Optimized Approaches for Object Detections on Raspberry Pi

Posted on March 3, 2019

You might have heard of Raspberry Pi, it’s a perfect mini computing machine that comes with decent hardware, a portable camera and can be used to for image processing and object detections in your IoT projects. You can even use your Raspberry Pi to train detections systems with your own datasets. In this post, we’ll look at different frameworks for setting up object detections on images and videos.

Raspberry Pi 3B which I’ll be using for this post comes with 1.2 GHz quad-core processor and 1 Gigabyte of RAM. If you have worked with detection systems like Darknet before you’ll be familiar that running detections on a GPU gives us a significant speed boost. However, Raspberry Pi does not contain a powerful enough GPU that can run detection algorithms in a short time. To overcome this constraint we’ll look at two methods which can take advantage of Pi’s multicore processor and we’ll be able to get pretty instant results.

Before we begin to discuss two optimized methods make sure you have the latest version of Raspbian OS(Lite or Full Desktop version) and the packages in your Raspbian installation are updated. If you aren’t sure you can run the following commands in the terminal to get the latest packages.

sudo apt-get update
sudo apt-get dist-upgrade

First Approach: Darknet-NNPACK

Darknet uses YOLO, a super fast classifier which works better than other detections networks. For Raspberry Pi, we’ll be using an optimized version of the original Darknet which is Darknet-NNPACK. There are multiple NNPACK optimized darknet repos on GitHub. We’ll be using the most tested and stable one.


Build Procedure


  • Step 1 – Install or Update Python-pip
      •  sudo apt-get install python-pip

Step 2 – Install PeachyPy and Confu

  • Step 3 – Install and configure Ninja
      • git clone https://github.com/ninja-build/ninja.git
      • cd ninja
      • git checkout release
      • ./configure.py --bootstrap
      • export NINJA_PATH=$PWD
      • Note: Your Pi might hang in the configure ninja step.
        If that happens you can stop the configuring process and start over.
        If the problem still persists I recommend increasing the swap space
        of your Raspberry Pi during the build process.
  • Step 4 – Build NNPACK for Darknet
      • cd
      • git clone https://github.com/digitalbrain79/NNPACK-darknet.git
      • cd NNPACK-darknet
      • confu setup
      • python ./configure.py --backend auto
      • $NINJA_PATH/ninja
      • sudo cp -a lib/* /usr/lib/
      • sudo cp include/nnpack.h /usr/include/
      • sudo cp deps/pthreadpool/include/pthreadpool.h /usr/include/

  • Step 5 – Build the Darknet
      • cd
      • git clone https://github.com/digitalbrain79/darknet-nnpack.git
      • cd darknet-nnpack
      • Make
  • Final Step – Download Weights and Test
  • To test your installation you first need to download YOLO weights file. YoloV2 and YoloV3-tiny work well on the Raspberry Pi. Run the following commands in terminal

If everything went well according to the build instructions above and the process didn’t run into any errors obviously! You’ll see detection results from the dog.jpg image in your terminal. As far as the detection time is concerned my device achieved a detection time of 2 seconds per image with YoloV3 tiny.


Second Approach: Optimized OpenCV

OpenCV is an open-source computer vision library written in C++ and it has bindings available in other languages for example Python. In the second approach, we’ll be using optimizing OpenCV for Raspberry Pi. Raspberry Pi uses an ARM processor which provides NEON and VFP as a part of it’s processing core. We’ll explicitly build OpenCV with flags for NEON and VFP to get the most of Pi’s processor.

Build Procedure

  • Step 1 – Download Dependencies for Optimized OpenCV
      • sudo apt-get install build-essential cmake pkg-config
      • sudo apt-get install libjpeg-dev libtiff5-dev libjasper-dev libpng12-dev
      • sudo apt-get install libavcodec-dev libavformat-dev libswscale-dev libv4l-dev
      • sudo apt-get install libxvidcore-dev libx264-dev
      • sudo apt-get install libgtk2.0-dev libgtk-3-dev
      • sudo apt-get install libcanberra-gtk*
      • sudo apt-get install libatlas-base-dev gfortran
      • sudo apt-get install python2.7-dev python3-dev
      • sudo apt-get install python3-pip
  • Step 2 – Download OpenCV source

  • Step 3 – Download NumPy
      • pip3 install numpy or pip install numpy(for python 2.7)
  • Step 4 – Build OpenCV
    • cd ~/opencv401/
    • mkdir build
    • cd build


-D OPENCV_EXTRA_MODULES_PATH=~/opencv_contrib-4.0.1/modules \






  • Step 5 – Compile OpenCV
      • Before this step make sure that you have at least 7-8 GB free space available on your SD card. You can also use an external HDD or USB pen drive, search for instructions on how to mount an external drive in Raspberry Pi.
      • Compiling OpenCV with all cores can take about 3 hours in total so you have to be patient.
      • If you run into virtual memory exhaust errors try to increase the swap space of Pi up to 1024 MB. Though, remember to bring it back to 100MB after completing the build.
      • Multicore compiling might not work so if the compiling process fails, again and again, you can try compiling it with one core. Compiling this way will take about 8 to 10 hours.
      • Keeping all the above instructions in mind you can start building opencv with  “make -j4” or “make” this is for single core build
  • Step 6 – Installing the compiled OpenCV
    • sudo make install
    • sudo ldconfig

After a successful build, you can test the installation with a simple python script, write import cv2 on the top of the script to verify installation. In my tests OpenCV performed better than Darknet-NNPACK, detections took 1.5 seconds per image. 


Phew! That was a long post. Anyway, there you have it two optimized ways of running object detections in Raspberry Pi. Both methods have their pros and cons. Darknet-NNPACK is easier to set up but it’s kind of rigid when you want to call it’s detect method in your python scripts, you might have to write a wrapper yourself to call darknet commands from a script. OpenCV on the other side tests your patience while setting up but it’s a lot easier to use with bindings such as python.

By Osama Asif

Osama is Enou's IOT Development Lead. He writes about new breakthroughs in the industry. He's passionate about the technologies of the future and loves Garlic Mayo Fries.

Enou is a design centered agency based in Lahore. Our mission is “To reinforce businesses and start-ups by creating value in their work, through our digital services” and the driving force that keeps us motivated is our desire for perfection. We truly believe in the transformative power of illustration and design and their ability to simplify communications, elevate experiences, engage and inspire people because in the end, what they see is what they believe!

Discuss your next big idea with us!

Got something you want us to look after? Or ready to shape up your thoughts? Speak up directly or drop us an email, we would love to reach you shortly.


  • Hotels
  • Hospitality
  • Startups
  • B2B Tech
  • Life Style

©2019 ENOU.       Terms & Conditions       Privacy Policy

Share This