Distracted driving is a leading cause of road accidents globally. Identification of distracted driving involves reliably detecting and classifying various forms of driver distraction (e.g., texting, eating, or using in-car devices) from in-vehicle camera feeds to enhance road safety. This task is challenging due to the need for robust models that can generalize to a diverse set of driver behaviors without requiring extensive annotated datasets. In this paper, we propose KiD3, a novel method for distracted driver detection (DDD) by infusing auxiliary knowledge about semantic relations between entities in a scene and the structural configuration of the driver's pose. Specifically, we construct a unified framework that integrates the scene graphs, and driver's pose information with the visual cues in video frames to create a holistic representation of the driver's actions. Our results indicate that KiD3 achieves a 13.64% accuracy improvement over the vision-only baseline by incorporating such auxiliary knowledge with visual information. The source code for KiD3 is available at: https://github.com/ishwarbb/KiD3.
Distracted driving is a leading cause of road accidents globally, posing significant challenges to road safety. According to the National Highway Trafic Safety Administration (NHTSA)1 approximately 3,308 people lost their lives in the United States in 2022 due to distracted driving, and nearly 290,000 people were injured. Almost 20% of those killed in distracted driving-related crashes were pedestrians, cyclists, and others outside the vehicle. In addition to the loss of lives and injuries, the financial burden from distracted driving crashes collectively amounts to $98 billion in 2019 alone, highlighting the urgency of developing efective detection methods. ing and computer vision techniques, including, but not limited to, object detection, pose estimation, and action recognition. On the other hand, recent advancements in knowledge infusion [1] and Neurosymbolic AI [2] provide new opportunities for challenging tasks in scene understanding [3, 4, 5] and context understanding [6]. Hence, we posit that there is valuable auxiliary knowledge that can be either computed/ derived from the visual inputs. Specifically, we hypothesize that by infusing such knowledge with current computer vision models would improve the overall detection capabilities and robustness while not requiring the heavy computation demands of ultra-high parameter models.
The task of identifying distracted driving involves re- method for distracted driver detection that infuses auxliably detecting and classifying various forms of driver iliary knowledge about inherent semantic relations bedistraction, such as texting, eating, or using other ob- tween entities in a scene and the structural configuration jects/devices from in-vehicle camera feeds. This task is of the driver’s pose. Specifically, we construct a unified challenging due to the need for robust models that can framework that integrates scene graphs and the driver’s generalize to a diverse set of driver behaviors without pose information with visual information to enhance requiring extensive annotated datasets. Traditionally, the the model’s understanding of distraction behaviors (see DDD task has been solved using various end-to-end learn- Figure 1).
our results indicate that incorporating such auxiliary knowledge with visual information significantly improves detection accuracy. KiD3 achieves a 13.64% accuracy improvement over the vision-only baseline, demonstrating the efectiveness of integrating semantic and cient and scalable solution that does not demand the use of expensive high-parameter models.
Contributions of this paper are as follows: action recognition model on each view and taking the average probability over all the views as the final output.
The outputs are then post-processed for predicting the action label and temporal localization of the predicted 1. A novel, simple method for distracted driver de- action. This work utilizes the X3D family of networks tection that incorporates the auxiliary knowl- [9] for video classification instead of relying on manual edge computed/estimated with vision inputs with- feature engineering. Wei Zhou et al. [10] improve upon out the need for high-parameter, computational this work by fine-tuning large pre-trained models instead heavy models. of training from scratch and by empirically selecting spe2. A demonstration of the efectiveness of infus- cific camera views for specific distracted action classes. ing diferent types of auxiliary knowledge over vision-only baselines using real-world distracted driving data.
one of 2 tasks: Action Recognition/Classification and Temporal Action Localization (TAL). Action recognition is a computer vision task that involves classifying a given image or a video into a set of pre-defined set of actions or classes. TAL, on the other hand detects activities being performed in a video streams and outputs start and end timestamps. In this paper, we focus on solving the action recognition task by classifying frames into various distracted driver activities. Here, we explore related work considering two directions: (1) methods for distracted driver identification and (2) methods for generating/encoding semantic graphs from visual scenes.
