paper/main.tex

\documentclass[runningheads]{llncs}

\usepackage[T1]{fontenc}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage{multirow}
\usepackage{amsmath}
\usepackage{url}
\usepackage{adjustbox}
\usepackage{tikz}
\usetikzlibrary{arrows.meta,positioning,shapes.geometric,fit}
\emergencystretch=2em

\title{DynaFall: Robust Skeleton-Based Fall Detection via Keypoint Dynamics and Graph-Based Benchmarking}
\titlerunning{DynaFall for Skeleton-Based Fall Detection}

\author{Anonymous Submission}
\authorrunning{Anonymous}
\institute{}

\begin{document}
\maketitle

\begin{abstract}
Automatic fall detection is a safety-critical problem for elderly-care and assisted-living environments, but vision-based systems must balance accuracy, privacy, robustness, and computational cost. This paper presents DynaFall-GCN, a skeleton-based fall-detection framework that converts RGB videos into 17-keypoint pose sequences and performs classification only on normalized skeleton features. Unlike generic action-recognition models that mainly learn joint trajectories, DynaFall-GCN explicitly combines three complementary streams: joint coordinates, bone vectors, and fall-oriented dynamics including velocity, acceleration, torso angle, hip-drop motion, body aspect ratio, and center-of-body displacement. A confidence-aware keypoint dropout strategy is further used to expose the model to pose-estimation failures during training. To make the evaluation stronger than a single-model report, we benchmark DynaFall-GCN against LSTM, ST-GCN, 2s-AGCN-style adaptive GCN, CTR-GCN-style refined GCN, PoseC3D-style heatmap CNN, and a TCN--Transformer baseline on URFD and MCFD. Experiments are conducted over three random seeds with video-level URFD splits and scenario-level MCFD splits. DynaFall-GCN achieves the best URFD F1 and macro-F1, and the highest MCFD fall recall among the evaluated methods. Ablation and efficiency results show that explicit dynamics are the most reliable contributor to clean-set performance, while confidence-aware dropout shifts the detector toward higher sensitivity under challenging multi-view conditions.
\keywords{Fall detection \and Skeleton action recognition \and Graph convolutional networks \and Keypoint dynamics \and Pose estimation}
\end{abstract}

\section{Introduction}
Falls are a major risk for older adults because delayed assistance after a fall can lead to severe injury, long-lie complications, loss of independence, and higher care cost. Vision-based fall detection is attractive because it does not require the user to wear or recharge a device, and recent surveys show that deep learning has become a dominant direction for non-intrusive fall monitoring~\cite{alam2022review}. However, raw RGB video also introduces practical concerns: appearance cues may encode private information, background and clothing can bias a classifier, and large video models can be expensive for continuous deployment.

Skeleton-based recognition is a pragmatic compromise. A pose estimator first converts each frame into body keypoints, after which the fall classifier only sees motion geometry rather than raw appearance. This representation is compact, privacy-friendlier, and naturally aligned with the physical structure of a fall: rapid center-of-body motion, torso rotation, hip descent, and a transition from upright to horizontal posture. At the same time, skeleton-based fall detection is not solved by simply applying a small recurrent model. Pose keypoints can be noisy near the floor, cameras can observe the same fall from very different viewpoints, and fall datasets are usually small and imbalanced. These issues motivate a careful benchmark, not only a proposed model.

This work focuses on two established fall datasets: UR Fall Detection (URFD)~\cite{urfd} and Multiple Cameras Fall Dataset (MCFD)~\cite{mcfd}. URFD is useful for controlled RGB-to-skeleton experiments, while MCFD stresses multi-view generalization. The complete processing pipeline is shown in Fig.~\ref{fig:pipeline}. We use a consistent protocol: all available videos are converted to pose clips, train/validation/test splits are formed before clip generation, and all compared methods use the same clips, optimizer, weighted focal loss, and seeds.

