Inspiration
Searching for a specific person across hours of CCTV footage can be time-consuming and challenging. Security teams often need to review multiple recordings manually, making it difficult to connect sightings across different cameras. This inspired us to develop SmartFind AI, an AI-powered person search and tracking system designed to make video investigation faster, more organized, and more efficient.
What it does
SmartFind AI aims to help authorized users search for a target person in video footage and retrieve potential matches with relevant timestamps and camera information.
Key features include:
- Target person image upload.
- Person detection and tracking in video footage.
- Face-based matching when facial information is usable.
- Appearance-based person re-identification.
- Similarity search to retrieve potential matches.
- Camera-wise event timelines and evidence review.
- A centralized interface for viewing search results.
The system is designed to support human investigation, not replace human verification. AI-generated matches are potential candidates and should be independently reviewed.
How we built it
SmartFind AI is designed around a computer vision and person re-identification pipeline.
Python and OpenCV: Video input, frame processing, and image analysis. YOLOv11: Person detection. ByteTrack: Tracking detected people across consecutive frames. InsightFace / ArcFace: Face feature extraction and similarity-based matching. OSNet: Appearance-based person re-identification. FAISS: Efficient similarity search over feature embeddings. FastAPI: Backend APIs for handling search requests and results. Database and evidence storage:Organizing event metadata, timestamps, and relevant evidence.
The architecture is designed to evolve toward centralized processing, multi-camera event correlation, cloud integration, and AI-assisted reporting. Individual components and integrations are validated as implementation progresses.
Challenges we ran into
- Handling variations in lighting, camera angles, and image quality.
- Dealing with partially hidden faces and occluded people.
- Balancing search speed with matching reliability.
- Combining face and body appearance information effectively.
- Organizing potential sightings across different cameras and timestamps.
- Managing video processing and evidence storage efficiently.
- Ensuring that uncertain matches are not treated as confirmed identities.
Accomplishments that we're proud of
We developed a clear modular architecture for combining person detection, tracking, face analysis, appearance-based re-identification, and similarity search in one person-search workflow.
We are also focused on making results more useful to investigators by organizing potential matches with timestamps, camera context, and relevant evidence rather than requiring them to review entire recordings manually.
What we learned
Through this project, we explored how computer vision, object tracking, feature embeddings, and vector similarity search can work together in a practical application. We also learned that real-world surveillance footage presents challenges that are not always visible in controlled demonstrations.
We recognized the importance of evaluating false matches, testing under different visual conditions, protecting sensitive footage, and keeping a human in the loop for important decisions.
What's next for SmartFind AI
Our future roadmap includes:
- Improving person matching under difficult visual conditions.
- Expanding from single-camera processing to multi-camera search and event correlation.
- Developing a centralized dashboard for search results and evidence review.
- Integrating cloud services for scalable processing and storage.
- Exploring GenAI-assisted natural-language queries and evidence summaries.
- Improving evaluation, access control, privacy protection, and auditability.
Our long-term vision is to help authorized security teams find relevant footage faster and make video investigation more efficient, transparent, and responsible.
Log in or sign up for Devpost to join the conversation.