Projects:2021s1-13252 Feral Animal Recognition Using Thermal and Depth Sensing

From Projects
Jump to: navigation, search
Top View of hardware, including thermal and depth camera

Abstract

A passive detection system to record feral animals was developed. A hardware demonstrator was developed to be deployed outside during nighttime to record the presence of any animal that moves in front of it. The system uses a thermal and depth camera to monitor for changes in temperature and distance and store these images in a dataset.

Introduction

Each year feral animals cause significant harm in Australia, killing both native wildlife and damaging the environment. Common pest species in South Australia include feral cats, rabbits, and foxes, but also larger species such as camels and buffaloes cause significant damage in areas of northern Australia [1]. This research project is an extension of work conducted last year aimed at developing an automated feral animal detection system that could be deployed remotely to monitor for the presence of pest species. This system was based on a Raspberry PI board and made use of optical, thermal (long-wave infrared) and depth sensors to detect the presence of an animal. Machine learning techniques were then used to attempt to distinguish a few simple types (eg. dogs, cats humans etc). The aim of this follow-on project is to: Build upon last year's project to improve detection capabilities and build a dataset for use in training a classification model Develop a hardware demonstrator which can be deployed remotely overnight

Project team

Project students

  • Benjamin Weichert
  • Daniel Rohling

Supervisors

  • Dr Danny Gibbins
  • Dr Said Al-Sarawai

Objectives

1. Improve the detection capability in a way that minimises power consumption but allows it to operate more effectively between dusk and dawn in the absence of a light source. 2. Examine techniques in 3D sensing and machine learning that could be used to improve recognition. 3. Collect data and develop a hardware demonstrator that could be deployed in the field overnight.

Method

Software

First, frames are extracted from each camera where detection programs decide if there is a warm, close subject in the image. The thermal and depth images containing the subject are mapped together and stored. Processing and inference programs then determine what, if any, animal is present.

Diagram of software design

Hardware

A weatherproof box encloses the system so it can operate outside for long time periods. Depth and thermal cameras are connected to a Raspberry Pi which runs the software. The system can be powered using either an external 12V battery or a power adapter.

Diagram of hardware design

Results

The images demonstrate the operation of the hardware demonstrator. The thermal camera images the dog using its temperature and the depth camera images the dog using its proximity.

Image of dog in thermal camera
Image of dog in depth camera












When deployed, the system can passively detect the presence of a warm subject, as shown in the graph to the right, by a statistically significant increase in the pixel values. If a detection occurs, that pixel is added to the binary detection mask, thereby isolating the warm subject from the background.

Results from the thermal camera, measuring thermal values against time. Clear detections can be seen


A similar approach is then taken with the depth camera where pixels with statistically close depth values are used to form a mask to fix a box around the subject and the real distance from the camera is used to calculate the subject’s height and width in centimetres. The system developed provides a robust, energy-efficient detection system, combining thermal and depth imagery to facilitate nocturnal gathering of data and in turn, develop a machine-learning classification system.

Result of a detection using the depth camera. The dog is clearly detected and a bounding box is fixed around it

Conclusion

In conclusion, the group was able to improve the detection and produce a system that could reliably detect and save data. The thermal and depth camera were successfully utilised and combined to produce a system that could operate outside, over long periods of time, and detect wild animals using both sensors. Data saved can be used by future students to train a classification model and determine the species of animals that are detected by the system.

References

[1] "Feral Animals." Invasive Species Council. https://invasives.org.au/our-work/feral-animals/ (accessed 11/05/2021, 2021).