Drone-Assisted Naloxone Delivery Simulation Dataset

• The paper introduces DANDSD as a pioneering dataset capturing drone-assisted naloxone delivery simulations to assess real-time bystander situational awareness.
• It employs a two-stage compositional framework combining YOLOv8, BoT-SORT, graph embeddings, and transformer models to integrate low-level perception and high-level comprehension.
• The dataset enables precise SA classification and temporal segmentation, offering actionable insights for enhancing human-autonomy teaming in emergency scenarios.

The Drone-Assisted Naloxone Delivery Simulation Dataset (DANDSD) is a public, annotated video dataset of drone-assisted naloxone delivery simulations designed to support real-time bystander situation awareness (SA) estimation during simulated opioid overdose emergency (OOE) scenarios. It captures human-autonomy teaming (HAT) from the observer/bystander perspective, with medically untrained college students acting as bystanders who attempt to administer intranasal naloxone (Narcan™) delivered by a drone to a mock overdose victim (mannequin). The dataset was introduced together with a video-based real-time SA assessment framework that combines graph embeddings and transformer models for SA prediction and temporal segmentation (Chang et al., 3 Oct 2025).

1. Dataset definition and research role

DANDSD was created to address a specific gap: there was no publicly available dataset for analyzing bystander behavior and SA in drone-assisted naloxone delivery during opioid overdose emergencies. The dataset is explicitly designed to support real-time assessment of the bystander’s situational awareness while an emergency response scenario unfolds. In the source paper, it is positioned as a first-of-its-kind dataset for real-time drone-assisted first aid, observer-rated SA annotations, temporal event boundary annotations, and downstream support for both SA prediction and temporal segmentation (Chang et al., 3 Oct 2025).

The motivating premise is that the success of drone-assisted emergency first aid depends not only on autonomous delivery, but also on the bystander’s ability to perceive the scene, understand what to do, and anticipate what happens next. This framing aligns the dataset with SA-centric HAT research rather than generic action recognition. A plausible implication is that DANDSD is intended less as a conventional vision benchmark than as a substrate for modeling cognitive state transitions during task execution.

The scenario captured by DANDSD is task-driven and spans the full rescue workflow: the bystander encounters the victim, the drone assists with naloxone delivery, the bystander interacts with the drone and scene, and naloxone is successfully administered. The recordings therefore encode action, instruction following, object interaction, and state transition within a single continuous emergency-response sequence.

2. Scenario construction, participants, and corpus statistics

DANDSD is derived from a prior simulation study comprising 17 simulated trials, from which the dataset contains 11 continuous, uninterrupted videos. Each video is 2–3 minutes long, recorded at 50 fps, with a total of 92,917 frames. The machine-learning split reported in the paper includes 5,575 sequences for training and 620 sequences for testing, with each sequence length = 15 frames; sequences are shuffled during training (Chang et al., 3 Oct 2025).

The human participants are college students without medical training, who serve as bystanders. Other scene entities are the mock overdose victim / mannequin, an instructor, and a drone. These four entities also define the object set used later in the interaction graph formulation. The bystander’s task is to recognize the emergency, interact with the drone-delivered naloxone, follow guidance, and administer intranasal naloxone to the mannequin.

Dataset component	Reported value
Simulated trials	17
Continuous videos	11
Video duration	2–3 minutes
Frame rate	50 fps
Total frames	92,917
Training sequences	5,575
Testing sequences	620
Sequence length	15 frames

Because the corpus consists of simulated OOEs rather than real emergency footage, it should not be construed as a naturalistic dataset of field opioid overdoses. The paper instead emphasizes its utility for controlled study of SA in drone-assisted first aid. This suggests that its principal methodological value lies in structured annotation, temporal continuity, and explicit HAT interaction, rather than in uncontrolled real-world variability.

3. Annotation design and SA label construction

DANDSD contains two annotation types: time interval annotations and SA rating annotations. For temporal structure, the videos were manually segmented into distinct events based on bystander actions. Two annotators independently reviewed the videos, reached consensus on event boundaries, and marked boundaries by specific movements or actions between events. The temporal segmentation task uses five predefined events (Chang et al., 3 Oct 2025).

For SA, each video was divided into 10 equally sized clips, yielding 110 clips total, with clip length = 30 seconds. Two domain experts independently rated each clip on a 1–5 scale for three SA levels:

1. Perception: “To what extent has the bystander observed all necessary visual cues?”
2. Comprehension: “How well does the bystander understand the situation and required actions?”
3. Projection: “To what degree does the bystander anticipate future developments and consequences?”

A critical methodological detail is that each expert gives one single rating per clip after reviewing the full clip. The reported interpretation is that each rating reflects the expert’s holistic judgment at the final timestamp of the clip. To obtain frame-level supervision, the authors linearly interpolated between rated points and reset SA to 0 at event boundaries, thereby producing a continuous SA trajectory over time.

