Edge computing in practice using IoT, Tensorflow and Google Cloud
2 likes895 views
This document discusses building an edge computing project using IoT cameras, TensorFlow, and Google Cloud. The project uses inexpensive WiFi cameras to capture images locally and process them on a Raspberry Pi gateway using TensorFlow object detection models. The processed data is then sent to Google Cloud for storage and remote access. Custom models can also be trained to classify objects like cats in the images. The project demonstrates how to leverage both local and cloud resources for scalable and flexible edge computing applications.
Ad
More Related Content
What's hot (20)
Similar to Edge computing in practice using IoT, Tensorflow and Google Cloud (20)
Ad
More from Alvaro Viebrantz (20)
Ad
Recently uploaded (20)
Edge computing in practice using IoT, Tensorflow and Google Cloud
- 1. EDGE COMPUTING IN PRACTICE USING IOT, TENSORFLOW AND GOOGLE CLOUD Alvaro Viebrantz Google Developer Expert for IoT and Product Engineer at Leverege aviebrantz.com @alvaroviebrantz
- 2. What will we cover ? What is Edge Computing ? Overview of a practical scenario. Show how to build it!
- 3. 3
- 4. Availability Scalability Maintenance Cost Savings 4
- 5. BUT WHAT IS EDGE COMPUTING ?
- 6. WHEN DOES IT MAKES SENSE ?
- 7. Privacy
- 9. Latency
- 10. Reliability
- 11. Costs
- 12. TALK IS CHEAP LET’S SEE IN PRACTICE
- 13. USING CAMERAS AS SENSORS
- 15. 15 Camera as sensors Our Edge Computing project • Use cheap Wifi Cameras • Local server receiving and processing images • Object detection using Machine Learning • Send processed data to the cloud • Provide a local and remote web interface to show the data
- 17. HARDWARE
- 18. 18 ESP32 Cam Raspberry Pi 3 10$ 35$
- 19. PROVIDE IMAGES OVER THE LOCAL NETWORK
- 20. 20 Camera Firmware github.com/alvarowolfx/gcloud-iot-edge-tensorflow/tree/master/esp32-camera-firmware • Connect to Wifi Network • Provide HTTP endpoint to acquire images • Read image from local camera (OV2640)
- 21. 21 Camera Discovery with mDNS Local Discovery Protocol • Find local cameras automatically on the local network {INSTANCE_NAME}.local/jpg • Service _camera and protocol _tcp
- 22. 22 Finding local cameras Testing it by searching locally using dns-sd cli • http://indoor-camera-ec5d.local/jpg • http://indoor-camera-60d8.local/jpg
- 23. 23
- 25. FETCH IMAGES, CLASSIFY AND SEND TO THE CLOUD
- 26. 26 ANY LANGUAGE CAN BE USED HERE
- 27. 27
- 28. DISCLAIMER I’M NOT A MACHINE LEARNING EXPERT
- 29. BUT MAYBE THIS IS BEST THING ABOUT THIS WHOLE PROJECT
- 30. tensorflow.org/js/models Pre-trained TensorFlow.js models that can be used in any project out of the box.
- 31. Classification Mobilenet model Object Detection CocoSSD model Cat, Dog, DuckCat Classification vs Object Detection
- 32. 32
- 35. DEMO - EDGE SERVER
- 36. !36
- 37. !37
- 38. !38 My Computer Raspberry Pi 3 TensorFlow.js Core 8 seconds per frame 45 seconds per frame TensorFlow.js Node 200 milliseconds per frame 1 second per frame Performance Use tfjs-node and tfjs-node-gpu if possible
- 39. CLOUD SETUP
- 45. INGESTING DATA
- 48. 48
- 49. STORING DATA
- 51. 51
- 52. 52 firebase deploy 🔥 🔥 🔥
- 53. !53
- 54. !54
- 55. !55
- 56. Article describing how to build this project bit.ly/gcloud-iot-edge
- 58. I love cats What if I could track my cats at home ?
- 59. Train custom object detection model First approach
- 60. Transfer Learning Using pre-trained models to train a custom model with smaller dataset
- 61. Building the dataset Usually the hardest part on any ML project
- 62. LOW ACCURACY ON MY MODEL 😭😿
- 63. Train a classification model Using the same dataset
- 64. Train a classification model Using the same dataset
- 65. Best of both worlds Use the pre-trained object detection model first them use the my custom classifier Detect Objects using CocoSSD Filter for cats and crop boxes Classify using custom model.
- 66. DEMO - CUSTOM MODEL
- 68. !68 Found a cat Found Jam
- 69. !69 Found a cat Found Berry
- 70. !70 Found two catsFound Jam Found Muffin Found Berry
- 71. FUTURE WORK
- 72. Edge Devices Better performance on the edge
- 73. Other Edge Devices Better performance on the edge Sipeed Maix Bit Sparkfun Edge
- 74. Face recognition Identify allowed person on a given place github.com/ageitgey/face_recognition
- 75. Pose estimation Posenet pre-trained model
- 76. WRAPPING UP
- 77. Wrapping up !77 Use device power on the edge Extract more info using Machine Learning Scalability, Flexibility and Ease of Usage Using both local and cloud computing
- 78. THANK YOU! AVIEBRANTZ.COM Alvaro Viebrantz Google Developer Expert for IoT and Product Engineer at Leverege aviebrantz.com @alvaroviebrantz