Deep learning gateway on Raspberry Pi
Clone or download
bafu Fix incorrect import module name.
Signed-off-by: Bofu Chen (bafu) <>
Latest commit a975198 Sep 11, 2018
Failed to load latest commit information.
apt Enable Movidius neural compute stick. Dec 4, 2017
berrynet Fix incorrect import module name. Sep 10, 2018
config Add dashboard cfg w/ MQTT-based image widget. Jul 26, 2018
doc Update document for Nest camera setup. May 30, 2017
inference Merge branch 'feature-engine-selection' into develop Mar 1, 2018
patch Make faster detection backend as systemd service. Sep 23, 2017
systemd Support detection by Movidius. Feb 8, 2018
test Init. Apr 27, 2017
tests Update TensorFlow engine unittest. Feb 17, 2018
utils Merge branch 'paulliu-caffe2-classify' into develop Jan 14, 2018
.eslintrc.yaml Use ESLint as linter. May 8, 2017
.gitignore Update gitignore. Mar 1, 2018
.gitlab-ci.yml Add CI/CD support Sep 4, 2018
.gitmodules Add darkflow as Git submodule. May 20, 2017
AUTHORS Update AUTHORS. Jan 14, 2018 Add (use DCO instead of CLA). Sep 4, 2017
LICENSE.txt Init. Apr 27, 2017 Manually merged Paul's Nest streaming support: Sep 30, 2017
berrynet-manager Remove LINE service from default services. Jul 8, 2018
broker.js Use ESLint as linter. May 8, 2017
camera.js Manually merged Paul's Nest streaming support: Sep 30, 2017
client.js Respect coding style. Sep 19, 2017
config.js Convert data collector's input text data format to JSON. Oct 9, 2017
configure Fix BN2 running issue Jul 8, 2018
data_collector.js Convert data collector's input text data format to JSON. Oct 9, 2017
journal.js Use the result image from object detection instead of camera May 30, 2017
line.js Merge branch 'feature-line-notification' into develop Oct 8, 2017
localimg.js Use ESLint as linter. May 8, 2017
mail.js Respect coding style. Sep 19, 2017
package.json Merge branch 'feature-line-notification' into develop Oct 8, 2017 Bump to v3.2.0 Aug 14, 2018
uninstall Add uninstallation script. Dec 13, 2017

BerryNet: Deep Learning Gateway on Raspberry Pi

This project turns Raspberry Pi 3 into an intelligent gateway with deep learning running on it. No internet connection is required, everything is done locally on the Raspberry Pi 3 itself. At DT42, we believe that bringing deep learning to edge devices is the trend towards the future. It not only saves costs of data transmission and storage but also makes devices able to respond according to the events shown in the images or videos without connecting to the cloud.

Figure 1

Figure 1 shows the software architecture of the project, we use Node.js, MQTT and an AI engine to analyze images or video frames with deep learning. So far, there are two supported AI engines, the classification engine and the object detection engine. Figure 2 shows the differences between classification and object detection.

Figure 2

One of the application of this intelligent gateway is to use the camera to monitor the place you care about. For example, Figure 3 shows the analyzed results from the camera hosted in the DT42 office. The frames were captured by the IP camera and they were submitted into the AI engine. The output from the AI engine will be shown in the dashboard. We are working on the Email and IM notification so you can get a notification when there is a dog coming into the meeting area with the next release.

Figure 3

AI Engines

The current supported AI Engines leverage work from the following projects:

The current supported classification model is Inception v3 [1] and the detection model is TinyYOLO [2]


$ git clone
$ cd BerryNet
$ ./configure

Start and Stop BerryNet

BerryNet is managed by systemd. You can manage BerryNet via berrynet-manager:

$ berrynet-manager [start | stop | status | log]


All the configurations are in config.js.

  • Choose AI Engine.

    • Two types of AI engines currently: object classifier and object detector.
  • Configure IP camera's snapshot access interface.

  • MQTT topics.


Open dashboard on RPi3 (with touch screen)

Open browser and enter the URL:


The default dashboard configuration file will be loaded.

Open dashboard on browser from any computer

Open browser and enter the URL:


Click the data sources, and change MQTT broker's IP address to the gateway's IP.

For more details about dashboard configuration (e.g. how to add widgets), please refer to freeboard project.

Provide Image Input

To capture an image via configured IP camera

$ mosquitto_pub -h localhost -t berrynet/event/camera -m snapshot_ipcam

To capture an image via board-connected camera (RPi camera or USB webcam)

$ mosquitto_pub -h localhost -t berrynet/event/camera -m snapshot_boardcam

To provide a local image

$ mosquitto_pub -h localhost -t berrynet/event/localImage -m <image_path>

To start and stop streaming from board-connected camera

$ mosquitto_pub -h localhost -t berrynet/event/camera -m stream_boardcam_start
$ mosquitto_pub -h localhost -t berrynet/event/camera -m stream_boardcam_stop

To start and stop streaming from Nest IP camera

$ mosquitto_pub -h localhost -t berrynet/event/camera -m stream_nest_ipcam_start
$ mosquitto_pub -h localhost -t berrynet/event/camera -m stream_nest_ipcam_stop

Enable Data Collector

You might want to store the snapshot and inference results for data analysis.

To enable data collector, you can set the storage directory path in config.js:

config.storageDirPath = '<data-storage-dirpath>';

and restart BerryNet.

Use Your Data To Train

The original instruction of retraining YOLOv2 model see github repository of darknet

In the current of BerryNet, TinyYolo is used instead of YOLOv2. The major differences are:

  1. Create file yolo-obj.cfg with the same content as in tiny-yolo.cfg
  2. Download pre-trained weights of darknet reference model, darknet.weights.12, for the convolutional layers (6.1MB)

The rest parts are the same as retraining YOLO.

If you use LabelMe to annotate data, utils/ can help convert the xml format to the text format that darknet uses.


Please refer to the Google Group for questions, suggestions, or any idea discussion.