trACT: temporal revelation Airborne Camera Trap

Abstract: Remote monitoring of free-ranging wildlife and anti-poaching surveillance across vast ecosystems are fundamentally constrained by environmental clutter, dynamic vegetation, camouflage, and the spatial rigidity of ground-based camera traps. Inspired by the hunting strategies of birds of prey, we present trACT (temporal revelation Airborne Camera Trap), an autonomous aerial sensing framework that converts consumer drones into mobile, high-altitude camera traps. By unifying Temporal Max Pooling (TMP) –a low-overhead signal processing method that extracts hidden motion trails across rolling integration windows– with self-supervised Noise-Contrastive Estimation (NCE), trACT reliably distinguishes subtle animal or person movement from background thermal noise and wind-induced foliage motion. To overcome mechanical and computational latency (946ms loop turnaround), the system combines motion prediction with automated gimbal zoom verification and multi-target balancing. Field validation across temperate forest, urban canopy surveillance grounds, and rainforest habitats demonstrates that trACT bridges wide-area aerial scanning with precise target verification, elevating a recall of 83.8%-87.6% when paired with AI vision-reasoning. Temporal Max Pooling Results: https://zenodo.org/records/23079962 Data S1. The dataset contains all trACT verification images for the wildlife monitoring experiments (both before and after outlier removal and visual reasoning-based species classification) along with manually selected and classified ground truth images. The data are organized by flight (date and flight number). A companion Python script for outlier removal and species classification is also provided. Data S2. The dataset contains all trACT verification images for the surveillance experiments (both before and after outlier removal) along with manually curated ground truth images. Because the flights were conducted on a single day, corresponding images are stored in the same folders. A companion Python script for outlier removal is also provided. Note that all faces have been blurred for privacy, which may prevent the reproduction of our outlier removal results. Movie S1. This video illustrates a hawk’s fixation strategy and demonstrates its application to a drone. It first shows how trACT uses motion signals to find and track a herd of red deer through dense forest occlusions and thermal clutter, achieving automatic detection and verification. It then presents technical verification experiments conducted under controlled conditions, focusing on motion anomaly detection, multi-target verification, and partial occlusion. Movie S2. This video presents two examples of manual detection and automatic verification using motion signals from wildlife monitoring field experiments. A roe deer (Capreolus capreolus) was detected in dense, low-tree vegetation, and a red deer stag (Cervus elaphus) was located hiding in a cornfield. Movie S3. This video presents eight examples of manual detection and automatic verification using motion signals from wildlife monitoring field experiments. A red deer stag (Cervus elaphus) and a roe deer (Capreolus capreolus) were detected in dense blackberry bushes; two European hares (Lepus europaeus) were found hiding in tall grass and open fields; two American minks (Neovison vison) chasing each other in an open field; two roe deer hiding in tall grass were captured in both daylight and night-vision modes; and a European badger (Meles meles) was detected under a tree. Movie S4. This video presents three examples of automatic detection and verification for multiple targets using motion signals. The footage features roe deer (Capreolus capreolus) hiding in grass in night-vision mode, and cattle (Bos taurus) in a pasture during daylight. Movie S5. This video presents multiple examples of manual and automatic detection and verification for both unoccluded individuals and people partially occluded by forest vegetation, covering single and multiple targets. Software S1. trACT software for DJI. M30, M400 series for unrestricted gimbal control (tilt and pan). M3, M4 series with limited gimbal control (tilt only). Software S2. trACT software for Windows (with Python source code available) demonstrating trACT's core functions (including TMP computations, motion detection, and multi-target verification) using pre-loaded sample videos, with support for custom video uploads.