STAGNet: Graph & LSTM for Dash-Cam Accident Anticipation

• The paper introduces STAGNet, a framework that uses spatio-temporal graphs and an LSTM to predict accidents early from dash-cam videos with improved Average Precision and Time-to-Accident.
• It constructs three nested graphs to model object interactions, spatial proximity, and temporal consistency, employing YOLO, VGG-16, and GATv2 for robust feature extraction.
• Empirical evaluations on multiple datasets demonstrate that STAGNet outperforms prior methods by achieving higher precision and earlier warnings, despite challenges like motion blur and variable lighting.

STAGNet is a Spatio-Temporal Graph and LSTM framework for accident anticipation from dash-cam video. It takes as input a short dash-cam video clip $(X_1,\dots,X_N)$ of $N$ consecutive frames and outputs, at each time $t$, a probability $p_t\in[0,1]$ that an accident will occur in the near future, together with an implied anticipation time $\tau$. The formulation is explicitly targeted at dash-cam-only operation for Advanced Driver Assistance Systems (ADAS), a setting described as practical because it is ubiquitous and inexpensive, but challenging because it lacks depth and radar/LiDAR cues and is affected by heavy motion blur and variable lighting (Vipulananthan et al., 21 Aug 2025).

1. Problem formulation and anticipation criterion

STAGNet is defined around accident anticipation rather than post-event recognition. For positive clips, if the ground-truth accident start time is $a$, the Time-to-Accident at frame $t$ for $t<a$ is

$\mathrm{TTA}(t)=a-t.$

The model is said to anticipate the accident when its predicted probability first exceeds a threshold $\alpha$. This makes the temporal location of the first confident warning operationally central to evaluation.

The paper positions this task as a video-only alternative to multi-sensor systems. The motivation is not that dash-cam input is easier; rather, it is more deployable. The description emphasizes the absence of explicit depth information, the lack of radar or LiDAR support, and the need to operate under unconstrained visual conditions. Within that constraint set, STAGNet aims to improve accident anticipation by incorporating better spatio-temporal features and aggregating them through a recurrent network.

2. Graph construction and feature representation

The architecture is described as building three nested graphs: an object graph per frame, a temporal link between objects across adjacent frames, and a frame-level graph. The detailed construction given for the object-level pathway begins by running YOLO at each frame $N$0 to detect up to $N$1 objects, indexed by $N$2 (Vipulananthan et al., 21 Aug 2025).

For each node $N$3, STAGNet extracts two feature types. The first is a visual feature from pretrained VGG-16, $N$4, with $N$5. The second is a word embedding of the class label from GloVe, $N$6, with $N$7. These are concatenated into the node feature

$N$8

Spatial edges are defined densely within a frame: every pair $N$9 is connected. The spatial adjacency $t$0 is weighted by inverse distance in pixels using the 2D centers $t$1 of the detected boxes,

$t$2

The stated interpretation is direct: closer objects receive stronger edges.

Temporal edges link object instances across adjacent frames. Node $t$3 at time $t$4 is linked to node $t$5 at time $t$6 if they share the same class label, $t$7. The temporal adjacency $t$8 is weighted by cosine similarity of the corresponding visual features:

$t$9

In practice, $p_t\in[0,1]$0, reflecting the short 5 s clips used in the study. This construction yields a local object-interaction graph within a frame and a short-range object-consistency graph across time.

3. GATv2 processing and recurrent temporal aggregation

The graph neural component applies two parallel GATv2 updates, one over the spatial adjacency and one over the temporal adjacency. Before message passing, the input features are reduced to a common embedding dimension $p_t\in[0,1]$1 by a trainable $p_t\in[0,1]$2, separately for visual and label embeddings. The reduced node representation is

$p_t\in[0,1]$3

The spatial and temporal object features are then computed as

$p_t\in[0,1]$4

$p_t\in[0,1]$5

At the node level, each GATv2 layer is written as

$p_t\in[0,1]$6

Self-loops are added so that $p_t\in[0,1]$7, and the nonlinearity is specified as $p_t\in[0,1]$8 or ELU. The spatial and temporal outputs are fused by concatenation,

$p_t\in[0,1]$9

In parallel with the object graph, STAGNet extracts one global feature per frame through a pretrained SlowFast network, followed by a fully connected layer and an LSTM (Vipulananthan et al., 21 Aug 2025). The frame feature pipeline is

$\tau$0

The reduced sequence $\tau$1 is then processed by an LSTM of hidden size $\tau$2, with the usual gates $\tau$3, cell state $\tau$4, and hidden state $\tau$5. The description states that $\tau$6 encodes global spatio-temporal context up to time $\tau$7. Final score prediction concatenates a pooled object-graph summary with $\tau$8 and applies a fully connected layer plus sigmoid:

$\tau$9

The architecture therefore combines explicit object-object structure, short-range temporal correspondence, and a recurrent global context pathway.

4. Objective, hyperparameters, and runtime profile

Training uses binary cross-entropy at each frame,

$a$0

where $a$1 is the ground-truth accident indicator. The description adds a notable labeling detail: in positive videos, all frames prior to the accident carry $a$2. No auxiliary losses are used beyond standard cross-entropy (Vipulananthan et al., 21 Aug 2025).