The contributions are threefold. First, we build a unified RGB-to-skeleton benchmark for URFD and MCFD with a leakage-aware split protocol. Second, we compare representative sequence, graph, heatmap, and temporal-attention baselines under the same experimental setting. Third, we propose DynaFall-GCN, a lightweight three-stream model that augments joint and bone features with fall-specific dynamics and confidence-aware keypoint dropout. Tables~\ref{tab:datasets}--\ref{tab:ablation} and Fig.~\ref{fig:runtime} report the dataset statistics, main results, ablation, and efficiency trade-offs.

\begin{figure}[t]
\centering
\begin{tikzpicture}[
  node distance=0.55cm,
  block/.style={draw, rounded corners, align=center, minimum height=0.72cm, minimum width=2.15cm, font=\scriptsize},
  arrow/.style={-{Latex[length=2mm]}, thick}
]
\node[block] (rgb) {RGB video};
\node[block, right=of rgb] (sample) {Frame sampling\\$T=32$};
\node[block, right=of sample] (pose) {YOLO pose\\17 keypoints};
\node[block, right=of pose] (norm) {Box-centered\\normalization};
\node[block, below=of norm] (feat) {Joint + bone\\+ dynamics};
\node[block, left=of feat] (models) {Baselines and\\DynaFall-GCN};
\node[block, left=of models] (pred) {Fall /\\non-fall};
\draw[arrow] (rgb) -- (sample);
\draw[arrow] (sample) -- (pose);
\draw[arrow] (pose) -- (norm);
\draw[arrow] (norm) -- (feat);
\draw[arrow] (feat) -- (models);
\draw[arrow] (models) -- (pred);
\end{tikzpicture}
\caption{Overall RGB-to-skeleton fall-detection pipeline used for all experiments.}
\label{fig:pipeline}
\end{figure}

\section{Related Work}
\textbf{Vision-based fall detection.}
Deep vision-based fall detection has moved from handcrafted motion and silhouette descriptors to CNN, recurrent, temporal-convolution, transformer, and pose-based models. A recent review emphasizes that non-intrusive vision systems are promising but still face dataset bias, class imbalance, viewpoint variation, occlusion, and real-time deployment constraints~\cite{alam2022review}. These issues motivate our evaluation protocol: Table~\ref{tab:main} reports accuracy together with precision, recall, fall-class F1, and macro-F1 instead of accuracy alone.

\textbf{Fall datasets and evaluation protocol.}
URFD combines falls and activities of daily living with RGB, depth, and inertial measurements~\cite{urfd}. MCFD contains multi-camera fall scenarios and confounding non-fall activities observed from eight cameras~\cite{mcfd}. Because these datasets are small, leakage prevention is critical. In MCFD, camera-level splitting can place different views of the same scenario into both training and testing, so we use scenario-level splitting as summarized in Table~\ref{tab:datasets}.

\textbf{Graph-based skeleton recognition.}
Skeleton action recognition provides stronger building blocks than classical fall-detection baselines. ST-GCN models joints and bones as a spatial-temporal graph~\cite{yan2018stgcn}. 2s-AGCN improves this by learning adaptive topology and adding a bone stream~\cite{shi2019agcn}, while CTR-GCN refines topology channel-wise for more flexible joint aggregation~\cite{chen2021ctrgcn}. These models motivate our joint and bone streams, but they are designed for general action recognition rather than fall-specific motion mechanics.

\textbf{Heatmap and temporal-attention alternatives.}
PoseC3D/PoseConv3D converts skeletons into 3D heatmap volumes and applies 3D CNNs, giving a non-GCN comparison for pose-based recognition~\cite{duan2022posec3d}. More recently, TCNTE combines temporal convolutions and Transformer Encoder modules for real-time skeleton-based fall detection and uses weighted focal loss for imbalance~\cite{yu2025tcnte}. Our TCN--Transformer baseline follows this motivation, while DynaFall-GCN adds explicit fall-oriented measurements such as torso angle and hip-drop motion.

\textbf{Position of this work.}
The main gap is that strong skeleton action-recognition models and fall-specific temporal models are rarely compared under one leakage-aware RGB-to-skeleton protocol. In addition, many methods rely on learned temporal features but do not explicitly encode fall mechanics. DynaFall-GCN is designed to fill this gap: it keeps the compactness of skeleton models, borrows the joint/bone reasoning that makes graph models effective, and adds dynamics features that are physically tied to falls. Fig.~\ref{fig:architecture} summarizes this design.

