with large-scale camera networks.
Therefore,
there is a need for an automated activity recognition system that can
process CCTV footage and classify human activities accurately and in real
time.
3. Objectives
- To
preprocess CCTV video data and extract relevant features.
- To
implement deep learning-based models (CNN, RNN, LSTM, 3D-CNN) for
activity recognition.
- To
classify activities such as walking, running, fighting, falling, or
suspicious movements.
- To
generate real-time alerts for abnormal or suspicious activities.
- To
evaluate the system using accuracy, precision, recall, F1-score, and
confusion matrix.
4. Proposed Approach
- Data
Collection & Preprocessing
- Use
public HAR datasets (UCF101, Kinetics, HMDB51, custom CCTV dataset).
- Perform
frame extraction, resizing, background subtraction, and normalization.
- Feature
Extraction
- Apply
CNN-based spatial feature extraction.
- Use
temporal modeling with RNN/LSTM or 3D-CNN for motion features.
- Activity
Recognition Model
- Train
and test deep learning models for activity classification.
- Fine-tune
models with transfer learning (e.g., ResNet, Inception, MobileNet).
- Anomaly
Detection
- Implement
unsupervised models (e.g., Autoencoders, One-Class SVM) for suspicious
activity recognition.
- System
Deployment
- Integrate
the trained model with CCTV video streams.
- Real-time
processing and alert generation for abnormal activities.
5. Expected Outcomes
- A
working HAR system capable of classifying normal activities
(walking, sitting, running) and detecting abnormal activities (fighting,
falling, intrusion).
- Improved
surveillance automation and reduced human workload.
- Enhanced
public safety and security monitoring.
- Real-time
activity recognition with high accuracy.
6. Tools & Technologies
- Programming
Languages: Python
- Deep
Learning Libraries: TensorFlow, Keras, PyTorch
- Computer
Vision Tools: OpenCV, MediaPipe
- Datasets:
UCF101, HMDB51, Kinetics dataset, custom CCTV dataset
- Deployment: Flask/Django
(for web integration), GPU-enabled environment for training
7. Applications
- Smart
city surveillance systems
- Crime
prevention and anomaly detection
- Crowd
monitoring in public places (railway stations, airports, malls)
- Elderly
care and fall detection in healthcare
- Workplace
safety monitoring (factories, construction sites)
8. Conclusion
This
project aims to develop a real-time Human Activity Recognition system
using CCTV footage and advanced deep learning models. The system enhances
surveillance by automatically detecting and classifying human activities,
thereby improving safety, security, and efficiency in monitoring environments.