Point-Based Driver Activity Recognition that extracts static and movement-based features from driver pose and facial features and trains a frame classification model for action recognition. Then, a merge procedure is used to identify robust activity segments while ignoring outlier frame activity predictions.
chronization across videos by training an ensemble 3D Previous works mainly focus on the use of sophisticated post-processing algorithms, use of larger encoder-decoder architectures and multi-view synchronization to improve action recognition and TAL performance. In contrast, our work aims to improve classification performance by incorporating auxiliary knowledge (e.g., semantic entities/relationships of a frame, pose information) that can be derived and infused as graphs into the encoder side of our architecture. Next, we will explore the state-of-the-art methods for scene graph generation.
tomatically mapping an image or a video into a semantic structural scene graph, which requires the correct labeling of detected objects and their relationships [11]. Yuren Cong et al. [12] pose SGG as a set prediction problem. They propose an end-to-end SGG model, RelTR, with an encoder-decoder architecture. In contrast to most existing scene graph generation methods, such as Neural Motif, VCTree, and Graph R-CNN, [13, 14, 15] which RelTR used as benchmarks, RelTR is a one-stage method that predicts sparse scene graphs directly only using visual appearance without combining entities and labeling all possible predicates. Due to its simplicity, eficiency and SOTA performance, we selected RelTR to generate SGGs for our experiments.
from the objects detected in the scene and the estimated pose of the driver.
Given a video frame x ∈ R× × 3 sampled from a video where denotes the height of the frame, denotes the width of the frame, and 3 corresponds to the color channels (RGB), the learning objective is to classify it into one of 18 predefined activities = {1, 2, . . . , 18}. ˆ = arg max∈ .
We define a classifier model : R× × 3 → [0, 1]18 that maps a video frame to a probability distribution over the 18 activities. Specifically, (x) = p, where p = [1, 2, . . . , 18] and represents the probability that the frame x belongs to class , such that ∑︀1=8 1 = 1 and 0 ≤ ≤ 1 ∀ ∈ {1, . . . , 18}. The predicted class ˆ for the frame x can therefore be determined by:
The real-world datasets for distracted driver identification typically include annotated video sequences from cameras mounted inside the vehicle. While several open datasets are available, such as StateFarmDataset2, we have selected SynDDv1 [17] to be used for experiments due to the higher number of distracted behavior classes and the diversity, including variations in lighting conditions, driver appearances, and the use of objects and
driver-detection people in the background. SynDDv1 consists of 30 video clips in the training set and 30 videos in the test set. The dataset consists of images collected using three in-vehicle cameras positioned at locations: on the dashboard, near the rear-view mirror, and on the top right-side window corner, as shown in Table 1 and Figure 1. The video sequences are sampled at 30 frames per second at a resolution of 1920×1080 and are manually synchronized for the three camera views. Each video is approximately 10 minutes long and contains all 18 distracted activities shown in Table 2. The driver executed these activities with or without an appearance block, such as a hat or sunglasses, in random order for a random duration. There are six videos for each driver: three videos with an appearance block and three videos without any appearance block.
sulting in 10 videos for training and 10 videos for testing. Sets of (frame, label) were created by sampling frames from the videos at regular intervals and obtaining the corresponding labels from the annotations. The publicly available dataset contains various inconsistencies in the annotation format provided as CSV files. These inconsistencies, such as diferent naming conventions, variations in capitalization, and extra spaces in names, have been resolved to ensure consistency across all data splits.