\section{Method}
\subsection{Pose Extraction and Normalization}
Each RGB frame is processed by YOLOv8 pose~\cite{yolov8}. If multiple people are detected, the person with the largest bounding box is selected because both URFD and MCFD focus on one dominant subject. Each frame is represented by 17 COCO keypoints with $(x,y,c)$ coordinates, where $c$ is the pose confidence. Coordinates are normalized by the visible person box:
\begin{equation}
x' = \frac{x - x_c}{w}, \qquad y' = \frac{y - y_c}{h},
\end{equation}
where $(x_c,y_c)$ is the box center and $(w,h)$ is the box size. Videos are converted into clips of 32 frames with stride 16; shorter sequences are padded by repeating the final frame.

\subsection{Feature Construction}
For a clip with normalized joints $P_t \in \mathrm{R}^{17 \times 3}$ at time $t$, DynaFall-GCN builds three complementary feature groups. The joint feature is the normalized keypoint tensor itself. The bone feature is computed from the COCO skeleton edges:
\begin{equation}
B_t^{(i,j)} = P_t^{j} - P_t^{i},
\end{equation}
where $(i,j)$ is a connected joint pair. This representation emphasizes body configuration, limb direction, and posture change. The dynamics feature is designed around fall mechanics. It includes first-order velocity $V_t=P_t-P_{t-1}$, second-order acceleration $A_t=V_t-V_{t-1}$, torso angle from shoulder and hip centers, hip vertical displacement, body aspect ratio, and center-of-body velocity. These descriptors encode the rapid downward movement and upright-to-horizontal transition that distinguish a fall from many activities of daily living.

\subsection{DynaFall-GCN}
DynaFall-GCN uses three streams as shown in Fig.~\ref{fig:architecture}. The joint stream receives normalized keypoints and learns spatial-temporal dependencies over the human-body graph. The bone stream receives bone vectors and focuses on posture and limb orientation. The dynamics stream receives the fall-oriented features above and models short-term motion using temporal convolutions followed by attention pooling. Joint and bone streams use lightweight graph-temporal encoders with adaptive adjacency so that the model can learn both physical skeleton links and task-dependent joint relations. The resulting embeddings are concatenated and classified by a multilayer perceptron.

\begin{figure}[t]
\centering
\begin{tikzpicture}[
  node distance=0.38cm and 0.7cm,
  block/.style={draw, rounded corners, align=center, minimum height=0.64cm, minimum width=2.25cm, font=\scriptsize},
  stream/.style={draw, rounded corners, align=center, minimum height=0.64cm, minimum width=2.55cm, font=\scriptsize},
  arrow/.style={-{Latex[length=2mm]}, thick}
]
\node[block] (input) {Pose clip\\$32 \times 17 \times 3$};
\node[block, below=of input] (drop) {Confidence-aware\\keypoint dropout};
\node[stream, below left=0.6cm and 1.8cm of drop] (joint) {Joint GCN\\coordinates};
\node[stream, below=0.6cm of drop] (bone) {Bone GCN\\bone vectors};
\node[stream, below right=0.6cm and 1.8cm of drop] (dyn) {Dynamics TCN\\fall cues};
\node[block, below=1.05cm of bone] (fusion) {Feature fusion\\concatenation};
\node[block, below=of fusion] (head) {MLP classifier};
\node[block, below=of head] (out) {Fall / non-fall};
\draw[arrow] (input) -- (drop);
\draw[arrow] (drop) -- (joint);
\draw[arrow] (drop) -- (bone);
\draw[arrow] (drop) -- (dyn);
\draw[arrow] (joint) -- (fusion);
\draw[arrow] (bone) -- (fusion);
\draw[arrow] (dyn) -- (fusion);
\draw[arrow] (fusion) -- (head);
\draw[arrow] (head) -- (out);
\end{tikzpicture}
\caption{DynaFall-GCN architecture with joint, bone, and fall-specific dynamics streams.}
\label{fig:architecture}
\end{figure}

\subsection{Confidence-Aware Keypoint Dropout}
Pose estimation can fail during occlusion, motion blur, or when the body is close to the floor. During training, DynaFall applies confidence-aware keypoint dropout: high-confidence joints are masked with probability 0.100, while joints with confidence below 0.300 are masked with probability 0.500. Masking sets coordinates, confidence, and derived motion values to zero. Weighted focal loss~\cite{lin2017focal} is used for all methods to reduce class-imbalance effects.

