XTrail-ID: An Explainable AI Human Footprint Trail Identification on Soil Substrate Using Unsupervised Machine Learning from UAV Imagery

1. What they did

Human tracks are valuable evidence in forensics, anti-poaching, military tracking and search and rescue, but most computer methods compare single tracks under controlled conditions. The authors developed XTrail-ID, a method that sorts barefoot tracks photographed from a drone into trails belonging to each person, without needing labelled training data.

Twenty adults walked barefoot, side by side or in single file, across compact and loose soft soil in Botswana, and 1,000 images were taken from drone video. The true trails were known from data collection and checked by an expert military tracker of the Botswana Defence Force. Tracks were located with an adapted footprint identification model, using either tilted boxes around each track or outlines of the track itself, then described by a pretrained vision model. Tracks were grouped by similarity, and heatmaps showed which parts of each track drove the grouping. Two experienced trackers checked the groupings and heatmaps.

2. What they found

• Using track outlines with cosine similarity gave the best results: 94.2% accuracy, against 91.5% using tilted boxes.
• Cosine similarity beat Pearson and Spearman correlations significantly for both methods.
• The outline method grouped tracks correctly for two to five people, including where grass partly hid the tracks and where people walked in single file.
• The box method worked for two or three people but became inconsistent with four or five, and overestimated the number of people when tracks overlapped on a shared path.
• The heatmaps highlighted parts of the track such as the heel, arch, outer edge and toes, which the authors interpret as differences in each person’s gait.

3. What they concluded

An unsupervised method using track outlines and cosine similarity can group barefoot tracks into individual trails from drone images, and heatmaps make the basis of each grouping visible, which matters in forensic work. The authors call for testing on more people, more soil types and harsher conditions, since there is no standard dataset yet.

4. How tracking was or was not used

Tracks are the subject. The method identifies which tracks belong to the same person and estimates how many people were present. An expert military tracker verified the true trails, and two experienced trackers independently examined the grouped trails, using track shape, trail continuity and the heatmaps, and explained why some tracks were grouped correctly or wrongly. No trailing in the sense of following a track was involved.

5. Tracker or observer certification

Reported but not validated. The ground truth was verified by an expert tracker specializing in military tracking in the Botswana Defence Force, and two experienced trackers evaluated the results, but no certification or test of their skill is reported.

6. Other technologies, skills or innovations

• Drone video (Inspired Flight IF1200A with a Sony Alpha 7R IV camera) at 8 m height.
• An adapted DeepFIT (Deep Learning Footprint Identification Technology) model, previously developed by the first author’s group for tracks on soil, used to locate each track.
• A self-supervised vision transformer (DINOv2) and a new Embedding Spatial Attribution Module to explain results with heatmaps.