The reported key hyperparameters are a learning rate of $a$3 with Adam, batch size $a$4 clips, one spatial GATv2 layer and one temporal GATv2 layer for the object graph, two parallel GATv2 layers for the frame graph, embedding dimension $a$5, LSTM hidden size $a$6, number of objects per frame $a$7, temporal window $a$8, and frame-graph window $a$9. The model is described as modular, comprising object-graph, frame-graph, and global sequence LSTM components, and as trainable end-to-end except for frozen detectors and feature extractors.

Execution speed is reported as approximately $t$0 FPS end-to-end on a Tesla P100 when using SlowFast. The limitations section attributes much of the runtime burden to the heavy VGG-16 and SlowFast modules.

5. Benchmarks, metrics, and empirical performance

The evaluation uses three publicly available datasets. DAD contains 1 750 5 s clips at 20 fps, with 620 positive and 1 130 negative clips, and fewer than 10 % ego-involved accidents. The ego subset of DoTA contains 1 490 clips at 10 fps, split 745/745. The ego subset of DADA contains 2 299 clips at 10 fps, with 938 positive and 1 361 negative clips. Performance is reported with Average Precision (AP), defined as the area under the precision-recall curve as threshold $t$1 varies, and mean Time-to-Accident (mTTA), defined as the average $t$2 over positive clips at the $t$3 that maximizes AP (Vipulananthan et al., 21 Aug 2025).

Dataset	STAGNet	Prior best
DAD (5-fold CV)	AP = 64.36 %, mTTA = 4.54 s	Graph(Graph): AP = 62.80 %, mTTA = 4.29 s
DoTA (ego)	AP = 92.73 %, mTTA = 3.23 s	Graph(Graph): AP = 83.50 %, mTTA = 3.14 s
DADA (ego)	AP = 95.39 %, mTTA = 3.08 s	Graph(Graph): AP = 84.04 %, mTTA = 2.92 s

Cross-dataset generalization is also reported. When trained on DoTA and tested on DADA, STAGNet achieves AP $t$4 versus $t$5 for Graph(Graph). When trained on DADA and tested on DoTA, it achieves AP $t$6 versus $t$7. The paper explicitly presents these results as evidence of improved generalization on ego-involved crashes.

The ablation study attributes distinct gains to several components. Replacing SlowFast with I3D drops AP by approximately $t$8–$t$9. Removing the global-LSTM drops AP by approximately $t<a$0–$t<a$1. Replacing GCN with GATv2 yields a minor gain of about $t<a$2. These ablations localize the strongest contribution to the global temporal aggregation pathway and the choice of frame-level feature extractor, with the attention-based graph update providing a smaller but positive increment.

6. Contributions, limitations, and plausible research trajectory

The stated novelty of STAGNet consists of four elements: the use of state-of-the-art SlowFast networks for frame-level spatio-temporal features, the aggregation of global features via an LSTM beyond the SlowFast temporal receptive field, the use of GATv2 rather than plain GCN so that each node can attend selectively to its neighbors, and a thorough cross-dataset evaluation on ego-involved crashes using DoTA and DADA (Vipulananthan et al., 21 Aug 2025).

The strengths claimed for the framework are substantially higher AP and earlier mTTA across three benchmarks, a modular decomposition into object-graph, frame-graph, and global sequence LSTM components, and end-to-end training except for frozen detectors and feature extractors. The experimental record reported in the paper is consistent with those claims.

The limitations are stated with equal clarity. Cross-dataset performance remains limited, and the remaining gap is used to point to domain bias in video quality and scene semantics. Inference speed, at approximately $t<a$3 FPS, is identified as another constraint. The future extensions proposed in the paper are segmentation-based object localization instead of bounding boxes, inclusion of approximate depth through camera intrinsics or monocular depth networks, and lightweight graph/LSTM variants for real-time embedded ADAS. A plausible implication is that the model’s current bottlenecks are not only architectural but also representational: object localization fidelity, latent geometry, and domain transfer are treated as open issues rather than solved components.

7. Terminological overlap and disambiguation

The label “STAGNet” is not unique to the accident-anticipation framework. In graph learning, the name also appears in a distinct line of work on stochastic aggregation, where a GNN replaces deterministic aggregation with a randomly re-weighted adjacency $t<a$4, and may further learn a variational posterior over the mask tensor $t<a$5 (Wang et al., 2021). That model is concerned with over-smoothing, discriminating power, and structural uncertainty in generic graph tasks such as citation-node classification and molecule regression, rather than dash-cam accident anticipation.

A separate remote-sensing model, “STAG-NN-BA,” has also been described as STAGNet. It constructs a supergraph from block-diagonal concatenation of region adjacency graphs across time and injects temporal bias by augmenting node features, targeting spatial and spatio-temporal classification with satellite imagery (Nazir et al., 2023). Its data pipeline is based on SLIC superpixels and region adjacency graphs rather than detected traffic actors in dash-cam video.

Accordingly, in current usage, STAGNet most precisely denotes the dash-cam accident-anticipation model only when the context is the 2025 paper subtitled “A Spatio-Temporal Graph and LSTM Framework for Accident Anticipation.” In technical discussion, disambiguation by arXiv identifier is often necessary because the same label has been applied to materially different architectures and problem domains.