\section{Experimental Setup}
\subsection{Datasets, Splits, and Metrics}
URFD is evaluated with video-level splits. MCFD is evaluated with scenario-level splits so that camera views of one scenario never appear in different splits. Table~\ref{tab:datasets} summarizes the processed RGB-to-pose data used in our experiments. We use seeds 7, 13, and 21. For each seed, the split ratio is 70\% train, 15\% validation, and 15\% test at the video or scenario level. Unless otherwise stated, precision, recall, and F1-score are computed for the positive fall class:
\begin{equation}
F1_{\mathrm{fall}}=\frac{2P_{\mathrm{fall}}R_{\mathrm{fall}}}{P_{\mathrm{fall}}+R_{\mathrm{fall}}}.
\end{equation}
Macro-F1 is the unweighted mean of the fall-class and non-fall-class F1 scores. Recall is reported explicitly because missing a true fall is usually more costly than producing a false alarm.

\begin{table}[t]
\centering
\caption{Dataset statistics after RGB video collection and pose extraction.}
\label{tab:datasets}
\small
\begin{adjustbox}{width=\textwidth}
\begin{tabular}{lrrrrrl}
\toprule
Dataset & Scenarios & Videos & Fall videos & Non-fall videos & Frames & Split unit \\
\midrule
URFD & -- & 70 & 30 & 40 & 11,936 & Video \\
MCFD & 24 & 192 & 176 & 16 & 261,139 & Scenario \\
\bottomrule
\end{tabular}
\end{adjustbox}
\end{table}

\subsection{Compared Methods}
All models use the same extracted pose clips, optimizer, loss, and seed protocol. The baselines are intentionally diverse: LSTM tests a simple sequence model, ST-GCN tests fixed graph reasoning, 2s-AGCN and CTR-GCN test stronger adaptive graph reasoning, PoseC3D tests a heatmap representation, and TCNTE tests a recent fall-oriented temporal design. DynaFall-GCN is evaluated as the proposed joint--bone--dynamics model.

\section{Results}
\subsection{Main Comparison}
Table~\ref{tab:main} reports mean and standard deviation across three seeds. On URFD, DynaFall-GCN obtains the best accuracy, fall-class F1, and macro-F1. This indicates that explicit dynamics are useful when pose quality is relatively stable and the clips are short. On MCFD, the scenario-level split is harder because the same physical fall is not shared across cameras between train and test. Under this stricter setting, AGCN gives the highest accuracy and macro-F1, while DynaFall-GCN reaches the highest fall recall and F1. This trade-off is important: DynaFall-GCN is more sensitive to falls, whereas AGCN is more balanced across both classes.

