Human Detection on RaspberryPi

Human Detection with a RaspberryPi 3 Model B+ with a SSD MobileNet V2 and a caffeemodel

115

Things used in this project

Hardware components

Raspberry Pi 3 Model B

Raspberry Pi Camera Module V2

Software apps and online services

Raspberry Pi Raspbian

OpenCV – Open Source Computer Vision Library OpenCV

Story

The goal of the project is to develop artificial intelligence that is able to identify humans to turn on the light using a raspberry pi with a night vision camera.

The Problem

If you have motion sensor in the front entrance or in the back yard, like us you will understand if we say we are annoyed that the light always turns on for everything that moves in front of it. Especially for these annoying trash pandas (raccoon). With a normal motion sensor it is impossible to ignore their movements.

Turning on the light in your front yard without a light switch is already possible with a motion sensor. However, the motion detector reacts to everything in its line of sight that emits a heat field or is moving depending on the sensor itself.

The Idea

To solve that issue, we wanted to use artificial network running an object detection on a raspberry pi with a pi camera attached to it. This will ensure we only detect people in our line of sight. The application should be small and convenient — just align the camera in your needed area that you want be illuminated.

Intro

This project is created in the context of the lecture “Applied Artificial Intelligence” at the University of applied science Esslingen. It covers the invention of an artificial intelligence and Application from scratch solving a real-world problem.

Our goal was to reach at least 10 FPS while reliably detecting humans in our field of view. Our thought process was the higher the fps the better we can track humans. Let's just be honest, more fps are just more awesome!

The final network

Our tests and research on the internet confirmed SSD is the way to go. It combines moderate accuracy with good speed while requiring low computing power. A perfect fit for our hardware. We were able to boost the computing power of the pi a lot by implementing the Intel compute stick.

Performance

In the best case scenario we reach 30 FPS and detect a human right in the second he crosses the camera field of view. Which means less than 50 degrees Celsius core temperature. In the worst case the pi reaches about 70 degrees Celsius and reduces its core speed resulting an approximate frame rate of 20.