Inspiration Crowd management is a critical aspect of public safety, especially in high-density areas like stadiums, airports, and public events. Existing solutions often rely on traditional surveillance and manual monitoring, which can be inefficient. We wanted to build an AI-powered system that provides real-time crowd flow analysis to enhance security and optimize crowd movement.
What it does Our system uses a deep learning model to analyze live video streams and detect crowd density and movement patterns. It integrates with NX Meta to process video frames and send actionable insights, such as congestion alerts and flow direction, to security teams in real time. This enables proactive crowd control measures and improves overall safety.
How we built it We set up live video streaming on a server using NX Meta. Developed a custom CNN + LSTM model trained on the PETS2009 dataset for crowd movement analysis. Optimized the model for CPU-only deployment to ensure efficient real-time processing. Created a Python script to process video externally and send results back to NX Meta via API. Explored direct AI model execution inside NX Meta for seamless integration. Challenges we ran into Optimizing the model for CPU-based inference while maintaining real-time performance. Handling large-scale video data without GPU acceleration. Ensuring smooth integration of AI results with NX Meta’s API. Training the model effectively using only bounding box coordinates instead of full video data. Accomplishments that we're proud of Successfully built and deployed a real-time crowd analysis system using NX Meta. Achieved efficient model performance on a CPU-only setup. Developed a unique CNN + LSTM architecture for crowd flow detection. Created a seamless pipeline to process video data externally and integrate results into NX Meta. What we learned The importance of dataset selection and annotation when training deep learning models. Techniques for optimizing AI models for CPU inference. How to work with NX Meta’s API for real-time AI integration. The challenges of real-time video processing in a resource-constrained environment. What's next for Pose Animator Further optimizing the model for improved speed and accuracy. Exploring alternative datasets to enhance generalizability. Adding real-time anomaly detection for identifying unusual crowd behaviors. Developing a user-friendly dashboard to visualize crowd movement insights in NX Meta.
Log in or sign up for Devpost to join the conversation.