Action Recognition in Law Enforcement: A Novel Dataset from Body
Worn Cameras
Sameer Hans
1
, Jean-Luc Dugelay
1
, Mohd Rizal Mohd Isa
2
and Mohammad Adib Khairuddin
2
1
EURECOM, 450 Route des Chappes, 06410 Biot, France
2
Universiti Pertahanan Nasional Malaysia (UPNM), Kem Perdana Sg. Besi, 57000 Kuala Lumpur, Malaysia
Keywords:
Body Worn Camera, Multimodal Dataset, Action Recognition, Surveillance.
Abstract:
Over the past decade, there has been a notable increase in the integration of body worn cameras (BWCs)
in many professional settings, particularly in law enforcement. BWCs serve as valuable tools for enhancing
transparency, accountability, and security by providing real-time, first-person perspective recordings of inter-
actions and events. These devices capture vast amounts of video data, which can offer critical insights into
the behaviors and actions of individuals in diverse scenarios. This paper aims to explore the intersection of
BWCs and action recognition methodologies. We introduce FALEBaction: a multimodal dataset for action
recognition using body worn cameras, with actions relevant to BWCs and law enforcement usage. We inves-
tigate the methodologies employed in extracting meaningful patterns from BWC footage, the effectiveness of
deep learning models in recognizing similar actions, and the potential applications and implications of these
advancements. By focusing on actions relevant to law enforcement scenarios, we ensure that our dataset meets
the practical needs of the authorities and researchers aiming to enhance public safety through advanced video
analysis technologies. The entire dataset can be obtained upon request from the authors to facilitate further
research in this domain.
1 INTRODUCTION
BWCs are becoming common in a variety of indus-
tries. They have been put into practice around the
globe, and they are an essential tool for law enforce-
ment to improve accountability (Suat Cubukcu and
Topalli, 2023), transparency (Choi et al., 2023), and
evidence gathering (Todak et al., 2024).
One of the most promising applications of BWC
data is in the field of action recognition, which fo-
cuses on the automated identification and classifica-
tion of human actions within video footage. Action
recognition in BWC footage can play a crucial role
in promoting transparency and accountability, ulti-
mately strengthening the trust between law enforce-
ment and the community. Advances in deep learn-
ing in recent times have greatly improved the perfor-
mance of action recognition systems. The ability to
extract spatial and temporal information from video
data has been significantly enhanced by models like
Convolutional Neural Networks (CNNs) (Tran et al.,
2015) and Recurrent Neural Networks (RNNs), par-
ticularly Long Short-Term Memory (LSTM) (Majd
and Safabakhsh, 2020) networks. These models can
learn complex patterns and dependencies, accurately
classifying a wide range of human actions.
Despite these advancements, action recognition
by BWCs remains a challenging task due to various
factors (Corso et al., 2018). The dynamic and often
unpredictable nature of body worn footage, character-
ized by varied perspectives, motion blur, and occlu-
sions, requires advanced algorithms capable of pro-
cessing and interpreting complex visual input. Ad-
ditionally, distinguishing between similar actions re-
quires models to be highly sensitive to subtle differ-
ences in motion and context.
This work introduces FALEBaction, a multi-
modal
1
dataset specific to the actions that are useful
for the applications of BWCs. This study is the first
to provide a publicly
2
available dataset with actions
specific to the usage of BWCs. It contains annotated
1
In the context of machine learning and data analysis,
a modality refers to a distinct type of data or information,
such as text, image, audio, video, sensor reading. A mul-
timodal dataset includes data from two or more of these
modalities.
2
To obtain the dataset, please visit
https://faleb.eurecom.fr/
Hans, S., Dugelay, J.-L., Isa, M. R. M. and Khairuddin, M. A.
Action Recognition in Law Enforcement: A Novel Dataset from Body Worn Cameras.
DOI: 10.5220/0013151900003905
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2025), pages 605-612
ISBN: 978-989-758-730-6; ISSN: 2184-4313
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
605

videos of 99 subjects with actions specific to law (di-
vided into 2 different scenarios), along with the meta-
data such as GPS position, and heart rate of the user.
