Human Fall Detection in Poor Lighting Conditions Using CNN-Based
Model
Md. Sabir Hossain
1 a
and Md. Mahfuzur Rahman
1, 2 b
1
Department of Information and Computer Science, King Fahd University of Petroleum & Minerals (KFUPM),
Dhahran, 31261, Saudi Arabia
2
Interdisciplinary Research Center for Intelligent Secure Systems (IRC-ISS),
King Fahd University of Petroleum & Minerals (KFUPM), Dhahran, 31261, Saudi Arabia
Keywords:
Elderly Care, Fall Detection, CNN, Low Lighting, Deep Learning.
Abstract:
Human fall detection for elderly care has become a crucial field of research as it can cause serious injuries and
impact the quality of life. In this article, we present a deep learning-based approach for human fall detection
in low-lighting conditions using a convolutional neural network (CNN). We trained and evaluated our model
on multiple datasets, both annotated for fall detection. The proposed architecture captures and analyzes the
falls-related features effectively, even in achieving a significant amount of precision, recall, and F1-scores for
human fall detection. Moreover, our proposed architecture outperforms (91% accuracy) several state-of-the-art
models, including ResNet50, InceptionV3, MobileNet, XceptionNet, VGG16, VGG19, and DenseNet. With a
reliable human fall detection architecture, this research significantly contributes to enhancing safety measures
for elderly individuals.
1 INTRODUCTION
Elderly people fall is a burning issue as it often leads
to serious health injuries, even death. World Health
Organization (WHO) reported that deaths among peo-
ple aged ≥65 are because of fall-related injuries
(Ageing and (AAH), 2008). According to the report,
approximately 28-35% of people aged ≥65 fall each
year, but increasing to 32-42% for those aged ≥70.
These phenomena affect individual’s health as well as
create challenges for healthcare domain. Again, ac-
cording to World Bank Data, the total amount of aged
(over 65) population is 10%. This large amount of
aged people all over the world puts an immense need
to address fall-related health issues. Frontier health
industries can play an important role in human fall
detection and leverage opportunities to take timely ac-
tion. Early fall detection can reduce health injuries in
a great context and improve the quality of living for
elderly people.
Existing fall detection techniques depend on sen-
sors such as accelerometers and gyroscopes (Chen
et al., 2022; Lian et al., 2021; Gomes et al., 2022).
a
https://orcid.org/0000-0003-4545-6872
b
https://orcid.org/0000-0002-2871-9119
These approaches generally ask the users to use dif-
ferent wearables to collect data. Though these tech-
niques achieved some success but lack in many cases.
Many users are not comfortable wearing device con-
stantly. Moreover, in dynamic environments, sensor-
based technologies perform very poorly. In this re-
search, we aim to develop a vision-based human fall
detection model using deep learning methods. Our
approach uses neural networks to analyze the visual
data to learn the fall-related features. We aim to im-
prove the accuracy of the fall detection model as well
as enhance its reliability. Additionally, we also intend
to deal with low-light environment challenges while
capturing images. Our final objective is to develop
an accurate and robust human fall detection system to
provide reliable healthcare services.
The rest of the paper is organized as follows. The
section 2 discusses the related research on human
fall detection. Our proposed CNN-based architecture
with overall methodology is described in section 3.
After presenting the results and comparisons with
other existing state-of-the arts models in section 4, we
conclude the paper in section 5. Additionally, future
research directions are added in the section 5.
414
Hossain, M. S. and Rahman, M. M.
Human Fall Detection in Poor Lighting Conditions Using CNN-Based Model.
DOI: 10.5220/0013502200003938
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 11th International Conference on Information and Communication Technologies for Ageing Well and e-Health (ICT4AWE 2025), pages 414-420
ISBN: 978-989-758-743-6; ISSN: 2184-4984
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

2 LITERATURE REVIEW
Elderly fall detection has drawn significant atten-
tion from the research community. It has made an
enormous impact on health and quality of life. Re-
searchers explored the use of sensors, wearable de-
vices, and cutting-edge technologies like machine
learning, computer vision, etc. to detect human falls.
This literature review focuses on the application of
deep learning techniques in fall detection among el-
derly people.
Alam et al. presented a comprehensive review
on vision-based human fall detection systems (Alam
et al., 2022). They classified existing techniques
into various deep learning models, including CNN,
LSTM, auto-encoders, MLP, and hybrid techniques.