Pose Estimator
Pose Information Scene Graph Generator
Scene Graph
GCN Graph
Encoder
Graph Encoding
Linear Classifier
Label 3.4. Image Encoding image embedding vector. The rationale for discarding the last 2 layers is that the final layer reduces the dimen3.4.1. Background sionality to only 18, which is insuficient for our needs. To classify video frames into one of the predefined activ- Additionally, the earlier layers capture more general feaities, the first step is to obtain robust image embeddings tures, which are beneficial for transfer learning. These that would efectively capture the visual features in raw embeddings are then used for further processing and pixel data into a more manageable and informative rep- classification tasks. resentation. Possible methods for this transformation include using pre-trained Convolutional Neural Networks 3.5. Scene Graph Generation and (CNNs) like VGGNet [18], ResNet [19], or Inception [20]. Encoding Out of these methods, we selected VGG16, a variant of VGGNet, due to its simplicity and efectiveness in ex- 3.5.1. Background tracting deep features from images. VGG16 has been extensively used and validated in various image classifi- Scene graphs structurally represent the relationships becation tasks, making it a reliable choice for our purpose. tween various objects in a given image. Each node in the graph represents an object, while edges denote the relationships between these objects; for example consider the 3.4.2. Technical Details triple: “« man holding phone »”. Scene graphs capture VGGNet, particularly VGG16, is a deep convolutional the high-level contextual and semantic information of the network known for its simple yet efective architecture, scene, going beyond pixel-level data. They are also essenconsisting of 16 weight layers. The network is struc- tial for scene understanding and reasoning and allow us tured with multiple convolutional layers followed by fully to explicitly inject knowledge into the pipeline. For examconnected layers. Each convolutional layer uses small ple, considering DDD task, a scene graph containing the receptive fields (3x3) and applies multiple filters to ex- triple “« person drinking_from bottle »” might indicate tract features at diferent levels of abstraction. The fully distracted driving activity. Modeling such important relaconnected layers then process these features for classifi- tions can otherwise be achieved implicitly using methods cation. VGG16’s design focuses on depth and simplicity, such as convolutional-network-based image encoders, making it an ideal candidate for transfer learning. with some uncertainty. 3.4.3. Pre-processing and Adaptation
obtain image embeddings. Specifically, we discarded the last 2 classifier layers of the pre-trained VGG16 model and retained the base model along with the first 4 classifier layers. This configuration results in a 4096-dimensional 3.5.2. Technical Details To generate the scene graph for a given frame, we use the RelTr architecture [12]. Then, we use a Graph Convolutional Network (GCN) [21] layer followed by a ℎ activation to obtain representations for each node in the graph. We take the mean of all the node embeddings to for classi cation. VGG16’s design focuses on depth and simplicity, making it an ideal candidate for transfer learning. art 2D pose estimation model, to extract pose information. OpenPose can detect and output a set of key points corresponding to various body parts, such as the head, shoulders, elbows, and hands.