We focus on the actions represented in 3.1, which are
useful for a law officer in real time. These actions
help in identifying if an officer is in a critical situation
(when the subject attacks and runs away) or when an
officer has made a significant step in their daily rou-
tine like making an arrest. We evaluate action recog-
nition models of C3D (Tran et al., 2015), I3D (Car-
reira and Zisserman, 2018), SlowFast network (Fe-
ichtenhofer et al., 2019), and TimeSformer (Berta-
sius et al., 2021), with an approach to improve the
accuracy in case of similar actions by following the
method of sequential fine-tuning (Chakraborty et al.,
2021) (a specialized form of transfer learning), along
with a comparative study on the two different scenar-
ios to evaluate the recognition performance in two dif-
ferent kinds of settings.
The paper is organized as follows. In section 2 we
survey related work on action recognition and BWCs.
In section 3, we introduce the steps followed in the
data collection for the activity. We report our exper-
imental setup and implementation in section 4. The
experiments and the results are presented in section
5. Finally, the conclusions and future work follow in
section 6.
2 RELATED WORKS
Existing action recognition datasets like HMDB51
(Kuehne et al., 2011), UCF101 (Soomro et al., 2012),
Something-something (Goyal et al., 2017), Kinetics
(Carreira and Zisserman, 2018), or EPIC-Kitchens
(Damen et al., 2020) offer diverse actions but lack the
specific context of BWC footage. While crowd scene
datasets like NWPU-Crowd (Wang et al., 2021) might
be relevant for scenarios with bystanders, they do not
capture the specific interactions between officers
3
and
suspects
4
.
The advancement of deep learning revolutionized
the field of action recognition by enabling end-to-end
learning of spatiotemporal features directly from raw
video data. CNNs were extended to process video
data, leading to the development of 3D CNNs such
as C3D. These models can capture both spatial and
temporal information by applying 3D convolutions
over the video frames (Duan et al., 2022). Table 1
shows performance of some of the models on existing
datasets.
3
Officer is the user of the camera.
4
The subject in question is the suspect.
Very limited work exists on action recognition by
BWCs. BWC footage presents unique challenges like
first-person viewpoint, low resolution due to camera
limitations, unbalanced data distribution across activ-
ities, privacy concerns over identifiable information,
and limited annotated training data. Moreover, the
existing work focuses on egocentric vision (N
´
u
˜
nez-
Marcos et al., 2022; Chen et al., 2019; Meng et al.,
2018) and not on plausible activities for action recog-
nition specific to BWCs like running, pushing, sitting
inside the car, and making an arrest.
The study (Chen et al., 2019) is based on ego-
activity recognition in first-person video in which
they propose a system for classifying ego-activities in
body worn video footage using handcrafted features
and a graph-based semi-supervised learning method.
They achieve comparable performance to supervised
methods on public datasets, however the challenges
include a lack of sufficient training data, and actions
specific to the usage of body cameras.
(Chao et al., 2022) presents a multimodal dataset
for human action detection that makes use of wearable
sensors and a depth camera to identify actions more
precisely. It has 880 sequences of 22 human acts car-
ried out by 5 participants. Based on depth images of
various actions, the recognition rate ranges from 58%
to 97%. Although the number of participants in this
dataset is extremely small, it is nonetheless helpful as
a starting point for the task of action recognition by
wearables.
In this paper, we introduce enough data for each
of the actions that are specific to the usage of BWCs
along with a sufficient number of subjects and a va-
riety of environmental conditions; which would be
helpful for law enforcement to assess the outcomes of
body camera footage. To the best of our knowledge,
no other study findings utilizing police BWCs in real
life scenarios consisting of relevant actions have been
published in the literature.
3 DATA COLLECTION
For the data collection, students from UPNM volun-
teered. The subjects were recorded using Cammpro
5
I826 Body camera. The recording took place over dif-
ferent sessions spread across a week. The camera was
fixed on the middle of the chest of the user (Bryan,
2020). All the recordings were done with a video res-
olution of 2304 × 1296 pixels at 30 fps.
The activity was recorded in an outdoor setting.