\begin{table}[t]
\centering
\caption{Main comparison over three seeds. Precision, recall, and F1 are computed for the fall class.}
\label{tab:main}
\small
\begin{adjustbox}{width=\textwidth}
\begin{tabular}{llccccc}
\toprule
Dataset & Method & Acc. & Prec. & Recall & F1 & Macro-F1 \\
\midrule
\multirow{7}{*}{URFD}
& LSTM & 0.762$\pm$0.051 & 0.519$\pm$0.131 & 0.654$\pm$0.196 & 0.565$\pm$0.118 & 0.700$\pm$0.069 \\
& ST-GCN & 0.742$\pm$0.059 & 0.480$\pm$0.126 & 0.614$\pm$0.167 & 0.531$\pm$0.130 & 0.676$\pm$0.081 \\
& AGCN & 0.652$\pm$0.206 & 0.467$\pm$0.223 & \textbf{0.813$\pm$0.162} & 0.556$\pm$0.143 & 0.626$\pm$0.185 \\
& CTR-GCN & 0.732$\pm$0.138 & \textbf{0.587$\pm$0.292} & 0.696$\pm$0.233 & 0.568$\pm$0.045 & 0.679$\pm$0.096 \\
& PoseC3D & 0.626$\pm$0.083 & 0.320$\pm$0.079 & 0.560$\pm$0.393 & 0.386$\pm$0.164 & 0.545$\pm$0.040 \\
& TCNTE & 0.702$\pm$0.169 & 0.489$\pm$0.265 & 0.718$\pm$0.156 & 0.558$\pm$0.203 & 0.665$\pm$0.170 \\
& DynaFall & \textbf{0.770$\pm$0.032} & 0.524$\pm$0.012 & 0.714$\pm$0.179 & \textbf{0.597$\pm$0.053} & \textbf{0.716$\pm$0.014} \\
\midrule
\multirow{7}{*}{MCFD}
& LSTM & 0.643$\pm$0.002 & 0.588$\pm$0.047 & 0.567$\pm$0.123 & 0.571$\pm$0.065 & 0.630$\pm$0.018 \\
& ST-GCN & 0.644$\pm$0.015 & 0.601$\pm$0.030 & 0.500$\pm$0.011 & 0.546$\pm$0.019 & 0.626$\pm$0.005 \\
& AGCN & \textbf{0.670$\pm$0.005} & \textbf{0.611$\pm$0.057} & 0.642$\pm$0.025 & 0.625$\pm$0.017 & \textbf{0.665$\pm$0.007} \\
& CTR-GCN & 0.587$\pm$0.047 & 0.510$\pm$0.054 & 0.768$\pm$0.077 & 0.613$\pm$0.062 & 0.581$\pm$0.043 \\
& PoseC3D & 0.553$\pm$0.086 & 0.522$\pm$0.145 & 0.583$\pm$0.270 & 0.509$\pm$0.118 & 0.538$\pm$0.077 \\
& TCNTE & 0.633$\pm$0.019 & 0.564$\pm$0.038 & 0.645$\pm$0.038 & 0.600$\pm$0.020 & 0.630$\pm$0.018 \\
& DynaFall & 0.616$\pm$0.057 & 0.541$\pm$0.059 & \textbf{0.772$\pm$0.115} & \textbf{0.631$\pm$0.048} & 0.607$\pm$0.060 \\
\bottomrule
\end{tabular}
\end{adjustbox}
\end{table}

\subsection{Ablation Study}
Table~\ref{tab:ablation} studies DynaFall-GCN components. Joint-only modeling is a reasonable baseline, but adding bone information improves URFD performance because bone vectors capture posture and body orientation. Adding explicit dynamics gives the most consistent clean-set gains: it improves F1 and macro-F1 relative to the joint-only and joint+bone variants on both datasets. Random dropout mainly increases recall on URFD, while confidence-aware dropout strongly increases MCFD recall but can reduce macro-F1. This supports a deployment-oriented interpretation: when missed falls are unacceptable, sensitivity-oriented training may be preferred, but threshold calibration is needed to control false alarms.