These key points are represented as coordinates in a 2D space. The process involves detecting the spatial locations of these joints and constructing a pose structure that re ects the driver’s body conguration. Mathematically, each key point can be represented as: k8 = (G8, ~8 ) where k8 denotes the 8-th key point with G8 and ~8 being its coordinates in the image frame. 3.4.3 Pre-processing and Adaptation. To adapt VGG16 for our task, we ne-tuned the model to obtain image embeddings. Speci cally, we discarded the last 2 classi er layers of the pre-trained VGG16 model and retained the base model along with the rst 4 classi er layers. This con guration results in a 4096-dimensional image embedding vector. The rationale for discarding the last 2 layers is that tohebtanianl laaygerrarepdhu-cleesvtehel rdeimpernessieonnatalittyiotonoannlyd1t8r,ewahticthhiissivnseucft-or acsietnhtefogr roauprhneeendsc.oAddidnigti.onally, the earlier layers capture more general features, which are bene cial for transfer learning. These embeddings are then used for further processing and classi cation t3as.k5s..3. Pre-processing and Adaptation 33..67.3. UPren-pirfoiceesdsinPgiapndeAldinapetation. To adapt the pose estimation data for our task, we pre-processed the key point coordinates Wobetacinoendstfrroumct OapseinmPposlee.mThaechkienyep-loeianrtsnwinegrepnipoermlinaleizetod caondmbstirnuectuthreed ltaotceonntsiestnenctolydirnepgrsesoenfttthhee darbivoevr’es pmosoed.ules. Each moAddudlietiotnaaklleys,waendiemrivaegdefeaastuirneps usutcahnads tphreodcisetsanseces biettiwneteona A scene graph output from RelTr [12] is in the form of mtheeahnanindsgafnudl evyeecst/foarcer,ethperaensgelnetfaortmioend.bWytehetheyeens wciothncthaetennecakt,e tt33wtih.o.r5e52ine.1pgsaghlreraeiSBpetptscsahancebokrnoelegftidprewsotretheueGseosneenrdfana.fnvrtopSaesdrrclhaimaeornntuGieioso(sbegnojtrnbeshajcepeetrch,tarwss,etsrhi=litneiaorlluetaan(ictgeotiduianov1grnne,aebn,dlrsle,yidmtEerwanen2gpeo)certeoe.)ewnsE.dtehahintEceenhthsregrtsenhemeloeand.tr1teieoiTliaananhl--nliyds, tsmteabhihnnofaeeythdtisadlnteetnrhhic(sverieitfeenderdpigi’pnessrtttspehaeasuccnoettetcfeimnveditiomthtbuadieeisatestsiilwgn.o’psgeeni.aeYpsbnAOeiulLtllihigstOnyieone[trh9og.ia]ta)nha.cdmTcfsuhearee1ansdtedse-uflfoeyocbacirtjnueiwtncreetacrssprtllwdriyeketMeroaeaunLctdprPluhicnctolioaenasslescfiltoofahyrrsesfhoiprsmbaettwiesencothnevseeorbtejedctst;ofoar elxisatmoplfe ecodngseidse,rwthheetrrieplee:d"g«emsanare hroeldpirnegspehnotneed»a".sScpeaniersgroapfhnsocadpetsu.reTthheishiigsh-plervoevl icdoendtexttouatlhe Algorithm 1 KiD3 Pipeline aGndCsNemeannctoicdienrfotromoatbiotanionf athgersacpenhe-,lgeovienlgrbeepyroensdepnitxaetli-olenve.l Require: Training Dataset, a collection of images and labels. data. They are also essential for scene understanding and reasoning for 8<064, ;014; in Training Dataset do and allow us to explicitly inject knowledge into the pipeline. For E8BD0;⇢=2>38=6 <064⇢=2>34A(8<064) e3xa.6m.pleP,coonsseideErinsgtiDmDDattaisko, nascene graph containing the triple B6⇢=2>38=6 (24=4⌧A0?⌘">3D;4 (8<064) "« person drinking_from bottle »" might indicate distracted driv- ?>B4 40CDA4B %>B4 =5 >A<0C8>=">3D;4(8<064) i3ng.6a.c1t.iviBtya.cMkogderloinugnsduch important relations can otherwise be abPcahoseisedeviemedsaitgmiempelianctcitioloydneurssii,snwgaimtcherstiohtmiocdeaslusnucccoehmrtaaspincootnyn.evnoltutiinonuanl-ndeetrwsotrakn-d- ;2>>6=82C0BC4=0(C>435C<0G[E(8"BD!0%;⇢(2=>2=>2308=C46=; 0B6C4⇢3=)2)>38=6; ?>B4 40CDA4B] ing the spatial configuration of a subject’s body, which ;>BB ⇠A>BB⇢=CA>?~(;>68CB, ;014;) 3i.n5.2thiTsecchansiecailsDtehtaeilsd.rTivoegre.nBeryatecathpetuscreinneggtrhapehpfoorsiatigoivnesnof fkraemye,bwoedyusepathretsR, eplTorsaercehsitteimctuarteio[2n].pTrhoevni,dweesuvsealauGabralephin- ;>BB.BackPropagate() ù Propagate errors to the linear CfoonrmvoaluttiioonnalaNbeotuwtortkhe(GdCrNi)v[e5r]’slaypeorsftoullroeweadnbdy ma)o0v=e⌘macetni-ts. classi er and GCNs vation to obtain representations for each node in the graph. We end for tTakheisthienmfoeramnoaftaiollnthiesneosdseeenmtbiaelddfionrgsatcocoubrtaatinelaygcralapshs-liefvyeilng rtehperedsernivtaetrio’sn aancdtivtrietaitesth.iVsvaerciotoursasmtheethgoradpshceanncobdeinegm.ployed for pose estimation, including 2D and 3D approaches.