It was divided into 2 scenarios. In the first scenario,
5
https://www.cammpro.com/
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
606

Table 1: Existing Dataset Evaluations as discussed in Related Works.
Dataset Number of Classes Model Accuracy [%]
UCF101 101 C3D 85.2
UCF101 101 VideoSwin (Liu et al., 2022) 98
HMDB51 51 C3D 65.4
KINETICS-400 400 VideoSwin 82.7
Something-something 174 VideoSwin 69.6
we include the actions of walking, talking, showing
hands, sitting, going forward and backward, standing,
pushing, and running away. The second scenario has
the same actions as the first one with some additional
actions (an arrest is made instead of the subject run-
ning away). So, the additional actions comprise hands
behind the head, turning around, sitting inside the car,
and opening and closing the car doors. The subjects
were provided structured scripts on how to act for the
scenes. In total, we have 47 subjects (all male) for the
first scenario, and 52 subjects for the second scenario
(32 male and 20 female). Garmin
6
vivoactive 5 was
used as an additional sensor to record GPS data and
heart rate of the user. There will be fluctuations in
the heart rate and sudden changes in GPS when the
user chases the subject, which are useful parameters
for the other officers to know when the user runs sud-
denly. These additional attributes are also useful in
identifying an action more accurately.
After collecting all the videos, they are annotated
using the CVAT
7
tool according to their actions for
each video of a subject. We get XML files for each
video, which consist of information like the frame
number and the associated action label for that frame.
Therefore, for every subject’s video, we have 8 ac-
tions for the scene 1 and 11 actions for the scene 2.
3.1 Actions
The scenes depict the suspect’s actions. However,
we also mention about the officer in the scenes as
the camera is inherently dynamic (due to the officer’s
movements).
• Backward: The suspect moves backward.
• Forward: The suspect moves forward and the of-
ficer is moving backward.
• Show Hands: The suspect is raising hands in sur-
render.
• Sit: The suspect sits on the ground.
• Stand: The suspect is standing and talking.
• Walk: The suspect is walking toward the officer
and the officer also walks toward the suspect. For
6
https://www.garmin.com/
7
https://www.cvat.ai/
scene 2 only, there is an additional walk where the
officer and the suspect walk together towards the
car.
• Push: (Scene 1 Only) The suspect aggressively
pushes the officer away and the officer moves
backward.
• Run: (Scene 1 Only) The suspect is running away
and the officer is chasing the suspect.
• Door open: (Scene 2 Only) We see the car door
opening.
• Door close: (Scene 2 Only) We see the car door
closing.
• Hands behind head: (Scene 2 Only) The suspect
complies by placing both hands behind their head.
• Turn around: (Scene 2 Only) The officer makes
the suspect turn around (to make him walk to-
wards the car).
• Sit inside car: (Scene 2 Only) The officer makes
the suspect sit inside the car.
4 SETUP
In this section, we discuss the steps followed for pre-
processing of the videos, the networks used, and the
implementation details of the networks.
4.1 Preprocessing
Each annotated video is split according to its action
labels. The dataset contains 13 unique action cate-
gories, where for the first scene, we have 8 action cat-
egories and for the second scene, we have 11 action
categories. Fig. 1 shows some samples of the actions
present in the dataset. In total, we obtain 954 individ-
ual video clips showcasing an action.
We divide the training, validation, and test set in
the ratio of 65:15:20. As we split the videos accord-
ing to the actions (and not the subjects), a subject can
appear in both train and test set. Clips are resized to
have a frame size of 128 × 171. On training, we ran-
domly crop input clips into 16 × 112 × 112 crops for
Action Recognition in Law Enforcement: A Novel Dataset from Body Worn Cameras
607

(a) Show Hands. (b) Push. (c) Run.
(d) Hands behind head. (e) Sit inside car. (f) Door close.
Figure 1: Samples of action frames from the dataset. The first row represents specific actions of scene 1, and the second row
represents some actions of scene 2.
spatial and temporal jittering. We also horizontally
flip them with 50% probability.
Table 2 shows the total duration of each action and
the proportion of that action in their respective scenar-
ios.
