Video analysis for the detection of animals using convolutional neural networks and consumer-grade drones
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
Determining animal distribution and density is important in conservation. The process is both time-consuming and labour-intensive. Drones have been used to help mitigate human-intensive tasks by covering large geographical areas over a much shorter timescale. In this paper we investigate this idea further using a proof of concept to detect rhinos and cars from drone footage. The proof of concept utilises off-the-shelf technology and consumer-grade drone hardware. The study demonstrates the feasibility of using machine learning (ML) to automate routine conservation tasks, such as animal detection and tracking. The prototype has been developed using a DJI Mavic Pro 2 and tested over a global system for mobile communications (GSM) network. The Faster-RCNN Resnet 101 architecture is used for transfer learning. Inference is performed with a frame sampling technique to address the required trade-off between precision, processing speed, and live video feed synchronisation. Inference models are hosted on a web platform and video streams from the drone (using OcuSync) are transmitted to a real-time messaging protocol (RTMP) server for subsequent classification. During training, the best model achieves a mean average precision (mAP) of 0.83 intersection over union (@IOU) 0.50 and 0.69 @IOU 0.75, respectively. On testing the system in Knowsley Safari our prototype was able to achieve the following: sensitivity (Sen), 0.91 (0.869, 0.94); specificity (Spec), 0.78 (0.74, 0.82); and an accuracy (ACC), 0.84 (0.81, 0.87) when detecting rhinos, and Sen, 1.00 (1.00, 1.00); Spec, 1.00 (1.00, 1.00); and an ACC, 1.00 (1.00, 1.00) when detecting cars.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.002 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it