Moreover, they described architectures and evalua-
tion metrics such as accuracy, sensitivity, and speci-
ficity. This study also analyzed different benchmark
datasets and measured performances of different tech-
niques on them. They also identified different limita-
tions such as lack of real fall data, privacy-preserving
issue, and detection in low-lighting or occlusions.
B. Luo proposed elderly fall detection for smart
home environments using Yolo networks (Bo, 2023).
The author demonstrates less memory usage and su-
perior accuracy (95%) with the Yolov5 network over
other networks. The proposed method is sensitive to
the camera field of view (FOV) for accurate detection.
The author suggested using other sensors with vision
sensors to improve the existing accuracy. Addition-
ally, the author suggested the integration of other sen-
sors with vision sensors further enhance the accuracy
of fall detection systems. X. Zi et al. also offered a de-
tection technique in poor lighting condition scenario
(Zi et al., 2023).
In contrast, X. Kan et al. laid a lightweight ap-
proach named CGNS-YOLO for human fall detection
by integrating the GSConv module and the GDCN
module with YOLOv5 network (Kan et al., 2023).
They also reduced the proposed model’s size and dis-
carded less pertinent information by incorporating a
normalization-based attention module (NAM). They
achieved 1.2% enhancement in detection accuracy
compared with the conventional YOLOv5s frame-
work. This paper also considered challenging envi-
ronments like different lighting conditions, and oc-
clusions in their research. However, this research re-
quires validation regarding the different lighting con-
ditions of fall detection.
Gunale et al. presented a novel way of using
CNN to detect falls to assist elderly people (Gunale
et al., 2023). They combined multiple datasets to
achieve generalization and used CNN for automatic
feature extraction. They performed both qualitative
and quantitative analysis and received 97.93% accu-
racy. However, the sensitivity value for combined
datasets (URFD, MCFD, FDD and SDU) is very low
(64.46%) compared to other state-of-the-art architec-
ture, indicating a performance constraint. Their pre-
dictive model also suffered from the scarcity of appro-
priate data to predict correctly. A few recent studies
also focused on using deep learning networks in hu-
man fall detection (Hoang et al., 2023; Alanazi and
Muhammad, 2022). Adri
´
an N
´
u
˜
nez-Marcos and Ig-
nacio Arganda-Carreras proposed a video fall detec-
tion system using transformer-based model (N
´
u
˜
nez-
Marcos and Arganda-Carreras, 2024). Their sug-
gested model determines whether or not a fall has oc-
curred based on a video clip. It uses a sliding window
style in a video stream to sound an alarm as soon as it
detects a fall.
In brief, the above articles discussed the applica-
tion of YOLO, Transformer, and Convolutional Neu-
ral Networks to detect human falls for elderly care in
different challenging environment. However, further
research is required to address the challenges and im-
prove real world fall detection performance.
3 METHODOLOGY
The proposed methodology for detecting human falls
in low lighting conditions consists of several steps;
pre-processing, model training, and prediction using
our CNN model and other pre-trained models [Fig-
ure 1]. Initially, the images are annotated with the
bounding box to locate the human subject. Later,
the images are resized to a standard dimension to
make them consistent throughout the dataset. For
model training, we have extracted the illumination-
invariant features using our proposed CNN architec-
ture. Then, the trained CNN model predicts based
on the test image to classify as a fall or non-fall sce-
nario. Furthermore, we have compared our CNN
model’s performance with state-of-the-art (SOTA)
pre-trained models trained on ImageNet namely
ResNet50 (He et al., 2016), InceptionV3 (Szegedy
et al., 2016), MobileNet (Howard et al., 2017), Xcep-
tionNet (Chollet, 2017), VGG16 (Simonyan and Zis-
serman, 2015), VGG19 (Simonyan and Zisserman,
2015) and DenseNet (Huang et al., 2017) to ensure
optimal detection capacity in low-lighting environ-
ments. By designing CNN-based architecture adap-
tive to different low-lighting conditions, this proposed
approach provides a robust and reliable human fall de-
tection and contributes to the safety and independent
living of elderly people.
Human Fall Detection in Poor Lighting Conditions Using CNN-Based Model
415

Figure 1: Proposed Methodology for Human Fall Detection.
Figure 2: Image samples from a) Fe2i Dataset and, b) Fall
Detection Dataset.
