Unmanned Aerial Vehicles (UAVs) and Artificial Intelligence Revolutionizing Wildlife Monitoring and Conservation

Felipe Gonzalez, Glen A. Montes, Eduard Puig, Sandra Johnson, Kerrie L. Mengersen, Kevin J. Gaston

TL;DR

This paper presents a system combining Unmanned Aerial Vehicles (UAVs) with thermal imaging and artificial intelligence to automate wildlife detection and counting. The system aims to overcome the resource-intensive and often imprecise nature of traditional wildlife monitoring methods, offering a more efficient and accurate approach for conservation efforts.

What Problem Does This Paper Address?

Accurately estimating wildlife populations is crucial for effective conservation and management, especially for threatened or invasive species. Traditional methods like camera traps, ground surveys, and GPS collaring are labor-intensive, costly, and can be imprecise, particularly when dealing with species that have large ranges, low population densities, elusive behavior, or are sensitive to disturbance. The challenge lies in efficiently surveying large areas and processing the vast amounts of data collected. While UAVs offer a promising aerial platform, the bottleneck remains in automating the analysis of the collected imagery to detect and count wildlife, which is often a manual and time-consuming process.

Methods

The researchers developed a system composed of two main segments: an airborne segment with a multirotor UAV equipped with a thermal camera and a gimbal system, and a ground segment with a laptop-based ground station for data processing.

The core innovation lies in the two algorithms developed for automated wildlife detection and classification:

  • Pixel Intensity Threshold (PIT) Algorithm: This algorithm leverages the heat signature of wildlife, which creates a contrast against the background. It uses image binarization (intensity thresholding) to segment potential objects of interest. Morphological operations (erosion and dilation) are then applied to clean up the image and separate overlapping objects. Finally, it searches for clusters of pixels within a defined size range to count objects. This method is effective for clear thermal signatures but struggles with distinguishing species or tracking single objects.
  • Template Matching Binary Mask (TMBM) Algorithm: This more advanced algorithm addresses the limitations of PIT. It involves:
  • Template Loading: Using pre-defined template images of target species (e.g., koalas).
  • Template Processing: Enhancing template contrast.
  • Template Matching: Searching for matches in video frames using a scoring system (score of 7 or higher for a high-quality match). Matches must persist for at least 10 consecutive frames to avoid false positives.
  • Coordinate Assignment & Mask Creation: Storing coordinates of matches and creating a mask to isolate the object of interest.
  • Logical Operation with Mask: Applying a logical AND to remove background.
  • Pixel Intensity Threshold: Further refining the detection.
  • Tracking: Identifying and distinguishing multiple objects based on their coordinates and ensuring they don't make large jumps or drastically change size between frames.
  • Counting: Numbering and displaying matches, counting consecutive frames an object is present, and tallying objects that leave the scene.

The system utilizes a FLIR Tau 2-640 thermal camera and a Mobius RGB camera for data acquisition.

Datasets

The paper does not explicitly mention training datasets. However, it uses footage collected from flights over a study area on the Sunshine Coast, Queensland, Australia. This footage, including both thermal and RGB imagery, serves as the evaluation dataset. The study area was a large rehabilitation enclosure with a known number of koalas, allowing for ground-truthing of the algorithm's performance.

Key Insights and Ideas

  • Thermal Imaging + AI is a Powerful Combination: The paper strongly demonstrates how combining the thermal signature detection capabilities of UAV-mounted cameras with AI-driven image processing can revolutionize wildlife monitoring. Thermal imaging excels at detecting warm-blooded animals against cooler backgrounds, especially at night or in dense foliage where visual detection is difficult.
  • Template Matching for Species Identification: The TMBM algorithm's use of templates is a crucial step towards species-specific detection, moving beyond simple object detection to classification. The ability to define templates for different species and adjust them for variations in size, shape, and orientation is key.
  • Robustness Through Multi-Frame Analysis: Requiring a match to persist for 10 consecutive frames is a smart way to significantly reduce false positives, a common issue in automated image analysis. This temporal consistency adds a layer of reliability.
  • Altitude Matters for Detection: The study highlights that detection accuracy and time are influenced by altitude. Lower altitudes (20-30m) yielded faster detection and fewer false positives, while higher altitudes (above 60m) presented challenges due to reduced resolution and thermal differentials.
  • UAVs as a "New Angle": The paper notes that even when a species is difficult to locate from a standard viewpoint, the UAV's ability to maneuver and provide different angles (e.g., vertical or lateral views) can significantly improve detection success.

Key Results

  • The TMBM algorithm achieved 100% accuracy in detecting koalas at altitudes of 20 m, 30 m, and 60 m when compared with ground-truth detections.
  • At 20 m and 30 m, the average detection time was approximately 1.5 seconds, with zero false positives after filtering.
  • At 60 m, the average detection time increased to 2.1 seconds, with a small number of false positives that were successfully filtered out by the tracking task.
  • The system successfully generated an orthomosaic thermal image with GPS locations of detected koalas, demonstrating its utility for mapping and population distribution analysis.

Limitations and Future Work

  • Species-Specific Templates: The TMBM algorithm's effectiveness is highly dependent on the quality and variety of templates. Developing comprehensive template libraries for multiple species and accounting for variations in their appearance (size, shape, color, thermal footprint) is an ongoing challenge.
  • Altitude Limitations: While 100% accuracy was achieved at lower altitudes, detection becomes more difficult and prone to false positives at higher altitudes (above 60m). This is attributed to lower resolution and reduced thermal contrast.
  • Distinguishing Species: Differentiating between thermal signatures of different species remains challenging, especially at higher altitudes or with lower camera resolution.
  • Dynamic Templates: The authors suggest implementing dynamic templates that can adapt in real-time to changes in object orientation, position, and size during flight.
  • Integration with Tagging: Combining the system with radio communication for matching thermal imagery with tagged animals is proposed as a future enhancement.
  • Broader Applications: The system's potential for detecting other species, including pests (wild dogs, pigs, dingoes), and for search and rescue missions is highlighted, requiring code and template adjustments.

Summary updated

Abstract

Surveying threatened and invasive species to obtain accurate population estimates is an important but challenging task that requires a considerable investment in time and resources. Estimates using existing ground-based monitoring techniques, such as camera traps and surveys performed on foot, are known to be resource intensive, potentially inaccurate and imprecise, and difficult to validate. Recent developments in unmanned aerial vehicles (UAV), artificial intelligence and miniaturized thermal imaging systems represent a new opportunity for wildlife experts to inexpensively survey relatively large areas. The system presented in this paper includes thermal image acquisition as well as a video processing pipeline to perform object detection, classification and tracking of wildlife in forest or open areas. The system is tested on thermal video data from ground based and test flight footage, and is found to be able to detect all the target wildlife located in the surveyed area. The system is flexible in that the user can readily define the types of objects to classify and the object characteristics that should be considered during classification.
BibTeX
@inproceedings{gonzalez2016unmanned,
  author = {Felipe Gonzalez and Glen A. Montes and Eduard Puig and Sandra Johnson and Kerrie L. Mengersen and Kevin J. Gaston},
  title = {{Unmanned Aerial Vehicles (UAVs) and Artificial Intelligence Revolutionizing Wildlife Monitoring and Conservation}},
  booktitle = {Sensors},
  year = {2016},
  url = {https://openreview.net/forum?id=GBe3_bes0__},
}

Loading related content...

Figure