The paper formulates SA classification in both binary and ternary forms. For each SA dimension, binary labels are thresholded at 3:

$$\text{High SA} \iff \text{rating} \ge 3,\quad \text{Low SA} \iff \text{rating} < 3$$

For frame $i$,

$$[Per_i, Com_i, Pro_i], \quad Per_i,Com_i,Pro_i \in \{0,1\}$$

where $1$ indicates high SA.

The ternary formulation uses Endsley’s hierarchical model, in which higher SA levels are meaningful only if lower levels are achieved. The binary levels are accumulated into a ternary class label $\{0,1,2\}$. The paper gives illustrative mappings: $[0,1,1]$ maps to class 0 because perception is not high, while $[1,1,1]$ maps to the highest class.

This annotation design matters because the dataset does not merely provide class labels for isolated clips; it provides a mechanism for deriving continuous SA curves linked to event structure. A plausible implication is that DANDSD supports hybrid inference over both cognition and action boundaries, which is unusual in task-specific HAT datasets.

4. Modalities, feature representation, and compositional framework

The associated framework is described as a two-stage compositional learning framework consisting of a Perception Module and a Comprehension Module. The architecture is intended to make SA prediction more interpretable and modular than an end-to-end black-box model (Chang et al., 3 Oct 2025).

The Perception Module handles low-level visual understanding, including object detection, tracking, and pose estimation. It uses YOLOv8 for object detection, BoT-SORT for multi-object tracking, and 2D keypoint estimation for body pose. Its outputs include bystander bounding boxes and pose keypoints. For each frame $i$, the bystander bounding box is

$b_i = \{x_1, y_1, x_2, y_2\}$

and the bystander’s body pose is represented with 17 2D keypoints:

$P_i := \{p_1, p_2, \dots, p_{17}\}.$

The Comprehension Module captures higher-level reasoning through object centers, depth estimates, graph embeddings, and transformer sequence modeling. The scene is modeled as a fully connected graph over four objects: bystander, instructor, patient, and drone. Node attributes include 2D coordinates, depth label, and body points for humans; the drone receives zero-padded pose features because pose estimation is not applicable. To estimate spatial relations, the method uses monocular depth estimation via DPT with ViT backbone. For each frame $i$0, a depth label $i$1 is derived from the disparity map $i$2 at the center point $i$3.

The graph convolution update is defined as

$i$4

where $i$5 denotes node features at layer $i$6, $i$7 the adjacency matrix, $i$8 learned weights, and $i$9 the activation function. A two-layer GCN autoencoder embeds the interaction graph into compact representations:

$$[Per_i, Com_i, Pro_i], \quad Per_i,Com_i,Pro_i \in \{0,1\}$$0

The resulting graph embedding $$[Per_i, Com_i, Pro_i], \quad Per_i,Com_i,Pro_i \in \{0,1\}$$1 is reported to capture spatial relations, pose context, relative object proximity, and interaction structure.

The final SA predictor is a Transformer-based sequence model, inspired by TrEP. For each frame $$[Per_i, Com_i, Pro_i], \quad Per_i,Com_i,Pro_i \in \{0,1\}$$2, the extracted features are concatenated as

$$[Per_i, Com_i, Pro_i], \quad Per_i,Com_i,Pro_i \in \{0,1\}$$3

and the input sequence is

$$[Per_i, Com_i, Pro_i], \quad Per_i,Com_i,Pro_i \in \{0,1\}$$4

where $$[Per_i, Com_i, Pro_i], \quad Per_i,Com_i,Pro_i \in \{0,1\}$$5 is the sequence length. The transformer projects features through a shared feed-forward layer, adds positional encoding, applies multiple transformer blocks, then flattens and classifies the output. Each transformer block uses multi-head self-attention and feed-forward sublayers. The binary SA task uses sigmoid activation and binary cross-entropy; the ternary SA task uses softmax activation and categorical cross-entropy.

5. Temporal segmentation through latent SA dynamics

A central contribution associated with DANDSD is the use of predicted SA trajectories to infer event boundaries. The stated logic is that SA changes reflect cognitive transitions; when SA resets or changes sharply, a new event is likely beginning. On that basis, latent SA cues are used to segment untrimmed video into meaningful actions (Chang et al., 3 Oct 2025).

This segmentation procedure depends directly on the annotation and label-construction scheme. Because SA is interpolated over time and reset to 0 at event boundaries, the trajectory is not treated merely as a static supervisory signal but as a temporal state variable. The model tracks SA over time, smooths it, and identifies transition points. The paper applies a 13-frame Gaussian filter, corresponding to an approximate human reaction time of 0.25 s; at 50 fps, 13 frames $$[Per_i, Com_i, Pro_i], \quad Per_i,Com_i,Pro_i \in \{0,1\}$$6 s, which the paper notes is consistent with the stated reaction time.

