CrowdVision: AI-Powered Crowd Analytics
Click to expand
Click to expand
Click to expand
Category:AI & Machine Learning / Web Application
Client:CrowdVision (Personal Project)
Duration:October 2025 - October 2025
Year:2026
My Approach:
Crafting Digital
Excellence
An advanced AI web application capable of estimating crowd sizes with high accuracy. Utilizing dual deep learning models (CSRNet & P2PNet) to generate density heatmaps and precise point localization for real-time crowd analytics.
The application provides two distinct modes of analysis:
- Density Mapping (CSRNet): Best for extremely dense crowds, generating a heatmap to visualize congestion levels.
- Point Localization (P2PNet): Best for precision counting, identifying individual head positions with high accuracy.
Tech Stack
- Core: Python 3.10, PyTorch
- Backend: FastAPI (High-performance ASGI framework)
- AI Models: CSRNet (Contextual Spatial & Channel-wise Attention), P2PNet (Point-to-Point Network)
- Infrastructure: Docker Containers, Fly.io Cloud Deployment
- Frontend: Vanilla JS & CSS for a lightweight, responsive UI
Key Features
- Dual Inference Engine: Users can switch between Density Map or Localization mode based on their image type.
- Robust Deployment: Deployed on Fly.io with optimized resource management (2GB RAM) to handle heavy PyTorch model inference.
- Dynamic Quantization: Implemented model quantization to reduce memory usage by 50% without significant accuracy loss.
- Interactive Visualization: Instant visual feedback with heatmaps and detection overlays.
Challenges & Solutions
One of the main challenges was deploying heavy Deep Learning models to a serverless/containerized environment with limited memory.
Solution: I optimized the Docker image by implementing multi-stage builds and utilized PyTorch's dynamic quantization to fit two complex models (VGG16 backbones) into a cost-effective cloud instance, ensuring fast cold-start times via custom health-check configurations.