4.2 Implementation Details
We used C3D, I3D, SlowFast network, and TimeS-
former models for our experiments. These models
were chosen for their respective strengths: C3D as a
baseline model, I3D for its popularity and proven per-
formance, SlowFast Network for its advancements in
the field, and TimeSformer for its novelty.
• C3D: We implement a pretrained C3D model,
trained on the Sports-1M dataset (Karpathy et al.,
2014), which consists of 1.1 million sports videos
belonging to one of the 487 sports categories.
This pretrained model is fine-tuned on both the
scene datasets to evaluate the performance of the
model and gain some insights on the videos by
BWCs. Training is done by SGD optimizer. The
learning rate is fixed as 0.001 after various ex-
periments. The initial layers of the model are
frozen, and we add fc8 layer to match the num-
ber of classes in the new dataset. This layer starts
with random weights and is trained from scratch.
The optimization is stopped after 50 epochs.
• I3D: We experiment with I3D architecture pre-
trained on Kinetics-400 dataset. This model is
used to initialize our network, where we replace
the final projection layer to match the number
of output classes corresponding to the actions in
the dataset. The model was trained using the
CrossEntropy loss function, and optimized using
the Adam optimizer with a learning rate of 0.001.
During training, both training and validation met-
rics, including loss, accuracy, precision, recall,
and F1-score, were tracked to evaluate perfor-
mance. We also ensured that each input video was
processed as a stack of frames, allowing the I3D
model to leverage its 3D convolutional layers to
capture temporal dynamics, improving its ability
to recognize actions across time.
• SlowFast Network: The SlowFast network op-
erates by processing video inputs through two dis-
tinct pathways: the slow pathway, which samples
frames at a lower frame rate to capture long-range
temporal patterns, and the fast pathway, which
processes higher frame-rate sequences to capture
finer motion details. In our implementation, the
SlowFast model is pretrained on Kinetics-400.
The final fully connected layer of the network is
replaced with a new layer corresponding to the
number of actions in the dataset. To accommodate
both fast and slow temporal dynamics, the video
frames are split into two pathways, with the slow
pathway subsampling every fourth frame, while
the fast pathway uses all the frames. The train-
ing process uses cross-entropy loss to minimize
classification error, with an Adam optimizer tuned
with a learning rate of 0.001.
• TimeSformer: The TimeSformer is a deep learn-
ing architecture designed specifically for video
understanding tasks like action recognition. It
applies transformers directly to the video’s spa-
tial and temporal dimensions. It processes video
frames as a sequence of patches, incorporating at-
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
608

Table 2: Class Proportion in FALEBaction.
Action Scene 1 Scene 2
Labels Duration Proportion Duration Proportion
[seconds] [%] [seconds] [%]
Backward 101.40 3.98 122.40 2.84
Forward 81.50 3.20 98.67 2.29
Show Hands 165.53 6.50 410.53 9.51
Sit 361.80 14.20 428.77 9.93
Stand 663.67 26.04 656.60 15.21
Walk 696.70 27.34 1422.37 32.96
Push 66.07 2.59 - -
Run 411.77 16.16 - -
Door close - - 307.70 7.13
Door open - - 189.60 4.39
Hands behind head - - 131.17 3.04
Sit inside car - - 260.10 6.03
Turn around - - 288.10 6.68
Total 2548.43 100 4316.00 100
tention mechanisms across both time and space,
allowing for more efficient and scalable learning
of video features. In our implementation, the
model is pretrained on Kinetics-400. We fine-tune
the model by modifying the classifier head. Dur-
ing training, the model’s performance is evaluated
on validation data after each epoch, and its per-
formance is seen on test data every 5 epochs to
monitor accuracy and loss, aiming to improve the
classification performance.
5 EXPERIMENTS
5.1 Scene 1 Analysis
In the first experiment (E1), we fine-tune the models
directly on the actions of scene 1. We have 231, 64,
and 80 videos for train, validation, and test sets re-
spectively. We receive high test accuracies of 95.7%,
81.25%, 81% ,and 76.25% for TimeSformer, C3D,
I3D, and SlowFast network respectively. However,
we observe that most errors occur in similar actions,
like moving backward and forward. As the camera
is not fixed and moves along with the user, this cre-
ates confusion for the model. For example, when the
suspect is moving forward, the officer moves slightly
backward and when the suspect is moving backward,
the officer remains still.