3.1 Data Collection and Annotation
In experimentation, we have used two datasets
namely Fe2i Fall Detection Dataset (Xing et al.,
2023) and Fall Detection dataset (Kandagatla, 2022).
Fe2i dataset is an annotated version of the original
video dataset, and contains 2996 sample images, each
annotated with relevant information for fall detection.
The original Le2i incorporated multiple scenarios like
living room, coffee room, office, and classroom. The
scenarios were recorded in different lighting condi-
tions. Each image annotation contains class labels
and the bounding box coordinates. The annotations
contain two classes: fall and upright. The annota-
tions were in XML format. We have converted it into
text format (.txt) to prepare for training and testing the
CNN model.
On the other hand, the Fall Detection dataset con-
sists of 374 images for training and 111 images for
validation. The images were labeled using the Make
Sense website (MakeSense.AI, 2024). After upload-
ing the images, a bounding box needs to be drawn.
Then, a class label from three (fall, sit, and walk)
is assigned. The class labels with bounding box co-
ordinates were exported as text files for each image.
The annotations process provides information about
the activities as well as enables training and testing
processes. Samples from both datasets are shown in
Figure 2.
3.2 Proposed CNN Architecture
Our proposed CNN architecture captures and ana-
lyzes features related to falls effectively including low
lighting conditions. The architecture includes several
layers that contribute to detecting fall detection accu-
rately. The proposed CNN architecture is depicted in
Figure 3. A brief description of each component and
its functionalities provided below:
The input layer takes the images as input of size
150 x 15 pixels with three color channels (Red, Green,
and Blue). The previously defined fixed size confirms
the compatibility with other layers of the model. In
the first convolutional layer, 32 filters with kernel size
IS4WB_SC 2025 - Special Session on Innovative Strategies to Enhance Older Adults’ Well-being and Social Connections
416

Figure 3: Proposed CNN Architecture to Detect Human Fall in Low Lighting Conditions.
3x3 are applied to the preprocessed images. The tar-
get of this component is to identify the crucial fea-
tures prevalent in images like edges and textures to
understand human shapes and movements. We have
introduced the ReLU activation function to add non-
linearity to the model to learn complex patterns in the
data. We used batch normalization after each convo-
lutional layer to normalize the activations. It results in
fast and more stable training of the network. This step
is also effective for varying characteristic images due
to low lighting conditions as it decreases the internal
covariate shift.
The first max pooling layers after convolutional
layers minimize feature maps by reducing spatial di-
mensions. However, this step keeps the most signifi-
cant information. As a result, the computational com-
plexity decreases, and the model becomes more ro-
bust to variations in input. Higher-level features are
learned in the subsequent convolutional layers from
the feature maps generated in immediate layers. In
these layers, the more abstract representations of the
input images are extracted which are significant for
detecting human falls in diverse environments and
lighting conditions. At the end of the architecture,
dense layers are introduced to learn complex patterns
from the flattened feature maps. There are 512 neu-
rons in dense layers and the ReLU activation function
is applied to capture complicated relationships in the
data. A dropout layer after dense layers prevents over-
fitting by randomly deactivating a fraction of neurons
during training. It improves model generalization to
unseen data and extends detection accuracy. The sin-
gle neuron with a sigmoid activation function in the
final layer generates a probability score between 0 and
1. A closer value to 1 indicates the likelihood of de-
tecting a human fall. We have provided layer-wise de-
tailed structure in Table 1 to confirm reproducibility.
It will facilitate the researchers to benchmark and fur-
ther enhance its performance for human fall detection
in real-world applications.
In brief, our proposed model is capable of detect-
ing human falls in low-lighting conditions effectively
by extracting and analyzing visual features from input
images. Several layers including convolutional, batch
normalization, activation functions, and dropout reg-
ularization are introduced so that the model learns
and generalizes to different environments to develop
a suitable real-world fall detection technique.
Table 1: Layer-wise details of the proposed CNN model.
Layer Kernel/Units Out. Shape Parameters
Input 150×150×3 150×150×3 -
Conv2D 3×3, 32 filters 148×148×32 896
BN - 148×148×32 128
ReLU - 148×148×32 0
MaxPool 2×2 74×74×32 0
Conv2D 3×3, 64 filters 72×72×64 18,496
BN - 72×72×64 256
ReLU - 72×72×64 0
MaxPool 2×2 36×36×64 0
Flatten - 82944 0
Dense 512 neurons 512 42,467,840
BN - 512 2,048
ReLU - 512 0
Dropout 0.5 512 0
Output 1 neuron (Sigmoid) 1 513
Figure 4: Model loss for each model for different epochs.
