 ICACT20210194 Slide.01
[Big Slide]
|
Chrome  Click!! |
 |
Hello everyone, my name is Ikromjanov Kobiljon Komil Ugli. For ICACT 2021, I have coauthored with four co-authors.The title of our presentation is "A Transfer Learning Approach for Identification of Distracted Driving". |
| ICACT20210194 Slide.02
[Big Slide]
|
Chrome Click!! |
 |
The agenda of the presentation is mentioned in this slide |
| ICACT20210194 Slide.03
[Big Slide]
|
Chrome Click!! |
 |
Introduction |
| ICACT20210194 Slide.04
[Big Slide]
|
Chrome Click!! |
 |
According to the World Health Organization, every year nearly 1.3 million people die and 50 million are injured as a result of traffic accidents
Distracted driving is dangerous, claiming 2,841 human lives in 2018 alone. Among them: 1730 drivers, 605 passengers, 77 bicyclists, and 400 pedestrians on the US roads.
One of the main causes of road fatalities and injuries is distracted drivers. (NHTSA)
|
| ICACT20210194 Slide.05
[Big Slide]
|
Chrome Click!! |
 |
Nowadays, some major companies like Intel are working to develop an Advanced Driver-Assistant System (ADAS) to tackle the issue by creating systems to warn the driver when detection activities are occurring. From this deep learning takes part in ADAS computer vision techniques.
|
| ICACT20210194 Slide.06
[Big Slide]
|
Chrome Click!! |
 |
Dataset description |
| ICACT20210194 Slide.07
[Big Slide]
|
Chrome Click!! |
.jpg) |
State Farm Distracted Driver Detection is used which is a publically available image dataset on Kaggle.com.
|
| ICACT20210194 Slide.08
[Big Slide]
|
Chrome Click!! |
 |
Examples |
| ICACT20210194 Slide.09
[Big Slide]
|
Chrome Click!! |
 |
DeepLearning Model |
| ICACT20210194 Slide.10
[Big Slide]
|
Chrome Click!! |
 |
Transfer learning is a deep learning technique that stores the knowledge gained while solving one issue and use it to a new but relate
Nine models based on Mobile Net, VGG16 and ResNet50 models were used in this research.
|
| ICACT20210194 Slide.11
[Big Slide]
|
Chrome Click!! |
 |
Fine-tuning is a way of utilizing or applying transfer learning. There is a different kind of fine-tuning technique, truncating the last layer, using a smaller learning rate, and freezing the weights of the layers are among these methods
|
| ICACT20210194 Slide.12
[Big Slide]
|
Chrome Click!! |
 |
Result |
| ICACT20210194 Slide.13
[Big Slide]
|
Chrome Click!! |
 |
Table highlights the first and the last fifteenth epochs with train accuracy, validation accuracy, train loss, and validation loss values of nine models.
Comparing all re-trained models, model_2 used Mobile Net architecture, and pre-trained weights with 80 frozen layers yield the highest accuracy of 99.33% and 99.22% and the lowest loss of 2.32% and 2.4% rate on train and validation sets respectively.
|
| ICACT20210194 Slide.14
[Big Slide]
|
Chrome Click!! |
 |
Table illustrates predictions of 500 unseen test data and their accuracy. On the testbed, model_2 predicted 495 images correctly and only making 5 mistakes out of 500 test images, scoring 99% as overall accuracy and 0.04169 multiclass logarithmic loss.
|
| ICACT20210194 Slide.15
[Big Slide]
|
Chrome Click!! |
 |
Conclusion |
| ICACT20210194 Slide.16
[Big Slide]
|
Chrome Click!! |
 |
It was found that pre-trained deep learning models on the ImageNet dataset can be fine-tuned to improve the accuracy to classify State Farm distracted driver dataset.
Three architectures and their weights were used and applied three different freezing techniques on the layers making them overall nine models.
After re-training models with distracted drivers image dataset, they are evaluated by predicting 500 unseen test sets.
As a result, Mobile Net model_2 showed the best 99% accuracy. It is recommended that this model could be used in real-time environments.
|
| ICACT20210194 Slide.17
[Big Slide]
|
Chrome Click!! |
 |
Thanks for your attention! Now! Floor is open for QnA. Welcome any comment or question. |
