SA-Co Dataset for UAV Collision Avoidance

• SA-Co dataset is a high-resolution collection of annotated image sequences capturing fixed-wing aircraft on a collision course for vision-based sense and avoid research.
• The dataset comprises 55,521 frames across 15 sequences recorded under varied environmental conditions with detailed per-frame centroid and range annotations.
• It supports rigorous benchmarking with a baseline HMM-based detection pipeline and evaluation protocols for mid-air collision avoidance in UAV systems.

The SA-Co dataset ("A Dataset of Stationary, Fixed-wing Aircraft on a Collision Course for Vision-Based Sense and Avoid") is a public, high-resolution collection of annotated image sequences specifically designed to enable development and benchmarking of vision-based sense-and-avoid (SAA) algorithms for small-to-medium unmanned aerial vehicles (UAVs) in collision-course scenarios. Comprising 55,521 frames, SA-Co captures fixed-wing aircraft on head-on approaches toward a stationary ground-based camera, providing both image data and per-frame ground-truth centroids and ranges. The dataset facilitates rigorous study and evaluation of detection pipelines for mid-air collision avoidance in the presence of real-world atmospheric and visual conditions (Martin et al., 2021).

1. Dataset Scope and Data Acquisition

SA-Co consists of 15 image sequences: seven single-aircraft encounters (S1–S7) and eight multi-aircraft runs (M1–M8). Each sequence records a medium-sized, fixed-wing Partenavia P68 aircraft, in some cases with additional aircraft present, flying a standard landing circuit and approaching a stationary camera head-on, with no apparent bearing shift in the image plane.

Data acquisition details include:

• Camera System: Marshall CV346, ProRes 4:2:2 recording, producing TIFF images at 1280×720 px and 25 fps. Auto-exposure is enabled; no on-board stabilization is present.
• Mounting and Geometry: Camera positioned 1.7 m above ground on a tripod, located 160 m east of the 28L runway threshold at Archerfield Airport (QLD, Australia; 27.57°S, 153.01°E), aligned with the threshold’s centerline.
• Approach Profiles: Aircraft fly standard patterns with final legs ranging 1.6–3.2 NM (≈3–6 km) from the camera. Typical approach speeds are 30–40 m/s (60–80 kt) at an altitude of ≈300 m AGL.
• Environmental Variability: Sequences include clear sky, haze, clouds, bright solar reflections, and negative contrast scenes. Some multi-aircraft sequences feature navigation or landing lights.

This setup is designed to capture the full approach, from when aircraft first emerge from complex cluttered backgrounds to the moment of arrival or crossing overhead.

2. Annotation, Labeling, and Ground Truth

Each frame is annotated with a CSV entry containing:

• frame_number
• Centroid coordinates $(u, v)$ marking the target aircraft's pixel position (origin at top-left, $u$ increasing right, $v$ increasing down)
• Range $r(t)$: Slant-range from camera to aircraft, derived from GPS/RTK logs.

Annotations use a centroid-only schema; no bounding boxes or segmentation masks are provided. In multi-aircraft sequences, only the aircraft on a collision course is labeled. Manual labeling is used at critical trajectory points (e.g., appearance, disappearance, lateral movements), with linear interpolation to approximate centroids elsewhere. Increased manual labeling density is used when the target is first visible or moves laterally.

The time-to-collision (TTC) for a frame can be computed as

$\mathrm{TTC}(t) = \frac{d(t)}{v_{rel}(t)},$

where $d(t)$ is the instantaneous slant range and $v_{rel}(t)$ (closing velocity) is typically approximated via $v_{closure}$ for near-constant speed approaches.

3. Organization, Access, and Integrity

SA-Co is released with the following directory structure:

Directory/Files	Contents	Example
README.md	Dataset guide	Root
checksums.txt	SHA-256 checksums for integrity verification	Root
S1/, ..., M8/	Sequence folders	S3/, M2/
images/00000000.tif ...	Image frames (zero-padded numbers)	images/00004422.tif
labels.csv	Per-frame annotations	labels.csv

No official train/validation/test split is imposed. Practitioners typically use S1–S4 + M1–M4 for training, S5–S6 + M5–M6 for validation, and S7 + M7–M8 for testing, or perform cross-validation across sequences.

