Apply Machine Learning to IoT using Tensorflow and Android Things

Discover how to apply Machin Learning to IoT using Android Things and Tensorflow. Learn how to build a custom model based on image classification using Tensorflow.

This project explores how to apply Machine Learning to IoT or if you prefer how to apply Tensorflow to IoT. In more details, this Machine Learning tutorial explores how to integrate Tensorflow with Android Things.

Android Things is now in a stable version named Android Things 1.x and it is ready to be used in the production systems. As you may know already Raspberry Pi is one of the supported platforms for development and prototyping with Android Things 1.x. This tutorial uses Android Things 1.x with Raspberry Pi. Anyway, you can easily move to other supported platforms without changing the code. During this tutorial that covers how to apply Machine Learning to IoT,  you will learn step by steps all the information you need to build a smart Android Things system that uses Tensorflow image classification.

Machine Learning with Android ThingsMachine Learning is, with the Internet of Things, one of the most interesting technological topics.  To give a simple definition of the Machine Learning, it is possible to the Wikipedia definition:

Machine learning is a field of computer science that gives computer systems the ability to “learn” (i.e. progressively improve performance on a specific task) with data, without being explicitly programmed.

In other words, after a training step, a system can predict outcomes even if it is not specifically programmed for them. On the other hands, we all know IoT and the concept of connected devices. One of the most promising topics is how to apply Machine Learning to IoT, building expert systems. Using Tensorflow with Android Things,  it is possible to develop a system that is able to “learn”. Moreover, this system uses this knowledge to control and manage physical objects. Before digging into the details of building a custom model with Tensorflow and use it in Android Things, you should have installed it on your device. If this is the first time you use Android Things, you can read this tutorial about how to install Android Things on your device.

There are several fields where applying Machine Learning to IoT produces  an important value, just to mention a few interesting fields, there are:

  • Industrial IoT (IIoT) in the predictive maintenance
  • Consumer IoT where the Machine earning can make the device intelligent so that it can adapt to our habits

The basic idea that stands behind this Android Things IoT project is exploring how to build a robot car that is able to recognize some basic images (like arrows) that can control in this way the robot car directions. We have already covered how to build robot car using Android Things, so I suggest you read the tutorial before starting this project.

Let us start!

Apply Machine Learning to IoT project steps

To apply Machine Learning to IoT, or in other words to integrate Android Things with Tensorflow, we have to follow these steps:

  • How to set up the TensorFlow environment using Docker
  • How to train the TensorFlow system building a custom model (image classification)
  • How to integrate TensorFlow with Android Things
  • How to control the robot car using TensorFlow result

This project is derived from Android Things TensorFlow image classifier.

How to use Tensorflow image recognition using a ML custom model

Before starting it is necessary to install and configure the TensorFlow environment. I’m not a Machine Learning expert, so I needed to find something fast and ready to use so that we can build the TensorFlow image classifier. For this reason, we can use Docker to run an image of TensorFlow. Follow these steps:

  1. Clone the TensorFlow repository:[shell] git clone
    cd /tensorflow
    git checkout v1.5.0[/shell]
  2. Create a directory (/tf-data) that will hold all the files that we will use during the project and they will be our dataset to train the system
  3. Run Docker:[shell]docker run -it \
    –volume /tf-data:/tf-data \
    –volume /tensorflow:/tensorflow \
    –workdir /tensorflow tensorflow/tensorflow:1.5.0 bash[/shell]

    Using this command, we run an interactive TensorFlow environment and we mount some directories that we will use during the project

How to Train TensorFlow to recognize images building a custom model

Before the Android Things system is able to recognize images, it is necessary to train the TensorFlow engine and building its model. For this purpose, it is necessary to gather several images that will be our dataset. As said before, we want to use arrows to control the Android Things robot car so that we have to collect at least four arrow types:

  • up arrow
  • down arrow
  • left arrow
  • right arrow

To train the system is necessary to create a “knowledge base” with these four different image categories. Create in /tf-data a directory called images and under it four sub-directories named:

  • up-arrow
  • down-arrow
  • left-arrow
  • right-arrow

Now it is time to look for the images. I have used Google Image search but you can use other approaches too. To simplify the image download process, you should install a Chrome plugin that downloads all the images with only one click. Do not forget more images you download better is the training process, even if the time to create the model could increase.

Open your browser and start looking for the four image categories:

TensorFlow image classifier

I have downloaded 80 images for each category.  Do not care about image extension.

Once all the categories have their images follow these steps (in the Docker interface):

[shell]python /tensorflow/examples/image_retraining/ \
–bottleneck_dir=tf_files/bottlenecks \
–how_many_training_steps=4000 \
–output_graph=/tf-data/retrained_graph.pb \
–output_labels=/tf-data/retrained_labels.txt \

It could take some time so be patient. At the end, you should have two files in /tf-data folder:

  1. retrained_graph.pb
  2. retrained_labels.txt

The first file contains our model as the result of the TensorFlow training process while the second file contains the labels related to our four image categories.