We opted to use a state-of-the-art 2D pose estimation technique to efectively capture the required spatial data. 3.7.1. Training 3.6.2. Technical Details We utilized OpenPose [22], a state-of-the-art 2D pose estimation model, to extract pose information. OpenPose can detect and output a set of key points corresponding to various body parts, such as the head, shoulders, elbows, and hands. These key points are represented as coordinates in a 2D space. The process involves detecting the spatial locations of these joints and constructing a pose structure that reflects the driver’s body configura- 4. Experiments tion. Mathematically, each key point can be represented as: k = (, ) where k denotes the -th key point We outline the following experimental setup to evaluate with and being its coordinates in the image frame. the proposed approach’s overall performance and the contribution of each sub-component.
distracted driver classification task to obtain task-suitable embeddings. During training, we freeze the Image Encoding and Pose Information modules and only train the linear classifier and the GCN graph encoder in the Scene Graph Encoding module. We use activation in the final layer of the feed-forward MLP and use the Cross-Entropy loss function. 3.6.3. Pre-processing and Adaptation To adapt the pose estimation data for our task, we pre- 4.1. Method 1 - Vision Only processed the key point coordinates obtained from Open- In the first experiment, we utilized existing computer viPose. The key points were normalized and structured to sion (CV) models to establish a baseline performance consistently represent the driver’s pose. for the frame classification task. We fine-tuned the
Additionally, we derived features such as the distance VGG-16 model to assess the performance of traditional between the hands and eyes/face, the angle formed by CV models. To achieve this, we froze the weights of the eyes with the neck, and the distance between the the entire model and unfroze only the classification hands and objects like a phone or bottle (if detected using layers (model.classifier[1...6]). The sixth classification YthOeLmOo[d2e3l]’)s. aTbhielistey fteoataucrceusrawteerlye cinrutecriparleftoranendhcalanscsiinfgy layer nn.Linear(4096, 1000) was replaced with the driver’s activities. intny.cLlaisnseesa.rT(h4e0m96o,difie1d8)mtoodeml awtcahs tthheennfinume-btuenreodf oacntiv
our classification task, allowing the classification layers to adapt to the specific features of our dataset.
Table 2 summarizes the results of our experiments on the 4.2. Method 2 - Vision + Scene Graphs test set and the ablation studies across diferent method variations. We evaluate the performance using two metIn the second experiment, we use the VGG-16 similar to rics: accuracy and the F1 score. The vision-only model how it was used in Method 1; however, out of the last achieves 79.64 overall accuracy and 0.81 F1 score, respecsix classifier layers, we discarded the last two layers and tively. With the inclusion of scene graphs, the accuracy used the base model with the first four classifier layers and the F1 score increased by 11.88% and 9.88%, respecto obtain a 4096-dimensional image embedding vector. tively. Finally, the complete model incorporating both The rationale is that the final layer could not be utilized scene graphs and pose information achieves the peak because it reduces the image embedding to only 18 di- performance of 90.5% accuracy and 0.91 F1 score, respecmensions, which is insuficient for capturing the rich tively. features needed for our task. Moreover, earlier layers in the network capture more general features beneficial for transfer learning. Then, we integrate image embeddings with scene graphs encoded using a Graph Convolutional Network (GCN) [21]. The embeddings derived from the GCN are concatenated with the image embeddings obtained from the VGG-16 model. Linear layers are used as a head to combine these information streams, forming a unified representation. This combined model was trained on the same classification objective, leveraging both the visual and relational features present in the data.