\begin{table}[t]
\centering
\caption{DynaFall-GCN ablation over three seeds. Precision, recall, and F1 are fall-class metrics.}
\label{tab:ablation}
\small
\begin{adjustbox}{width=\textwidth}
\begin{tabular}{llccccc}
\toprule
Dataset & Variant & Acc. & Prec. & Recall & F1 & Macro-F1 \\
\midrule
\multirow{5}{*}{URFD}
& A1 Joint & 0.666$\pm$0.040 & 0.394$\pm$0.050 & 0.731$\pm$0.226 & 0.507$\pm$0.098 & 0.623$\pm$0.030 \\
& A2 Joint+Bone & 0.748$\pm$0.115 & 0.563$\pm$0.192 & 0.658$\pm$0.253 & 0.558$\pm$0.118 & 0.687$\pm$0.098 \\
& A3 +Dynamics & \textbf{0.769$\pm$0.095} & \textbf{0.618$\pm$0.337} & 0.614$\pm$0.063 & \textbf{0.576$\pm$0.118} & \textbf{0.708$\pm$0.095} \\
& A4 +Random Dropout & 0.727$\pm$0.087 & 0.507$\pm$0.112 & \textbf{0.752$\pm$0.297} & 0.565$\pm$0.059 & 0.675$\pm$0.020 \\
& A5 +Confidence Dropout & 0.663$\pm$0.128 & 0.411$\pm$0.077 & 0.703$\pm$0.274 & 0.503$\pm$0.085 & 0.611$\pm$0.092 \\
\midrule
\multirow{5}{*}{MCFD}
& A1 Joint & 0.631$\pm$0.008 & 0.581$\pm$0.053 & 0.517$\pm$0.083 & 0.543$\pm$0.040 & 0.617$\pm$0.015 \\
& A2 Joint+Bone & 0.630$\pm$0.075 & 0.589$\pm$0.074 & 0.567$\pm$0.161 & 0.562$\pm$0.071 & 0.611$\pm$0.062 \\
& A3 +Dynamics & \textbf{0.649$\pm$0.013} & \textbf{0.591$\pm$0.045} & 0.605$\pm$0.061 & 0.595$\pm$0.016 & \textbf{0.642$\pm$0.007} \\
& A4 +Random Dropout & 0.605$\pm$0.025 & 0.536$\pm$0.033 & 0.566$\pm$0.059 & 0.550$\pm$0.034 & 0.598$\pm$0.025 \\
& A5 +Confidence Dropout & 0.597$\pm$0.106 & 0.537$\pm$0.118 & \textbf{0.708$\pm$0.030} & \textbf{0.604$\pm$0.071} & 0.594$\pm$0.109 \\
\bottomrule
\end{tabular}
\end{adjustbox}
\end{table}

\subsection{Runtime and Efficiency}
Fig.~\ref{fig:runtime} replaces the runtime table with a visual efficiency summary. All methods are lightweight compared with RGB video models because they operate on 32-frame skeleton clips. PoseC3D has the smallest parameter count in this compact implementation, but it also gives the lowest mean F1. DynaFall-GCN has 0.307M parameters, remains above 1000 clips/s on URFD and 4200 clips/s on MCFD, and achieves the best average fall-class F1 across the two datasets. This makes it a practical compromise between accuracy and efficiency.

\begin{figure}[t]
\centering
\resizebox{\textwidth}{!}{%
\begin{tikzpicture}[x=0.68cm,y=3.6cm]
\scriptsize
\draw[->] (-0.45,0) -- (6.85,0);
\draw[->] (-0.45,0) -- (-0.45,0.66);
\foreach \y/\lab in {0.0/0.000,0.2/0.200,0.4/0.400,0.6/0.600}
  \draw (-0.50,\y) -- (-0.40,\y) node[left=1pt] {\lab};
\foreach \name/\x/\f in {LSTM/0/0.568,ST-GCN/1/0.539,AGCN/2/0.590,CTR-GCN/3/0.590,PoseC3D/4/0.447,TCNTE/5/0.579,DynaFall/6/0.614} {
  \draw[fill=black!55] (\x-0.24,0) rectangle (\x+0.24,\f);
  \node[rotate=35, anchor=east] at (\x,-0.035) {\name};
}
\node[rotate=90] at (-1.05,0.33) {Mean fall-class F1};
\begin{scope}[xshift=6.8cm,x=0.68cm,y=0.82cm]
\draw[->] (-0.45,0) -- (6.85,0);
\draw[->] (-0.45,0) -- (-0.45,5.0);
\foreach \y/\lab in {1/1k,2/2k,3/3k,4/4k}
  \draw (-0.50,\y) -- (-0.40,\y) node[left=1pt] {\lab};
\foreach \name/\x/\u/\m in {LSTM/0/1.235/4.268,ST-GCN/1/1.126/4.249,AGCN/2/1.081/4.363,CTR-GCN/3/1.060/4.231,PoseC3D/4/1.023/3.978,TCNTE/5/1.086/4.334,DynaFall/6/1.024/4.285} {
  \draw[fill=black!25] (\x-0.22,0) rectangle (\x-0.03,\u);
  \draw[fill=black!65] (\x+0.03,0) rectangle (\x+0.22,\m);
  \node[rotate=35, anchor=east] at (\x,-0.24) {\name};
}
\node[fill=black!25, minimum width=0.30cm, minimum height=0.12cm] at (4.45,4.80) {};
\node[right] at (4.65,4.80) {URFD};
\node[fill=black!65, minimum width=0.30cm, minimum height=0.12cm] at (5.45,4.80) {};
\node[right] at (5.65,4.80) {MCFD};
\node[rotate=90] at (-1.05,2.5) {clips/s};
\end{scope}
\end{tikzpicture}
}
\caption{Runtime and efficiency summary. Left: mean fall-class F1 across URFD and MCFD. Right: inference throughput in clips/s.}
\label{fig:runtime}
\end{figure}

