My Senior Project utilizes a Jetson Orin NX (16GB) that will be accessible via a weblink to see the camera's perspective in real-time, with the goal of the using the AI computing power to detect objects and people on screen, to label those objects and track their movement.
A self-hosted & privacy-focused security camera system built on edge computing hardware, which all AI inference runs locally on the Jetson meaning there is no data is sent to third-party cloud servers. Detected objects are labeled in real time, events are logged with snapshots and video clips, and everything is viewable through a dashboard in the browser.
Built as a two-semester senior capstone (CSC 495/496) at Concordia University Irvine.
- Real-time yolo26n object detection (TensorRT-optimized) running on-device
- Sub-500ms WebRTC video stream from Jetson to browser
- Rolling buffer recording — 5s pre-detection, 10s post-detection MP4 clips
- Snapshot and event logging per detection (class, confidence, timestamp, bounding box)
- Archives dashboard with search, snapshot viewer, and clip playback
- Google OAuth authentication
- Jetson-to-VM networking over Tailscale (no public IP and/or port forwarding)
- Dockerized frontend/backend on Google Cloud Engine
- Dockerized backend on Jetson Computer
The CSI camera feeds frames into the Jetson, where GStreamer handles capture and YOLOv8 runs inference. Annotated frames are sent over a raw TCP socket through a Tailscale tunnel to a GCE VM (open-cv-frontend.taila8654.ts.net), which re-streams them via WebRTC using aiortc. The same VM hosts a FastAPI backend that handles event storage, snapshot serving, and clip playback, backed by PostgreSQL running in Docker. The React frontend connects to the WebRTC signaling server at /ws/signaling and displays the live feed, with the Archives page pulling from /api/events, /api/snapshots, and /api/clips.
The edge device is a Seeed Studio reComputer J4012 running a Jetson Orin NX 16GB, paired with a Seeed Studio CSI camera. Inference runs locally using yolo26n with TensorRT and CUDA acceleration, and frames are captured via GStreamer. The backend is Python/FastAPI with PostgreSQL and aiortc, all running in Docker on a Google Cloud Engine VM. The frontend is React and Vite. Auth is handled by Google OAuth 2.0, and the Jetson connects to the VM privately over the mesh VPN, Tailscale.
Log in with Google. The Live page shows the annotated real-time feed from the Jetson. The Archives page lets you browse past detection events, view snapshots, and play back video clips. Events can be filtered by class name (person, car, etc.) or date using the search bar.