For the second experiment (E2), we group the sim-
ilar actions and see the performance of the models.
So, ”backward” and ”forward” are grouped as a sin-
gle action (Motion). In total, we now have 7 actions
for this scene. In this case, we have 203 videos in the
train set, 56 videos in the validation set, and 70 videos
in the test set. While fewer actions are considered,
there is a substantial improvement in test accuracies
as compared to experiment 1.
On evaluating the confusion matrix, the models
are seen performing poorly on similar actions (back-
ward and forward), and also because they have a very
low proportion in the scene 1 actions as compared
to others (3.98% and 3.20% for backward and for-
ward respectively). To potentially improve the perfor-
mance, for the third experiment (E3), we fine-tune the
model in two sequential phases, first on just the con-
fusing ”backward” and ”forward” actions, and then
on the full set of 8 actions. This two-phase approach
mimics a hierarchical learning process, where the
model initially concentrates on differentiating subtle
distinctions between closely related actions, and then
expands its knowledge to the remaining classes in the
second phase. This approach shows improvement in
the performance of the model, particularly in the case
of similar actions. We obtain accuracies of 88.75%,
and 86.25% for C3D and I3D models respectively.
There is a significant increase in the accuracy as com-
pared to the first experiment when we follow this ap-
proach. For the TimeSformer model, this optimiza-
tion was not very relevant and had similar results as
the first experiment.
Table 3 shows the results of the 3 approaches dis-
cussed above.
5.2 Scene 2 Analysis
Similar to scene 1 experiments, we carry out the ex-
periments for the scene 2 actions. For the first exper-
iment (T1), there are 362 videos in the train set, 97
Action Recognition in Law Enforcement: A Novel Dataset from Body Worn Cameras
609

Table 3: Scene 1 Experiments.
Models Test Accuracy [%]
E1 E2 E3
(8 (7 (8
actions) actions) actions)
C3D 81.25 92.19 88.75
I3D 81.00 90.00 86.25
SlowFast 76.25 78.75 76.00
TimeSformer 95.70 - -
videos in the validation set, and 120 videos in the test
set. There is a decrease in the accuracy as compared to
previous scene tests. We receive accuracies of 88.5%,
80.83%, 66.67%, and 60% for TimeSformer, C3D,
I3D, and SlowFast respectively. The confusion ma-
trix for this experiment shows that the models pro-
duce the most errors in the actions of backward, and
forward (confused as ”walk”), and show hands, and
hands behind head. The first three actions (backward,
forward, and walk) are very similar to each other and
the last two (show hands and hands behind head) are
identical and difficult to differentiate.
In the second experiment (T2), we merge the ac-
tions of ”backward” and ”forward” into a single ac-
tion (Motion) and the actions of ”show hands” and
”hands behind head” into a single action (Hands). We
now have 9 actions for this scene, where there are 303
videos in train set, 81 videos in validation set, and 99
videos in test set. When we fine-tune the models on
the actions, again there is a significant increase in the
test accuracy as compared to the first experiment.
In the final test (T3), we follow the approach of
sequential fine-tuning again, where we first fine-tune
the model on the confusing actions only (backward,
forward, walk, show hands, and hands behind head),
and then fine-tune this new model on the entire 11
actions for this scene dataset. We see a notable im-
provement in the performance, especially when com-
paring actions that are similar. We are able to improve
the test accuracy to 88.33% as compared with the first
experiment for the C3D model.
Table 4 shows the results of the 3 approaches dis-
cussed above.
Table 4: Scene 2 Experiments.
Models Test Accuracy [%]
T1 T2 T3
(11 (9 (11
actions) actions) actions)
C3D 80.83 89.90 88.33
I3D 66.67 74.75 68.44
SlowFast 60.00 70.54 63.64
TimeSformer 88.50 - -
5.3 Cross-Scene Analysis
It is essential to ensure that our model can gener-
alize well across different environments. In both
scenes, we have some common actions between them
(Backward, Forward, Show Hands, Sit, Stand, Walk).
