Hi everyone,
Enmin Lin and Zijie Ning
Time flies! After weeks of compiling, debugging, and hacking, we have reached the finish line of the Axelera AI Smarter Spaces Project Challenge.
Today, we are thrilled to share the final results of FlowSentry-Wake.
First off: We thoroughly enjoyed this journey. Building on top of the Voyager SDK has been an incredibly fun and rewarding experience. Translating our concept into a fully accelerated dual-stream system on the Metis M.2 has been an intense but rewarding engineering journey. We’re finally seeing the parallel processing power we aimed for.
Here is a quick wrap-up of what makes our final system special, how it performs in the real world, and where we see it going next.
🛠️ The Engineering Highlights
We didn't just want to run a model; we wanted to build a bulletproof system. Here are the key technical milestones we achieved:
1️⃣ Pioneering Optical Flow on Voyager SDK
To our knowledge, we are one of the first teams to successfully build a complete, end-to-end Optical Flow streaming pipeline on this platform.
We deployed EdgeFlowNet alongside YOLOv8, running them in parallel. We engineered custom DataAdapters, bypassed non-symmetric padding constraints without retraining, and wrote custom C++ GStreamer plugins to ensure real-time, low-latency fusion.
2️⃣ The "Brain": Our Custom Triage State Machine
Running two heavy models is great, but making sense of them is harder. We designed a customized Triage Finite State Machine (FSM). It cross-checks the motion volume (from optical flow) with the semantic identity (from YOLO).
For example:
- If YOLO sees nothing, but the optical flow detects a massive moving blob? Alarm. (Physical camouflage defeated!)
- If optical flow detects movement, but YOLO confirms it's just a football rolling by? Ignore. (False alarms eliminated!)
3️⃣ Bridging Edge AI with IoT (Home Assistant)
A security system is useless if it can't yell. We integrated FlowSentry directly with Home Assistant over the local network. The moment our FSM detects a disguised intruder, it triggers a webhook that instantly wakes up the Apple HomePod in the room to blast a siren and send an alarm to you. Zero cloud dependency. Lightning fast.
🏡 Real-World Testing: We Sleep Better Now
The system is no longer just a script on a desk. We have actually deployed it in our own living space for testing.
We tried everything to fool it: walking normally, rolling objects, and yes, crawling on the floor completely covered by a giant black blanket. Standard YOLO went completely blind, but FlowSentry caught the movement instantly and set off the HomePod siren.
Honestly, we sleep much better now. 😂 It’s incredibly reassuring to know that even if an intruder tries to use the most bizarre physical camouflage, our Metis M.2 is watching exactly how they move.
🌍 The Grand Vision: Privacy-First Intelligent Spaces
While we tested this at home, the vision for FlowSentry-Wake goes much further.
Imagine deploying this in:
- Server Rooms & Data Centers: Where any unauthorized, unusual movement (even disguised) needs instant flagging without uploading sensitive footage to the cloud.
- Museums & Art Galleries: Guarding priceless artifacts after hours, immune to shadows or simple camouflage tricks.
- Libraries & Archives: Providing high-level security while ensuring 100% privacy for the patrons, as everything stays strictly on-device.
FlowSentry proves that Edge AI can provide true spatial intuition—understanding the physics of a space, not just the pixels.
🎬 Check Out Our Demo & Code!
Seeing is believing. We’ve recorded a single-take, one-shot demo showing exactly how FlowSentry reacts to different types of intrusions and camouflage in real-time.
📺 Watch the Final Demo Video Here:
https://www.youtube.com/watch?v=MPbFn8jpanw!-->
💻 Explore the Open-Source Code:
https://github.com/mm0806son/FlowSentry-Wake
📖 Catch Up on Our Development Journey:
If you missed our previous updates, you can read the full story of how we built this here:
- 🚀FlowSentry-Wake: Selected for the Axelera AI Smarter Spaces Final 10🎉 | Community
- 🚀 FlowSentry-Wake | Stage 2 Update | Community
- 🚀 FlowSentry-Wake | Stage 3 Update | Community
A huge thank you to the Axelera AI team for providing such a powerful compute platform and an excellent SDK, to the EdgeFlowNet team for their open-source foundation, and to this amazing community for the support and inspiration.
Stay curious, and keep building at the edge! ✨
Enmin & Zijie