Human Fall Detection in Poor Lighting Conditions Using CNN-Based Model
417

Figure 5: Model accuracy for each model for different
epochs.
4 RESULTS AND ANALYSIS
4.1 Training Loss and Accuracy Results
The training loss and accuracy statistics for different
models laid an idea about the superiority of our pro-
posed model. For both cases, our proposed model
shows faster convergence [Figure 4] and higher ac-
curacy [Figure 5] compared to state-of-the-art models
like ResNet50, InceptionV3, MobileNet, Xception,
VGG16, VGG19, and DenseNet.
Figure 4 represents that our proposed CNN model
efficiently minimizes classification errors. The higher
loss values for InceptionV3 (115.27) and Xception-
Net (133.25) in the initial epoch indicate that these
models did not learn the relevant features for falling
detection. Furthermore, ResNet50, VGG16, and
DenseNet demonstrated significantly higher loss val-
ues compared to our proposed CNN model. Although
these models finally reduced their losses over time,
they are still higher. On the other hand, our proposed
model gradually reached a final loss of 0.0397, in-
dicating its ability to learn fall-related features effi-
ciently, optimize parameter updates effectively, and
generalize the training data. Thus, our proposed
model can detect human fall under low-light condi-
tions.
The training accuracy depicted in Figure 5
also supports the effectiveness of the proposed
CNN model in fall detection. Few models like
ResNet50 (0.986), VGG16 (0.993), and VGG19
(0.992) achieved significant amounts accuracy in later
epochs but didn’t achieve highest accuracy (1.00).
However, other models, InceptionV3 and Xception-
Net did not reach optimal accuracy levels, with
0.7707 and 0.7118, respectively. In contrast, our pro-
posed model achieved maximum accuracy showing
it’s effectiveness in human fall detection in challeng-
ing situation like low lighting.
4.2 Model Performance
The performance of our proposed deep learning-
based model for human fall detection in low lighting
conditions is presented in Table 2. We use precision,
recall, and F1-score as evaluation metrics to assess the
effectiveness of our model.
Our model achieved a precision score of 0.90 for
detecting class 0 (fall) and 0.93 for class 1 (non-fall).
Precision indicates the percentage of true positive de-
tection over total positively identified instances in-
cluding true positives and false positives. With the
high precision score, false positive detection is min-
imized. Our proposed model’s precision score indi-
cates that it can detect falls and non-falls with min-
imal errors, making it reliable for practical applica-
tions.
Additionally, the recall values of our proposed
model are also significant, 0.88 for class 0 and 0.94
for class 1. Recall value is measured by the ratio of
the true positive detection out of all actual positive
instances (true positive and false negative). The re-
call value is very crucial to detect the actual falls and
non-falls to ensure the safety of elderly people. Our
proposed model achieved high recall scores which in-
dicates the effectiveness of our model in detecting
the majority of falls, even in challenging low-lighting
conditions.
The F1-score presents a balanced measure of the
model’s performance by calculating the harmonic
mean of precision and recall. Our model performed in
a balanced manner and achieved F1-scores of 0.89 for
class 0 and 0.94 for class 1. The high F1-scores are
an indication of the model’s overall effectiveness in
accurately and reliably detecting human falls. More-
over, our model achieved 0.92 overall accuracy, which
indicates that the model correctly identifies falls and
non-falls.
In brief, the performance measurements of our
proposed deep learning-based model show its effec-
tiveness in human fall detection, even in low-lighting
conditions. The high precision, recall, F1-scores, and
overall accuracy denote that our model provides a
reliable solution for enhancing the safety and well-
being of elderly individuals.
Table 2: Precision, recall, and F1-score for each class.
Class Precision Recall F1-Score
0 0.90 0.88 0.89
1 0.93 0.94 0.94
Accuracy 0.92 (111 samples)
Macro Avg 0.91 0.91 0.91
Weighted Avg 0.92 0.92 0.92
IS4WB_SC 2025 - Special Session on Innovative Strategies to Enhance Older Adults’ Well-being and Social Connections
418

4.3 Comparative Analysis
The Table 3 presents the comparative analysis of var-
ious state-of-the-art models used in human fall detec-
tion. Our CNN-based proposed model outperforms
other notable models listed in the table, achieving an
accuracy of 0.9099.