How to test the Tensorflow custom model

If you want to test the model to check if everything is working you can use this command:

[shell]python scripts.label_image \
–graph=/tf-data/retrained-graph.pb \

Optimizing the model

Before we can use this TensorFlow custom model in the Android Things project it is necessary to optimize it:

[shell]python /tensorflow/python/tools/ \
–input=/tf-data/retrained_graph.pb \
–output=/tf-data/opt_graph.pb \
–input_names=”Mul” \

That’s all. We have our Tensorflow custom model. We will use this model to apply Machine Learning to IoT or in more details to integrate Android Things with TensorFlow. The goal is applying to the Android Things app the intelligence to recognize arrow images and react consequently controlling the robot car directions.

If you want to have more details about TensorFlow and how to generate the model look at the official documentation and to this tutorial.

How to integrate Android Things and TensorFlow

Once the TensorFlow data model is ready, we can move to the next step: how to integrate Android Things with TensorFlow.  To this purpose, we can split this task into two steps:

  1. The hardware part, where we connect motors and other peripherals to the Android Things board
  2. Implementing the app

Android Things Schematics

Before digging into the details about how to connect peripherals, this is the list of components used in this Android Things project:

  1. Android Things board (Raspberry Pi 3)
  2. Raspberry Pi Camera
  3. One LED
  4. LN298N Dual H Bridge (to control the motors)
  5. A robot car chassis with two wheels

I do not cover again how to control motors using Android Things because we have already covered in the previous post.

Below the schematics:

Integrating Android Things with IoT

In the picture above, the camera is not shown. The final result is:

Integrating Android Things with TensorFlow

Implementing the Android Things app with TensorFlow

The last step is implementing the Android Things app. To this purpose, we can re-use the example available in Github named sample TensorFlow image classifier. Before starting, clone the Github repository so that you can modify the source code.

This Android Things app is different from the original app because:

  1. it does not use the button to start the camera to capture the image
  2. It uses a different model
  3. It uses a blinking led to notify that the camera will take the picture after the LED stops blinking
  4. It controls the motors when TensorFlow detects an image (arrows). Moreover, it turns on the motors for 5 seconds before starting the loop from step 3

To handle a blinking LED, use the following code:

[java]private Handler blinkingHandler = new Handler();
private Runnable blinkingLED = new Runnable() {
public void run() {
try {
// If the motor is running the app does not start the cam
if (mc.getStatus())
return ;

Log.d(TAG, “Blinking..”);
if (currentValue <= NUM_OF_TIMES) {
else {
currentValue = 0;;
} catch (IOException e) {

When the LED stops blinking, the app captures the image.

Now it is necessary to focus on how to control the motors according to the image detected. Modify the method:

public void onImageAvailable(ImageReader reader) {
final Bitmap bitmap;
try (Image image = reader.acquireNextImage()) {
bitmap = mImagePreprocessor.preprocessImage(image);

final List<Classifier.Recognition> results =

“Got the following results from Tensorflow: ” + results);

// Check the result
if (results == null || results.size() == 0) {
Log.d(TAG, “No command..”);;
return ;

Classifier.Recognition rec = results.get(0);
Float confidence = rec.getConfidence();
Log.d(TAG, “Confidence ” + confidence.floatValue());

if (confidence.floatValue() < 0.55) {
Log.d(TAG, “Confidence too low..”);;
return ;

String command = rec.getTitle();
Log.d(TAG, “Command: ” + rec.getTitle());

if (command.indexOf(“down”) != -1)
else if (command.indexOf(“up”) != -1)
else if (command.indexOf(“left”) != -1)
else if (command.indexOf(“right”) != -1)

In this method, after the TensorFlow returns the possible labels matching the image captured, the app compares the result with the possible directions and controls the motors consequently.

Finally, it is time to use the model created at the beginning. Copy the opt_graph.pb and the reatrained_labels.txt under the assets folder replacing the existing files.

Open the and modify the following lines:

[java]public static final int IMAGE_SIZE = 299;
private static final int IMAGE_MEAN = 128;
private static final float IMAGE_STD = 128;
private static final String LABELS_FILE = “retrained_labels.txt”;
public static final String MODEL_FILE = “file:///android_asset/opt_graph.pb”;
public static final String INPUT_NAME = “Mul”;
public static final String OUTPUT_OPERATION = “output”;
public static final String OUTPUT_NAME = “final_result”;

Run the app and have fun showing arrows to the camera and check the result. The robot car has to move according to the arrow shown.


At the end of this tutorial, we have discovered how to apply Machine Learning to IoT using Android Things and TensorFlow. We can control the robot car using images and make it moving according to the image shown.