crowd-vision 0.1.2

An advanced library for crowd detection, tracking, and analysis in videos.

Crowd Vision: Advanced Crowd Tracking & Analysis 🚶‍♂️🚶‍♀️ crowd_vision is a high-performance Python library for detecting, tracking, and analyzing crowds in video streams. It uses a hybrid tracking approach with Kalman filters and appearance features to maintain stable object IDs even in dense and complex scenes.

The entire pipeline is wrapped in a simple-to-use API, allowing you to get detailed crowd analytics from a video file with just a few lines of code.

✨ Features Robust Person Detection: Utilizes a pre-trained Faster R-CNN model from torchvision for accurate person detection.

Hybrid Tracking Algorithm: Combines IOU (Intersection over Union) and appearance-based matching (color histograms) for reliable tracking that resists ID switches.

Kalman Filter Motion Prediction: Predicts object motion to handle temporary occlusions and maintain consistent tracks.

Motion-Aware Crowd Clustering: Uses the DBSCAN algorithm on both spatial and velocity features to intelligently identify and group individuals into crowds.

Easy-to-Use API: Process an entire video and get detailed crowd analytics with a single function call.

Detailed Output: Generates an annotated output video and a CSV log file with frame-by-frame data on crowd size and composition.

⚙️ Installation Installation is a two-step process. For the best results, it's highly recommended to install the core deep learning and computer vision libraries first, as pip can sometimes struggle with their complex dependencies.

Step 1: Install Core Dependencies (Recommended)

Follow the official instructions for your specific system (e.g., with or without a CUDA-enabled GPU).

PyTorch: Go to the PyTorch Official Website and select the correct options for your OS and package manager. This will give you the correct command to run.

OpenCV: pip install opencv-python

Step 2: Install crowd_vision

Once the core libraries are installed, you can install the package directly from PyPI:

pip install crowd_vision

🚀 How to Use Using the library is designed to be straightforward. You provide a path to a video file and an output directory, and the analyze_video function handles the entire pipeline.

from crowd_vision import analyze_video import os

1. Define the path to your video and where to save the results

video_input_path = "path/to/your/video.mp4" output_directory = "results/run_01"

2. Create the output directory if it doesn't exist

os.makedirs(output_directory, exist_ok=True)

3. Check if the video file exists before running

if os.path.exists(video_input_path): print(f"Starting analysis for: {video_input_path}")

# 4. Run the full analysis pipeline
# You can customize many parameters here (see below)
analyze_video(
    video_path=video_input_path,
    output_dir=output_directory,
    confidence_threshold=0.5,
    motion_weight=0.2,
    app_weight=0.5
)
print(f"✅ Analysis complete! Results saved in '{output_directory}'")

else: print(f"❌ Error: Video not found at '{video_input_path}'")

Customization You can tweak the tracker's behavior by passing optional arguments to analyze_video:

confidence_threshold: The minimum score for a person detection to be considered (e.g., 0.5).

tracking_max_age: How many frames a track can be lost before it's deleted (e.g., 30).

motion_weight: How much velocity matters when clustering people into groups (e.g., 0.2).

app_weight: The balance between IOU and appearance for matching tracks (e.g., 0.5).

📊 Output The analyze_video function will generate two files in your specified output directory:

crowd_analysis_output.mp4: The original video, annotated with:

Bounding boxes and unique IDs for each tracked person.

A colored convex hull drawn around identified crowds.

Text indicating the number of people in each crowd.

crowd_analysis_data.csv: A CSV file containing a frame-by-frame log of all detected crowds, with the following columns:

Frame

Time_Seconds

Crowd_ID

Person_Count

Track_IDs (a semicolon-separated list of the IDs of people in the crowd)