ResNet50 is one of the promising models in com-
puter vision, achieved 0.8834 accuracy. ResNet50 is a
popular option for many image processing tasks, spe-
cially classification, due to its deep architecture and
residual connections. Though it’s a robust model,
it performed approximately 2.65% less compared to
our proposed model accuracy. Another widely used
model, InceptionV3, performed the poorest among
the compared models with an accuracy of 0.4577. In
spite of its efficiency and accuracy in various appli-
cations, this poor performance indicates that Incep-
tionV3 might not be well-suited for human fall detec-
tion, especially in low-lighting conditions.
MobileNet, another popular model designed for
mobile and embedded vision applications, achieved
an accuracy of 0.7460. Due to the presence of low
lighting conditions, it is compared to other models
(ResNet50, VGG16, VGG19, and DenseNet) espe-
cially, 16% less than our proposed model. Xcep-
tionNet, a model created to improve efficiency by
expanding on the Inception module achieved an ac-
curacy of 0.6469. While its unique design is com-
mendable the results suggest that XceptionNet may
not excel much in this task when compared to our
suggested method. Interestingly, VGG16, VGG19,
and DenseNet, these three models achieved an identi-
cal accuracy of 0.8434. These models are capable to
capture image details. Though these models outper-
formed than InceptionV3 MobileNet and Xception-
Net, they still fall short of our proposed model by,
approximately 6.65%.
The results demonstrated that the CNN-based ar-
chitecture we proposed performed better than all other
state-of-the-art models tested. The higher accuracy
rate (0.9099) denotes its effectiveness in detecting
human falls, in challenging low-light settings. The
exceptional performance is achieved due to the cus-
tomized design and fine-tuning of the CNN network,
which probably improves its capability to recognize
and understand the characteristics of falls.
5 CONCLUSIONS
In this research, we have developed an accurate
and reliable CNN-based human fall detection model
to enhance elderly care through early and reliable
Table 3: Comparison with the state-of-the-art models with
the proposed model.
Model Accuracy
ResNet50 0.8834
InceptionV3 0.4577
MobileNet 0.7460
XceptionNet 0.6469
VGG16 0.8434
VGG19 0.8434
DenseNet 0.8434
Proposed Network 0.9099
fall detection. By achieving 91% overall accuracy
our proposed model outperforms most state-of-the-art
(SOTA) models. Furthermore, our proposed CNN-
based customized model performs accurately in dif-
ferent challenging situations including low lighting.
Our proposed CNN-based model is a notable contri-
bution in the advancement in the area of elderly care.
Detecting human falls at the earliest possible time
may avoid severe injuries and contribute to the life
of elderly individuals, leveraging a safe environment
at living places.
Though our proposed model has competing re-
sults, there are still a few areas where researchers can
contribute in the future. One of them is validating the
model in in diverse environments and varying lighting
conditions. In future research, we can consider amal-
gamation of deep learning-based models with other
sensors, such as accelerometers and gyroscopes. It
will improve the accuracy and reliability of fall de-
tection systems. This diverse method may leverage
more comprehensive data which will lead to better de-
tection capabilities. Another future research direction
might be real-world deployment and continuous im-
provement based on user feedback. As fall detection
in real-world scenarios is crucial, it can be refined by
constant monitoring and providing input from real-
world deployments. By contributing to the above-
mentioned areas, we can improve the accuracy, reli-
ability, and applicability of the fall detection model.
Finally, the enhanced model can be an indispensable
tool in elderly care and other safety-critical applica-
tions.
ACKNOWLEDGMENT
The authors would like to thank Dr. Saeed Anwar
for his comments on the project presentation which
improved the quality of the manuscript.
Human Fall Detection in Poor Lighting Conditions Using CNN-Based Model
419

6 FUNDING STATEMENT
This research was fully funded by the Interdisci-
plinary Research Center for Intelligent Secure Sys-
tems (IRC-ISS) through the Deanship of Research,
King Fahd University of Petroleum and Minerals,
Dhahran 31261, Saudi Arabia. The IRC-ISS research
grant #INSS2516 supported the required data and
computing resources for this experiment.