For this experiment, we experimented with C3D and
TimeSformer models as they had the best perfor-
mance in previous experiments. The training and val-
idation sets are made from one scene setting and the
test set from different setting to evaluate the model’s
ability to generalize to unseen settings for the same
actions, which is crucial for real-world deployment.
Table 5 shows the performance of the models across
different scene environments on the test set. The first
test is training on scene 1 environment and testing
on scene 2 environment. For C3D model, there is a
big drop in the test accuracy which becomes 75.76%.
When the training set is scene 2 environment, we see a
significant improvement in the accuracy value (85%).
The model generalizes better from scene 2 to scene
1 (85%) as compared to scene 1 and scene 2. TimeS-
former model performs slightly better than C3D (with
similar behavior of generalizing better from scene 2
to scene 1). Although this experiment had less num-
ber of actions as compared to both scenes, the models
had lower accuracy as compared to both scene analy-
sis. Scene 2 has more variation and longer durations
for certain actions like walking and showing hands.
This increased variation leads the model to learn more
robust features, which helps in better generalizing to
new scenes.
Table 5: Cross-Scene Experiments.
Experiment Test Accuracy [%]
C3D TimeSformer
Train Scene 1
Test Scene 2
75.76 80.00
Train Scene 2
Test Scene 1
85.00 86.36
6 CONCLUSION
This work introduces FALEBaction, a multimodal an-
notated dataset for action recognition using BWCs,
which is the first of its kind that is publicly available
and based on actions relevant to the daily usage of
law enforcement scenarios. The dataset is created by
only using BWCs for images and videos, and an addi-
tional sensor to record metadata (GPS and heart rate
of the user). This dataset addresses a critical gap in
the current research landscape, providing comprehen-
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
610

sive and detailed annotations for actions such as mak-
ing an arrest, attacks on officers, and suspects flee-
ing, which are integral to an officers’ daily duties.
Although fine-tuning the model on the 2 scenes pro-
duces good results, we see improvement in the recog-
nition performance following the approach of trans-
fer learning. For cross-scene tests, the model gener-
alizes well in the case of training from scene 2 and
testing on scene 1 due to more variations and longer
durations for the actions. TimeSformer outperformed
the traditional models such as C3D, I3D, and Slow-
Fast in recognizing complex law enforcement-related
actions. Future efforts should focus on tailoring the
models to better handle variations between environ-
ments to enhance cross-scene action recognition. Ad-
ditionally, custom architectures or domain-adaptive
layers could be introduced to better capture the con-
textual details of complex law enforcement scenarios,
making the model more robust and capable of gener-
alizing across different environments and action dy-
namics. Apart from action recognition with BWCs,
the application of BWCs for face recognition has also
been explored only to a limited extent, and there is
tremendous potential for advancements in this area.
ACKNOWLEDGEMENTS
This research work is a collaborative study between
UPNM-EURECOM and Naval Group.
REFERENCES
Bertasius, G., Wang, H., and Torresani, L. (2021). Is space-
time attention all you need for video understanding?
In Proceedings of the International Conference on
Machine Learning (ICML).
Bryan, J. (2020). Effects of Movement on Biometric Fa-
cial Recognition in Body-Worn Cameras. PhD thesis,
Purdue University Graduate School.
Carreira, J. and Zisserman, A. (2018). Quo vadis, action
recognition? a new model and the kinetics dataset.
In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition (CVPR).
Chakraborty, S., Mondal, R., Singh, P., Sarkar, R., and
Bhattacharjee, D. (2021). Transfer learning with fine
tuning for human action recognition from still images.
Multimedia Tools and Applications, 80.
Chao, X., Hou, Z., and Mo, Y. (2022). Czu-mhad: A multi-
modal dataset for human action recognition utilizing a
depth camera and 10 wearable inertial sensors. IEEE
Sensors Journal, 22:1–1.
Chen, H., Li, H., Song, A., Haberland, M., Akar, O.,
Dhillon, A., Zhou, T., Bertozzi, A. L., and Branting-
ham, P. J. (2019). Semi-supervised first-person activ-
ity recognition in body-worn video. arXiv preprint
arXiv:1904.09062.