In the final experiment, we further enrich the scene representation by incorporating pose information, enhancing its ability to understand the driver’s activities. The pose details included the location of objects via bounding boxes and the outline of the human skeleton with coordinates of key points such as the eyes, nose, and ifsts. We engineered additional features based on external knowledge, including the distance between the hand and face and the distance between the hand and a phone or bottle (if detected using YOLO [23]). These engineered features were added to the concatenation of image embeddings and scene graph embeddings. The model is then re-trained on the classification task with these additional features, providing a holistic understanding of the driver’s activities.
are particularly efective in identifying classes such as Eating (class 5), Adjusting Control Panel (class 10), and Singing with Music (class 17). We interpret this as evidence that our approach successfully incorporates auxil- References iary knowledge, enhancing our model’s performance for these classes. [1] A. Sheth, M. Gaur, U. Kursuncu, R. Wickramarachchi, Shades of knowledge-infused learning for enhancing deep learning, IEEE Internet Com6. Discussion puting 23 (2019) 54–63. doi:10.1109/MIC.2019. 2960071.
Our results clearly support the initial hypothesis that [2] A. Sheth, K. Roy, M. Gaur, Neurosymbolic artificial the inclusion of valuable auxiliary knowledge with vi- intelligence (why, what, and how), IEEE Intelligent sual features would enhance the performance of the DDD Systems 38 (2023) 56–62. doi:10.1109/MIS.2023. task. The ablation study further establishes each auxiliary 3268724. knowledge type’s role in the overall performance. Scene [3] R. Wickramarachchi, C. Henson, A. Sheth, graphs provided the most significant auxiliary knowl- Knowledge-infused Learning for Entity Prediction edge, highlighting the importance of explicitly encoding in Driving Scenes, Frontiers in Big Data 4 (2021) semantic information and infusing it with visual features. 759110. doi:10.3389/fdata.2021.759110. By incorporating pose information of driver actions, we [4] R. Wickramarachchi, C. Henson, A. Sheth, were able to further enrich overall accuracy and robust- Knowledge-based entity prediction for improved ness. However, several limitations to our approach war- machine perception in autonomous systems, rant further investigation. IEEE Intelligent Systems (2022). doi:10.1109/MIS. 2022.3181015. 6.1. Limitations [5] R. Wickramarachchi, C. Henson, A. Sheth, Clue-ad: A context-based method for labeling unobserved One limitation is the reliance on annotated data for train- entities in autonomous driving data, Proceedings of ing. While we used a combination of supervised and un- the AAAI Conference on Artificial Intelligence 37 supervised learning techniques to mitigate this issue, the (2023) 16491–16493. URL: https://ojs.aaai.org/index. availability of annotated data remains a key constraint. php/AAAI/article/view/27089. doi:10.1609/aaai. Additionally, our method may struggle with complex and v37i13.27089. highly variable driving scenarios where the relationships [6] A. Oltramari, J. Francis, C. Henson, K. Ma, R. Wickbetween objects and actions are less clear. Finally, we ramarachchi, Neuro-symbolic architectures for conhave not considered using foundation models like Vi- text understanding, in: Knowledge Graphs for eXsion Language Models (VLMs) for our experiments. Our plainable Artificial Intelligence: Foundations, Apmain focus in this work is to evaluate the impact of aux- plications and Challenges, IOS Press, 2020, pp. 143– iliary knowledge on the DDD task without the need for 160. complex, high-parameter models. [7] A. Vats, D. C. Anastasiu, Key point-based driver activity recognition, in: 2022 IEEE/CVF Confer7. Conclusions and Future Work ence on Computer Vision and Pattern Recognition Workshops (CVPRW), 2022.
