Edge Device Image Classification in the Field

Development of stand-alone edge devices with in-field animal detection and classification in key regions for real-time data collection and analysis

Table of Contents

Project Overview

To extend the analysis of ecosystems such as the Kalahari, a self-sustaining edge device is being developed, which is to be installed in key regions. The first step is to use motion sensors to detect movements and changes in the environment and an infrared camera to record a specific scene. Subsequently, the recordings will be evaluated by an AI algorithm, animals and features will be recognized and the information will be transmitted to conservationists in Kuzikus.

As a long-term goal, the edge device will support the monitoring of discontinuities in animal behavior that can’t be captured continuously. Furthermore, the migration behavior of animals in the wildlife reserve is not sufficiently known since a high scanning frequency of certain areas is currently not given. In addition, the setup is designed to prevent animals from escaping from protected areas and endangering themselves outside the boundaries by other wildlife, poachers, or commercial farmers which are sensitive to wildlife destroying their farmland.

Hardware and Software Development

The microcontroller used in the first experiments is one of the latest Raspberry Pis, the 4B. Besides a 4GB RAM (up to 8GB commercially available), this Raspberry Pi model features Gigabit Ethernet, along with onboard wireless networking and Bluetooth. The fanless, energy-efficient Raspberry Pi runs silently and uses far less power than other computers.

To provide power to the Edge Device when it is deployed in the field, a PiJuice HAT (Hardware Attached on Top), which includes a 1820 mAh battery from the factory, is tested as the battery management system. To get a first impression about the runtime of the PiJuice HAT with a fully charged battery and the Raspberry Pi in idle state, the battery metrics of the standard 1820 mAh battery have been extracted and visualized. The unfiltered metrics over time are shown in the following graph and include the charge status, the temperature of the battery as well as the voltages and currents of the battery, and the GPIOs (General Purpose Inputs/Outputs).

The black curve in the graph, which represents the battery’s state of charge in percent, shows that a charging process from 5 % to 95 % with the provided power adapter takes about 3 hours, whereas the battery is discharged after 1,5 hours. With an emphasis on the fact that the measurements of the metrics were carried out in the idle state and that the microcontroller is to run autonomously with other components later in the field, it becomes apparent that this runtime is much too short to be able to monitor an area over a sufficient period of time. Therefore, the next step is to test a 12000 mAh lithium-ion battery, which promises a runtime of up to 13 h and can be tested with the PiJuice HAT without any problems.
In order to ensure a purely self-sufficient operation without having to replace the battery continuously, a solar panel is connected to the Raspberry to sufficiently charge the battery during the day. Accordingly, the battery must be designed in such a way that, taking into account a safety factor and an average time of at least 9 hours of sunshine a day, the system is continuously supplied with power for one night and at least one cloudy day.
Furthermore, some components require more energy at certain times. For example, the infrared camera to be used requires more energy at night to supply the infrared LEDs with power in the dark. For a better overview of the components, the Raspberry Pi, the PiJuice, the lithium-ion battery and the camera are shown in the following picture.

In the picture on the right, these components are already installed in a prototypical box to simplify transport and to put less mechanical stress on the battery while testing.

On the software side, a PIR sensor (Pyroelectric Infrared Sensor) has already been tested in order to be able to detect simple movements and thus trigger the camera. Continuous operation of the camera should be avoided in order to extend the battery lifetime. Furthermore, methods were tested to transmit the field data to a receiver. The following image shows possible notification options, which can be customized for different scenarios like a detection event or a critical battery state.

What’s next?

Due to the bad weather conditions in the last weeks in Aachen, only a few data for a suitable solar panel could be obtained so far, so this has to be dealt with in the future when the weather is better. Based on this data, the design of the battery has to be updated considering all consumers at the edge device. Furthermore, one of the biggest tasks remains to develop a classification algorithm that allows different animal species to be reliably detected in front of the camera. In addition, it must be decided how the transfer of the data can be ensured so that they can be evaluated in the conservation headquarters and remain protected from poachers.


The Edge Device will be used to collect information about animal behavior in addition to evaluating the Kalahari using drone data in the field. Because the animals feel unobserved and are not disturbed by a manual measurement by humans, we expect that the animals will behave more naturally and thus the data can be used to better understand different species and their interaction with each other. Another goal is to use low-cost components to install multiple modules in different locations to advance the digitalization of the wildlife resort. The resulting network of data sources can also be used to analyze more complex relationships such as herd migrations.


Christoph Spurk
Katrin Koppe

Do you have any comments,  ideas or hints for this post?

Then contact us!

This website uses cookies to ensure you get the best experience on our website.