7 DATA AND CODE
AVAILABILITY
Both datasets used in this study are publicly available
on Kaggle, provided in (Xing et al., 2023), (Kanda-
gatla, 2022). The code for this study is available from
the authors upon reasonable request.
REFERENCES
Ageing and (AAH), H. (2008). Who global report on
falls prevention in older age. https://www.who.int/
publications/i/item/9789241563536. Accessed: Mar.
03, 2024.
Alam, E., Sufian, A., Dutta, P., and Leo, M. (2022). Vision-
based human fall detection systems using deep learn-
ing: A review. Computers in biology and medicine,
146:105626.
Alanazi, T. and Muhammad, G. (2022). Human fall detec-
tion using 3d multi-stream convolutional neural net-
works with fusion. Diagnostics, 12(12).
Bo, L. (2023). Human fall detection for smart home car-
ing using yolo networks. International Journal of Ad-
vanced Computer Science and Applications, 14(4).
Chen, Z., Wang, Y., and Yang, W. (2022). Video based fall
detection using human poses. In CCF Conference on
Big Data, pages 283–296. Springer.
Chollet, F. (2017). Xception: Deep learning with depthwise
separable convolutions. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recogni-
tion (CVPR), pages 1251–1258.
Gomes, M. E. N., Mac
ˆ
edo, D., Zanchettin, C., de Mattos-
Neto, P. S. G., and Oliveira, A. (2022). Multi-human
fall detection and localization in videos. Computer
Vision and Image Understanding, 220.
Gunale, K. G., Mukherji, P., and Motade, S. N. (2023). Con-
volutional neural network-based fall detection for the
elderly person monitoring. Journal of Advances in In-
formation Technology, 14(6):1169–1176.
He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-
ual learning for image recognition. In Proceedings of
the IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), pages 770–778.
Hoang, V. H., Lee, J. W., Piran, M. J., and Park, C. S.
(2023). Advances in skeleton-based fall detection in
rgb videos: From handcrafted to deep learning ap-
proaches. IEEE Access, 11:92322–92352.
Howard, A. G., Zhu, M., Chen, B., Kalenichenko, D.,
Wang, W., Weyand, T., Andreetto, M., and Adam,
H. (2017). Mobilenets: Efficient convolutional neu-
ral networks for mobile vision applications. arXiv
preprint arXiv:1704.04861.
Huang, G., Liu, Z., Maaten, L. V. D., and Weinberger, K. Q.
(2017). Densely connected convolutional networks.
In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), pages 4700–
4708.
Kan, X., Zhu, S., Zhang, Y., and Qian, C. (2023). A
lightweight human fall detection network. Sensors,
23(22).
Kandagatla, U. K. (2022). Fall detection dataset. https:
//www.kaggle.com/datasets/uttejkumarkandagatla/
fall-detection-dataset. Accessed: Dec. 20, 2024.
Lian, J., Yuan, X., Li, M., and Tzeng, N. F. (2021). Fall
detection via inaudible acoustic sensing. Proc ACM
Interact Mob Wearable Ubiquitous Technol, 5(3).
MakeSense.AI (2024). Makesense.ai - free annotation tool
for ai and machine learning. https://www.makesense.
ai/. Accessed: 2024-12-20.
N
´
u
˜
nez-Marcos, A. and Arganda-Carreras, I. (2024).
Transformer-based fall detection in videos. En-
gineering Applications of Artificial Intelligence,
132:107937.
Simonyan, K. and Zisserman, A. (2015). Very deep con-
volutional networks for large-scale image recogni-
tion. International Conference on Learning Represen-
tations (ICLR).
Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., and Wo-
jna, Z. (2016). Rethinking the inception architecture
for computer vision. In Proceedings of the IEEE Con-
ference on Computer Vision and Pattern Recognition
(CVPR), pages 2818–2826.
Xing, Z., Chaturvedi, K., Braytee, A., Li, J., and
Prasad, M. (2023). Fe2i fall detection dataset
(annotated). https://www.kaggle.com/datasets/starxz/
fe2i-fall-detection-datasetannotated. Accessed: Dec.
21, 2024.
Zi, X., Chaturvedi, K., Braytee, A., Li, J., and Prasad, M.
(2023). Detecting human falls in poor lighting: Ob-
ject detection and tracking approach for indoor safety.
Electronics (Switzerland), 12(5).
IS4WB_SC 2025 - Special Session on Innovative Strategies to Enhance Older Adults’ Well-being and Social Connections
420