\section{Discussion}
The results show two operating regimes. On URFD, DynaFall-GCN performs best because pose quality is stable and explicit posture-transition cues are reliable. On MCFD, scenario-level splitting makes the problem harder: test clips come from unseen scenarios and viewpoints. AGCN therefore gives stronger macro-F1, while DynaFall-GCN still gives the highest fall recall. This is useful in safety monitoring, where a sensitivity-oriented model can be preferable if decision thresholds are calibrated to control false alarms.

The ablation in Table~\ref{tab:ablation} explains the model behavior. Bone features encode body configuration, while dynamics features encode fast downward motion and loss of upright support. Confidence-aware dropout is more nuanced: it does not always improve clean-set macro-F1, but it raises MCFD recall, making it suitable when missed-fall reduction is the priority.

\section{Conclusion}
This paper presented DynaFall-GCN, a skeleton-based fall detector that combines joint, bone, and fall-specific dynamics features with confidence-aware keypoint dropout. Under a unified RGB-to-skeleton protocol on URFD and MCFD, DynaFall-GCN achieves the best URFD fall-class F1 and macro-F1 and the highest MCFD fall recall, while remaining lightweight. Future work should test real non-staged falls, improve cross-view calibration, and integrate on-device pose estimation.

\begingroup
\footnotesize
\begin{thebibliography}{10}
\setlength{\itemsep}{0pt}
\setlength{\parsep}{0pt}
\bibitem{alam2022review}
Alam, E., Sufian, A., Dutta, P., Leo, M.: Vision-based human fall detection systems using deep learning: A review. Comput. Biol. Med. \textbf{146}, 105626 (2022)

\bibitem{urfd}
Kepski, M., Kwolek, B.: Human fall detection on embedded platform using depth maps and wireless accelerometer. Comput. Methods Programs Biomed. \textbf{117}(3), 489--501 (2014)

\bibitem{mcfd}
Auvinet, E., Multon, F., Saint-Arnaud, A., Rousseau, J., Meunier, J.: Fall detection using body geometry and human motion analysis. In: ICADI (2011)

\bibitem{yan2018stgcn}
Yan, S., Xiong, Y., Lin, D.: Spatial temporal graph convolutional networks for skeleton-based action recognition. In: AAAI (2018)

\bibitem{shi2019agcn}
Shi, L., Zhang, Y., Cheng, J., Lu, H.: Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In: CVPR (2019)

\bibitem{chen2021ctrgcn}
Chen, Y., Zhang, Z., Yuan, C., Li, B., Deng, Y., Hu, W.: Channel-wise topology refinement graph convolution for skeleton-based action recognition. In: ICCV (2021)

\bibitem{duan2022posec3d}
Duan, H., Zhao, Y., Chen, K., Shao, D., Lin, D., Dai, B.: Revisiting skeleton-based action recognition. In: CVPR (2022)

\bibitem{yu2025tcnte}
Yu, X., Wang, C., Wu, W., Xiong, S.: Real-time skeleton-based fall detection using temporal convolutional networks and transformer encoder. Pervasive Mob. Comput. \textbf{102}, 102016 (2025)

\bibitem{yolov8}
Ultralytics: YOLOv8. \url{https://github.com/ultralytics/ultralytics} (2023)

\bibitem{lin2017focal}
Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: ICCV, pp. 2980--2988 (2017)
\end{thebibliography}
\endgroup

\end{document}