Transportation safety enhancement by implementing computer vision based models in intelligent transportation systems

Thumbnail Image
Ketabchi Haghighat, Arya
Major Professor
Sharma, Anuj
Sarkar, Soumik
Day, Christopher
Dong, Jing
Krishnamurthy, Adarsh
Committee Member
Journal Title
Journal ISSN
Volume Title
Research Projects
Organizational Units
Journal Issue
Is Version Of
Civil, Construction, and Environmental Engineering
By emerging deep learning techniques in recent years, most fields from medical science to all engineering branches implemented these techniques in different ongoing challenges. The popularity of deep learning techniques in this research is due to their strength in finding a comparable solution with more analytical approaches using large data. Intelligent Transportation Systems is also one of the fields which flourished in this situation. In this thesis, I implemented different computer vision and deep learning models to enhance traffic safety. In the first chapter of this thesis, I surveyed recently published literature on different deep learning applications in Intelligent Transportation Systems and mentioned some of the gaps in the field which still have the potential for more focus and research. I proposed an augmented annotation pipeline inspired by imitation learning in the next chapter. In this pipeline, I proposed a pipeline that can save lots of time on annotating new data for training object detection models to detect traffic objects in naturalistic driving data and traffic surveillance data. By visualizing the model's performance at the end of each iteration, this iterative pipeline lets us recognize the model's weakness to detect particular objects in particular locations. This knowledge provides the annotators to focus more on those particular situations. This pipeline showed significant reductions in the number of false positives - negatives and an increase in the number of true positives after the fourth iteration of the augmented annotation process. Then in the next step, we proposed a fully automated pipeline to detect wrong-way driving on highways using Pan-tilt-zoom traffic cameras. A scalable solution is proposed in this pipeline by combining different multi-object detection and tracking models with a deep model. Contrary to existing solutions, which required exogenous specification of the camera as a separate parameter, in the suggested solution, camera orientation is considered a variable. Camera rotation detection performs by the model automatically. The model adopts new decision criteria accordingly by learning them from a neural network model. We showed the proposed solution could detect WWD with the precision of 0.99, where the recall is 0.97. Finally, in the next chapter, I proposed a deep model to predict cognition disability by observing the driver's driving behavior while driving. For this manner, a deep learning model with combination of CNN and RNN layers proposed. This model using a Convolutional Neural Network model extracted features from the video, and sets of Long-Short Term Memory cells have been implemented to analyze these features over time. This chapter discusses the results, performance, and limitations of this model. We showed that using this model we can predict driver’s cognition stat with over 70% accuracy.