Choi, S., Michalski, N. D., and Snyder, J. A. (2023). The
“civilizing” effect of body-worn cameras on police-
civilian interactions: Examining the current evidence,
potential moderators, and methodological limitations.
Criminal Justice Review, 48(1):21–47.
Corso, J. J., Alahi, A., Grauman, K., Hager, G. D., Morency,
L.-P., Sawhney, H., and Sheikh, Y. (2018). Video anal-
ysis for body-worn cameras in law enforcement. arXiv
preprint arXiv:1604.03130.
Damen, D., Doughty, H., Farinella, G., Fidler, S., Furnari,
A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T.,
Price, W., and Wray, M. (2020). The epic-kitchens
dataset: Collection, challenges and baselines. IEEE
Transactions on Pattern Analysis and Machine Intel-
ligence, PP:1–1.
Duan, H., Zhao, Y., Chen, K., Lin, D., and Dai, B. (2022).
Revisiting skeleton-based action recognition. In 2022
IEEE/CVF Conference on Computer Vision and Pat-
tern Recognition (CVPR), pages 2959–2968.
Feichtenhofer, C., Fan, H., Malik, J., and He, K. (2019).
Slowfast networks for video recognition. In 2019
IEEE/CVF International Conference on Computer Vi-
sion (ICCV), pages 6201–6210.
Goyal, R., Kahou, S. E., Michalski, V., Materzynska, J.,
Westphal, S., Kim, H., Haenel, V., Fruend, I., Yian-
ilos, P., Mueller-Freitag, M., Hoppe, F., Thurau, C.,
Bax, I., and Memisevic, R. (2017). The “something
something” video database for learning and evaluat-
ing visual common sense. pages 5843–5851.
Karpathy, A., Toderici, G., Shetty, S., Leung, T., Suk-
thankar, R., and Fei-Fei, L. (2014). Large-scale video
classification with convolutional neural networks. In
CVPR.
Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., and Serre,
T. (2011). HMDB: a large video database for human
motion recognition. In Proceedings of the Interna-
tional Conference on Computer Vision (ICCV).
Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., and
Hu, H. (2022). Video swin transformer. In 2022
IEEE/CVF Conference on Computer Vision and Pat-
tern Recognition (CVPR), pages 3192–3201.
Majd, M. and Safabakhsh, R. (2020). Correlational con-
volutional lstm for human action recognition. Neuro-
computing, 396:224–229.
Meng, Z., S
´
anchez, J., Morel, J.-M., Bertozzi, A. L., and
Brantingham, P. J. (2018). Ego-motion classification
for body-worn videos. In Tai, X.-C., Bae, E., and
Lysaker, M., editors, Imaging, Vision and Learning
Based on Optimization and PDEs, pages 221–239,
Cham. Springer International Publishing.
N
´
u
˜
nez-Marcos, A., Azkune, G., and Arganda-Carreras, I.
(2022). Egocentric vision-based action recognition:
A survey. Neurocomputing, 472:175–197.
Soomro, K., Zamir, A., and Shah, M. (2012). Ucf101: A
dataset of 101 human actions classes from videos in
the wild. CoRR.
Suat Cubukcu, Nusret Sahin, E. T. and Topalli, V. (2023).
The effect of body-worn cameras on the adjudication
Action Recognition in Law Enforcement: A Novel Dataset from Body Worn Cameras
611

of citizen complaints of police misconduct. Justice
Quarterly, 40(7):999–1023.
Todak, N., Gaub, J. E., and White, M. D. (2024). Testing the
evidentiary value of police body-worn cameras in mis-
demeanor court. Crime & Delinquency, 70(4):1249–
1273.
Tran, D., Bourdev, L., Fergus, R., Torresani, L., and Paluri,
M. (2015). Learning spatiotemporal features with 3d
convolutional networks. In 2015 IEEE International
Conference on Computer Vision (ICCV), pages 4489–
4497.
Wang, Q., Gao, J., Lin, W., and Li, X. (2021). Nwpu-crowd:
A large-scale benchmark for crowd counting and lo-
calization. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 43(06):2141–2149.
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
612