Temporal segmentation is evaluated using Mean over Frames (MoF) and Intersection over Union (IoU). The reported definitions are

$$[Per_i, Com_i, Pro_i], \quad Per_i,Com_i,Pro_i \in \{0,1\}$$7

where $$[Per_i, Com_i, Pro_i], \quad Per_i,Com_i,Pro_i \in \{0,1\}$$8 is the total number of frames, $$[Per_i, Com_i, Pro_i], \quad Per_i,Com_i,Pro_i \in \{0,1\}$$9 the true label at frame $1$0, $1$1 the predicted label at frame $1$2, and $1$3 the indicator function; and

$1$4

where $1$5 is the predicted segmentation and $1$6 the ground-truth segmentation.

The paper notes that MoF is frame-wise accuracy but can be sensitive to class imbalance. This is relevant because DANDSD contains imbalanced SA labels. For ternary labels, the counts are 34,453 for class 0, 33,587 for class 1, and 24,877 for class 2. For binary labels, the reported counts are 46,498 / 46,419 for Perception, 48,838 / 44,079 for Comprehension, and 55,885 / 37,032 for Projection.

6. Evaluation protocol and empirical findings

For SA prediction, the framework is evaluated using Accuracy, AUC, F1, Precision, and Balanced Accuracy (BAcc). The reported protocol uses random sampling for balanced representation, 10-fold cross-validation, and early stopping with tolerance of 5 epochs (Chang et al., 3 Oct 2025).

For ternary SA prediction, the paper compares the proposed method against C3D, I3D, X3D, and SlowFast. The reported results are as follows:

Model	Accuracy	AUC	F1
C3D	0.54	0.50	0.33
I3D	0.55	0.52	0.47
X3D	0.60	0.57	0.55
SlowFast	0.60	0.58	0.56
Ours	0.63	0.64	0.62

The corresponding Precision values are 0.32 for C3D, 0.45 for I3D, 0.54 for X3D, 0.54 for SlowFast, and 0.62 for the proposed method. The paper states that the model improves over the best baseline by about 3% to 8% across metrics.

Feature ablation further tests Bbox, Pose, Graph, and their combinations. For binary per-class top-1 accuracy, the best result comes from Bbox + Pose + Graph: 0.71 for Perception, 0.60 for Comprehension, and 0.74 for Projection. For ternary classification, the best result is again Bbox + Pose + Graph, with Accuracy = 0.63, Balanced Accuracy = 0.62, and F1 = 0.62. The paper additionally reports that pose alone performs better than bbox alone in ternary SA, that graph alone is weak by itself, and that graph adds value when combined with other features. In the paper’s interpretation, this supports the importance of interaction structure for understanding situational awareness.

For temporal segmentation, the SA-informed model is compared with TW-FINCH baselines. The reported results are:

Method	MoF	IoU
TW-Finch (cls = 6)	0.41	0.23
TW-Finch (cls = 7)	0.49	0.29
TrSA	0.58	0.34

The paper states that TrSA outperforms TW-FINCH (cls = 7) by 9% in MoF and 5% in IoU. In the framing of the work, this result indicates that latent SA predictions can improve event segmentation, not merely SA classification.

7. Interpretation, significance, and recurrent points of clarification

DANDSD is significant in several explicitly stated ways. For human-autonomy teaming, it provides a concrete benchmark for studying how AI systems can support humans in high-stakes collaborative tasks. For drone-assisted first aid, it supports systems that assess whether bystanders are prepared to follow life-saving instructions during naloxone delivery. For adaptive emergency-response systems, it opens the door to future drones that can detect confusion, adjust instructions, time guidance based on bystander SA, and support more effective first intervention. The paper also states that the dataset is publicly available, together with source code, to enable reproducibility and future work in SA modeling, temporal segmentation, emergency robotics, and human-drone interaction (Chang et al., 3 Oct 2025).

Several clarifications are important for correct interpretation. First, DANDSD is not a dataset of real OOEs; it is a dataset of simulated opioid overdose emergency scenarios with a mock overdose victim (mannequin). Second, its SA labels are not instantaneous self-reports from participants. They are observer-rated, clip-level expert judgments assigned after review of the full clip, then transformed into continuous supervision by interpolation and boundary reset. Third, the paper does not present graph structure as sufficient on its own; rather, graph alone is weak by itself, but contributes when fused with bounding-box and pose information. Fourth, the framework’s temporal segmentation is not derived from direct event classification alone; it is derived from latent SA changes, which the paper treats as informative about event transitions.

The broader conceptual claim advanced by the work is that SA is not just an outcome label; it is also a useful latent signal for understanding temporal structure. Within that view, DANDSD functions as a dataset for jointly studying perception-level cues, comprehension-level scene relations, and projection-related temporal evolution in an emergency HAT setting. A plausible implication is that the dataset may be most valuable where the central research question concerns how human cognitive state and embodied action co-evolve under real-time autonomous assistance.