“A Mask Detection Smart Camera Using the Nvidia Jetson Nano: System Architecture and Developer Experience,” a Presentation from BDTI and Tryolabs
For the full video of this presentation, please visit: https://www.edge-ai-vision.com/2021/07/a-mask-detection-smart-camera-using-the-nvidia-jetson-nano-system-architecture-and-developer-experience-a-presentation-from-bdti-and-tryolabs/ Evan Juras, Computer Vision Engineer at BDTI, and Braulio Ríos, Machine Learning Engineer at Tryolabs, co-present the “A Mask Detection Smart Camera Using the Nvidia Jetson Nano: System Architecture and Developer Experience” tutorial at the May 2021 Embedded Vision Summit. MaskCam is a prototype reference design for a smart camera that counts the number of people wearing masks in its field of view and reports statistics to the cloud. Based on the Nvidia Jetson Nano, all AI processing is handled on the device. In this presentation, BDTI and Tryolabs share their story of developing MaskCam, and how the two companies went from concept to working prototype with a small team in just a few months. Juras and Ríos explain the hardware and software design of the camera and present their companies’ experiences working with Jetson Nano (both the Dev Kit and the system-on-module), NVIDIA’s DeepStream framework and containerization via balenaOS/balenaCloud. The solution features a custom tracking component, static video file serving, video streaming over RTSP, and MQTT communication to report statistics and receive commands using a web interface. Working with Jabil Optics, Juras and Ríos also present estimated manufacturing costs for volume production.
Recommended
Recommended
More Related Content
What's hot
What's hot (20)
Similar to “A Mask Detection Smart Camera Using the Nvidia Jetson Nano: System Architecture and Developer Experience,” a Presentation from BDTI and Tryolabs
Similar to “A Mask Detection Smart Camera Using the Nvidia Jetson Nano: System Architecture and Developer Experience,” a Presentation from BDTI and Tryolabs (20)
More from Edge AI and Vision Alliance
More from Edge AI and Vision Alliance (20)
Recently uploaded
Recently uploaded (20)
“A Mask Detection Smart Camera Using the Nvidia Jetson Nano: System Architecture and Developer Experience,” a Presentation from BDTI and Tryolabs
- 1. © 2021 BDTI and Tryolabs MaskCam: A Jetson Nano AIoT Mask Detection Camera Evan Juras Braulio Ríos BDTI Tryolabs
- 2. © 2021 BDTI and Tryolabs Agenda 2 About us MaskCam, an AI-enabled smart camera based around Jetson Nano Software design and containerization Hardware design and high volume productization Conclusions and resources
- 3. © 2021 BDTI and Tryolabs About us 3
- 4. © 2021 BDTI and Tryolabs MaskCam: A mask detection smart camera COVID-19 creates a need to monitor crowd size and face mask usage in public areas MaskCam provides mask-wearing statistics in an indoor or outdoor area Reference design for AI-enabled embedded vision applications 4
- 5. © 2021 BDTI and Tryolabs MaskCam in action! 5 https://youtu.be/VofHurI7EZE
- 6. © 2021 BDTI and Tryolabs Original camera video: courtesy of EarthCam Pose Estimation + Classification Object Detection Object classes: mask, no_mask, not_visible Tracking (Norfair) Software: AI model selection
- 7. © 2021 BDTI and Tryolabs FP-16 Software: Optimized object detection
- 8. © 2021 BDTI and Tryolabs Hardware acceleration & parallelization Why DeepStream Pipeline? Maximum performance (inference bottleneck) Fast development (Python) + high performance (C, CUDA) Software: Video pipeline
- 9. © 2021 BDTI and Tryolabs ● Web server • Postgres DB, FastAPI backend, Streamlit frontend ● Remote communication • MQTT protocol: send statistics, receive commands ● Video streaming, save video snapshots • Python multiprocessing • UDP video packets for internal sharing Other Requirements
- 10. © 2021 BDTI and Tryolabs Why containers? Thin layer update for source code changes Dependencies bundled together Better reproducibility Hardware decoupling Easy over-the-air updates Easily link to balenaCloud Containerization
- 11. MaskCam Hardware and Productization
- 12. © 2021 BDTI and Tryolabs Hardware: Design 12 Carrier Board Camera Module Connectivity Module (WiFi or 4G LTE) Jetson Nano SOM Primary components ● Jetson Nano SOM ● Carrier board ● Camera module ● Connectivity module NVIDIA resources ● Jetson Partner Hardware Products list ● Jetson Partner Supported Cameras list ● elinux.org/Jetson_Nano
- 13. © 2021 BDTI and Tryolabs Hardware: Thermal testing 13 Heatsink Final CPU Temp Final GPU Temp Average Inference FPS Passive 97°C 93°C 13.1 Active 48°C 45°C 14.2 Conclusion: final production design requires active heatsink No fan: thermal throttling causes Degraded performance Active fan: stable temperature, maximum performance
- 14. © 2021 BDTI and Tryolabs From prototype to high-volume production 14 Developer Kit Off-the-shelf prototype kit Production model Worked with Jabil to determine next steps for turning MaskCam into a high- volume product and get a preliminary cost estimate
- 15. © 2021 BDTI and Tryolabs Estimated BOM and production cost 15 Item Cost (10K/year, USD) Cost (100K/year, USD) Jetson Nano $99.00 $89.00 Camera Module $42.36 $33.84 Carrier Board $54.82 $53.13 Enclosure, Fan $41.20 $35.29 Total BOM Cost $237.38 $211.26 Production Cost (MVA) $118.69 $84.51 Overall Cost $356.07 $295.77
- 16. © 2021 BDTI and Tryolabs MaskCam is open source! 16 github.com/bdtinc/maskcam (MIT License) Step 1. Pull the Docker container Step 2. Run MaskCam nano@nano-desktop:~$ sudo docker pull maskcam/maskcam-beta nano@nano-desktop:~$ sudo docker run --runtime nvidia --privileged --rm -it -- env MASKCAM_DEVICE_ADDRESS=<your-jetson-ip> -p 1883:1883 -p 8080:8080 -p 8554:8554 maskcam/maskcam-beta
- 17. © 2021 BDTI and Tryolabs Available resources 17 Us! bdti.com tryolabs.com Independent report bdti.com/maskcam GitHub repository github.com/bdtinc/maskcam Come see our demos and talk to us live!
- 18. © 2021 BDTI and Tryolabs 18 Questions? Email us: maskcam@bdti.com