The dataset (version v1.0) is available for direct HTTP/HTTPS download at https://qcr.github.io/dataset/aircraft-collision-course/, accompanied by checksums for data verification.

4. Benchmarking and Evaluation Protocols

The dataset includes results from a baseline detection pipeline:

• ISD-4I Detector: This baseline employs a hidden Markov model (HMM) detector using a "close minus open" morphological pre-filter, replacing the conventional top-hat operator. Four parallel HMM filters operate on motion patches.
• Detection Thresholding: Detection thresholds are tuned to achieve zero false alarms (ZFA) across the full set of 15 sequences.
• Detection Ranges: For S1–S7 (single-aircraft runs), the mean ZFA detection range is 1,635.4 m (±128 m standard error).
• Evaluation Metrics: Each sequence is evaluated for:
  • $\mathrm{Precision} = \frac{TP}{TP + FP}$
  • $\mathrm{Recall} = \frac{TP}{TP + FN}$
  • $$\mathrm{F1\ Score} = \frac{2 \,\mathrm{Precision} \cdot \mathrm{Recall}}{\mathrm{Precision} + \mathrm{Recall}}$$
  • $$\text{Mean Detection Range} = \frac{1}{N} \sum_i D_i$$
  • where $D_i$ is the true-positive detection distance in sequence $i$.

System Operating Characteristic (SOC) curves are provided, plotting mean detection distance versus mean false alarms as a function of detection threshold. Frame-level false alarms are measured as the number of frames with test statistics over the given threshold. The recommended protocol is to report both detection ranges at fixed low-to-moderate false alarm rates (e.g., 0.01 FPF, 0 FPF) and complete SOC curves.

5. Applications and Analytical Opportunities

SA-Co is intended for:

• Training and evaluating vision-based SAA pipelines, including morphological pre-processing, temporal filtering, and deep learning approaches incorporating time-series structure.
• Studying the "stationary target" failure mode of frame-differencing detection schemes.
• Benchmarking new detection algorithms on true head-on approach scenarios (characterized by zero relative bearing rate, challenging for motion-based detection).
• Investigating challenging visual features such as variable target saliency (bright/dim targets), intermittent light sources, and cloud occlusions.

A plausible implication is that SA-Co enables focused research into detection challenges unique to zero apparent motion targets and background clutter conditions common in real-world airspace.

6. Limitations and Usage Considerations

Principal limitations of SA-Co include:

• Stationary Camera: The fixed ground-based perspective excludes ego-motion cues and rotational dynamics present in cockpit or airborne systems.
• Single Approach Geometry: Only head-on (zero bearing) collision courses are represented; cross-wind, off-angle, and drift effects are absent.
• Annotation Scope: Only centroids are provided—bounding boxes or pixel-perfect segmentation are not available; centroid labels may be subjective and are designated for verification rather than precise training.
• Aircraft and Scene Diversity: Only the Partenavia P68 aircraft is used, and speed/appearance regimens are limited; dynamic ground clutter and heavy occlusions are not represented.

These aspects should be considered when interpreting algorithmic performance on the dataset or transferring trained models to alternate sensing scenarios.

7. Summary and Availability

SA-Co delivers 55,521 frames at 25 fps, covering 15 high-fidelity approach sequences under challenging environmental conditions, with comprehensive per-frame centroid and range labels. A working HMM-based baseline with detection metrics provides immediate benchmarking capability. The dataset and supporting documentation are available for download, with integrity verification via provided SHA-256 checksums. SA-Co constitutes the first public resource for rigorous study of medium-sized fixed-wing aircraft on a collision course with a stationary observer, directly supporting advancements in robust vision-based sense-and-avoid systems for UAVs (Martin et al., 2021).