Abstract

The People Counting System for Smart Building Management is an AI-driven solution designed to detect and track the number of people entering or exiting a designated area, such as a building or room. By leveraging computer vision techniques and real-time object detection, this system enables efficient occupancy monitoring and can be integrated with building management systems to automate actions based on occupancy levels, such as controlling lighting, ventilation, and security. The project was developed as part of the NVIDIA GRIL Training, which explores real-world applications of AI and computer vision technologies. This paper outlines the methodologies used for developing the system, the experiments conducted to test its efficacy, and the results that demonstrate its potential as a smart building management tool.

Introduction

The People Counting System is designed to solve the challenges of monitoring human occupancy in various environments, particularly for smart building management. The system provides real-time tracking and counting of individuals as they enter or exit a building, room, or other specified area. It leverages computer vision and AI technologies to detect and count people, track their movements, and enable automated actions based on the occupancy status.

This project aims to address key issues related to energy consumption and security within smart buildings. By detecting the number of occupants in a space, the system can optimize lighting and ventilation systems, reduce energy waste, and enhance security measures by alerting when unusual activity or overcrowding occurs. Additionally, the system offers a scalable solution to monitor people in real-time, offering critical data for managing spaces such as office buildings, hospitals, malls, and more.

** Requirements:

To implement the People Counting System, the following tools and technologies are required:

• Hardware: A camera (webcam or surveillance camera) for capturing live video feeds.

• Software: Python 3.8+, OpenCV for real-time video processing, TensorFlow for training deep learning models, and Flask for the web interface to display real-time data.

• Environment: The system is best run on a machine equipped with sufficient computational power to process video streams, preferably utilizing GPU acceleration for faster performance.

The project also requires access to a dataset consisting of images of people, annotated with bounding boxes to define the regions of interest (the people) in the images.

Methodology

The People Counting System operates on the principle of real-time object detection using a trained deep learning model. Below is a detailed breakdown of the methodology employed in this project.

** Data Collection and Annotation**

Accurate annotations play a crucial role in training the detection model. To facilitate object detection, images are annotated with bounding boxes around each person. These bounding boxes are used to indicate the location of individuals within an image or video frame.

The following annotation process was followed:

• Tools: CVAT were used to manually annotate images.

• Annotation Format: The bounding box annotations were saved in either the Pascal VOC format (XML) or JSON format, which are compatible with TensorFlow.

*Annotation Categories: Each annotation specifically marks the region around individuals to aid the model in identifying and tracking them.

** Model Selection and Training**

For this system, the object detection model chosen was a pre-trained model available in TensorFlow. The model was fine-tuned with the annotated dataset to optimize its performance for detecting people in video frames.

The key steps in this phase include:

• Dataset Preparation: The annotated images were split into training and validation sets. The training set was used to train the model, while the validation set was used to evaluate its performance.

• Model Training: The training was conducted using the TensorFlow framework. The model was fine-tuned for the task of detecting and counting people in images.

• Model Evaluation: After training, the model was evaluated based on its accuracy, precision, and recall to ensure its reliability in real-world scenarios.

Real-Time Detection and Counting

The system continuously analyses incoming video streams to detect and count people.

Using OpenCV, the following steps are performed:

• Object Detection: The trained model processes each frame from the video feed to detect people in real time.

• Tracking: Once people are detected in a frame, they are tracked as they enter or exit the camera's field of view.

• Counting: The system maintains a count of individuals entering or exiting the area, providing accurate occupancy data.

Integration with Smart Building Systems

Once the People Counting System is operational, it can be integrated into existing smart building systems. This allows automatic actions based on occupancy, such as adjusting lighting or ventilation based on the number of people in a room.

Experiments

To evaluate the system’s effectiveness, several experiments were conducted:

• Testing with Different Room Sizes: The system was tested in rooms of various sizes to check its performance across different spatial configurations.

• Lighting Conditions: The system was evaluated under different lighting conditions, including low-light scenarios, to test its robustness.

• Real-Time Performance: The system was tested for its ability to detect and count people in real-time without lag, utilizing a webcam and surveillance cameras in various environments.

Commands for running the system:

(Ensure ssh server and scp commands for connecting personal systems and DGX servers are run before on terminal.)

Set up the environment on a DGX server:

sudo apt update
sudo apt install python3.8-venv
python -m venv newenv
source newenv/bin/activate
pip install -r requirements.txt

Run the system:
python3 app.py

Results

The People Counting System demonstrated successful real-time people detection and counting across various test environments. Key results include:

• Accuracy: The model was able to detect and count people with a high degree of accuracy (over 90%) in environments with moderate lighting.

• Real-Time Performance: The system processed video feeds in real-time with minimal delay, allowing for effective live monitoring.

• Scalability: The system proved to be scalable, capable of being deployed in larger building systems for broader occupancy management.

Conclusion

The People Counting System successfully demonstrated its capability to detect and track the number of people entering or exiting a designated area. This system offers a robust solution for smart building management, helping optimize energy consumption and enhancing security measures. The integration of machine learning and computer vision into real-time occupancy monitoring presents a valuable tool for modern infrastructure.

Through this project, we gained hands-on experience in applying deep learning techniques to solve practical challenges. The system also provided insights into the importance of real-time processing and the scalability of AI-driven applications in smart buildings. Future improvements could include refining the model for higher accuracy in varying lighting conditions and expanding its use for larger-scale environments.

The complete codebase for the Intelligent Speed Monitoring and Alert System is available on GitHub. Visit the GitHub repository to explore the source code, annotated datasets, and implementation details.

GitHub Link

Thank you for taking the time to go through my project! I am truly grateful for the opportunity to participate in the global Computer Vision Projects Expo 2024 and share my work with such an inspiring community.