In this paper, we proposed a novel, simple approach to [8] M. T. Tran, M. Quan Vu, N. D. Hoang, K.-H. distracted driver detection by infusing two types of aux- Nam Bui, An efective temporal localization iliary knowledge with visual information. Our method method with multi-view 3d action recognition for leverages scene graphs and estimated pose information untrimmed naturalistic driving videos, in: 2022 with visual embeddings to comprehensively represent IEEE/CVF Conference on Computer Vision and driver actions. Our experimental results showcase the ef- Pattern Recognition Workshops (CVPRW), 2022, fectiveness of infusing each type of auxiliary knowledge pp. 3167–3172. doi:10.1109/CVPRW56347.2022. with visual features to achieve 90.5% peak performance 00357. on the DDD task. [9] C. Feichtenhofer, X3D: expanding architec
Future work will address the limitations mentioned tures for eficient video recognition, CoRR above, such as the reliance on annotated data and the abs/2004.04730 (2020). URL: https://arxiv.org/abs/ handling of complex driving scenarios. Additionally, we 2004.04730. arXiv:2004.04730. plan to explore the integration of other types of knowl- [10] W. Zhou, Y. Qian, Z. Jie, L. Ma, Multi view action edge representations, such as temporal graphs, to further recognition for distracted driver behavior localenhance the performance of distracted driver detection ization, 2023. doi:10.1109/CVPRW59228.2023. systems Further, we plan to investigate the role of VLMs 00567. in this task. [11] G. Zhu, L. Zhang, Y. Jiang, Y. Dang, H. Hou, P. Shen, M. Feng, X. Zhao, Q. Miao, S. A. A. Shah, M. Bennamoun, Scene graph generation: A comprehensive survey, 2022. arXiv:2201.00443. [12] Y. Cong, M. Y. Yang, B. Rosenhahn, Reltr: Relation transformer for scene graph generation, 2023.
arXiv:2201.11460. [13] R. Zellers, M. Yatskar, S. Thomson, Y. Choi, Neural motifs: Scene graph parsing with global context, in: Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), 2018. [14] K. Tang, H. Zhang, B. Wu, W. Luo, W. Liu, Learning to compose dynamic tree structures for visual contexts, CoRR abs/1812.01880 (2018). URL: http: //arxiv.org/abs/1812.01880. arXiv:1812.01880. [15] J. Yang, J. Lu, S. Lee, D. Batra, D. Parikh, Graph r-cnn for scene graph generation, in: Proceedings of the European Conference on Computer Vision (ECCV), 2018. [16] P. Ping, C. Huang, W. Ding, Y. Liu, M. Chiyomi, T. Kazuya, Distracted driving detection based on the fusion of deep learning and causal reasoning, Information Fusion 89 (2023) 121– 142. URL: https://www.sciencedirect.com/science/ article/pii/S1566253522001014. doi:https://doi.
org/10.1016/j.inffus.2022.08.009. [17] M. S. Rahman, A. Venkatachalapathy, A. Sharma,
J. Wang, S. V. Gursoy, D. Anastasiu, S. Wang, Synthetic distracted driving (syndd1) dataset for analyzing distracted behaviors and various gaze zones of a driver, Data in Brief 46 (2023) 108793. doi:https: //doi.org/10.1016/j.dib.2022.108793. [18] K. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale image recognition, arXiv preprint arXiv:1409.1556 (2014). [19] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), 2016. [20] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed,
D. Anguelov, D. Erhan, V. Vanhoucke, A. Rabinovich, Going deeper with convolutions, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. [21] T. N. Kipf, M. Welling, Semi-supervised classification with graph convolutional networks, 2017.
arXiv:1609.02907. [22] Z. Cao, T. Simon, S.-E. Wei, Y. Sheikh, Realtime multi-person 2d pose estimation using part afinity ifelds, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. [23] J. Redmon, S. Divvala, R. Girshick, A. Farhadi, You only look once: Unified, real-time object detection, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.