Computer Vision: Real-time Object Detection & Classification with Deep Learning on Raspberry Pi

AISOMA Case Study:

Real-time object detection and classification with DeepLearning on the Raspberry Pi 3 B+ without connecting to any external web/cloud services

Our approach:

1. Object detection
HoughCircle Detection (OpenCV)

2. Object classification
A specially optimized deep learning network is used, which is also performant on computers with limited resources without GPU and can nevertheless achieve relatively high accuracies. Only a small training and test data set are available for the eight different euro coins: 1281 photos for training, 707 for testing. Since the training data set is minimal, so-called transfer learning is used. A pre-trained deep learning network is used, which was trained on the ImageNet training dataset (approx. 1.2 million images from 1000 categories). To make the classification more robust against rotations, brightness, contrast etc., the training images were additionally rotated randomly, and brightness and contrast were changed (data augmentation).

Achieved validation accuracy: 93.3%*
Test accuracy achieved: 93.4%*
Runtime for classification: 50ms
*strongly dependent on lighting conditions and camera settings

More info:

Please follow our channel and do not miss any trends and innovative developments around edge computing, computer vision and AI based solutions:

You can also follow us on

[aisomaag]

Let us know what video you'd like to see next in the area of AI technologies!

[ahmadasghar]

Hi,
You guys did a great job. Furthermore, I have read so many articles and came to the conclusion that the fastest you can achieve using Raspberry Pi is 1FPS from the live video feed. I would like to request you to guide us through the technique/approach you used to achieve this speed and accuracy. And it would be wonderfully great if you share the project with us. Thanks :)

[FunManHD]

Which SaaS/cloud service do you use? Azure/Cognitive Services from Microsoft?

[reo]

Hi! I new in this area.
I want know more this features in production line. Detecting small part of electronic board “in movement”
Detecting small chips, number, condensers, etc.
if this is broken,
Camera Distance of about 10 cm.

[Nunocesarsa]

Hey can you give me the details on the camera? i would like to loook into using this for nature stuff and not just industrial settings