Authors :
Deniz Sen
Volume/Issue :
Volume 8 - 2023, Issue 10 - October
Google Scholar :
http://tinyurl.com/42a26rhv
Scribd :
http://tinyurl.com/3prm7dx5
DOI :
https://doi.org/10.5281/zenodo.10390839
Abstract :
Erroneous driver behavior is one of the most
dangerous aspects of driving in traffic, which can be
prevented if the dangerous actions are detected
immediately and the necessary actions are taken.
However, in order to be able to implement a hazard
detection framework, we first need a model that can
perceive the related objects in the traffic scene, which is
not a trivial task considering the diversity of the outside
conditions. In this paper, we propose a computer vision-
based system that only requires dash camera footage to
compute binary drivable area segmentation, dashed and
straight lane line segmentation, and traffic object
detection. We utilize two YOLOv5s and YOLOP, a
multitask architecture, to create a complete panoptic
perception model and train the member models using
the public BDD100K, traffic sign dataset, and our
private dataset. To reduce the significant annotation
overhead of the segmentation tasks, we use semi-
supervised learning techniques and a different
annotation approach for lane line labeling. We also
present 2 lane violation detection algorithms and
temporal smoothing techniques for the segmentation
tasks. We managed to achieve remarkable results in all 3
of our tasks and showed the usability of our system
under real-world scenarios.
Keywords :
YOLOP; Lane Detection; Traffic Object Detection; Drivable Area Segmentation; Traffic Violation.
Erroneous driver behavior is one of the most
dangerous aspects of driving in traffic, which can be
prevented if the dangerous actions are detected
immediately and the necessary actions are taken.
However, in order to be able to implement a hazard
detection framework, we first need a model that can
perceive the related objects in the traffic scene, which is
not a trivial task considering the diversity of the outside
conditions. In this paper, we propose a computer vision-
based system that only requires dash camera footage to
compute binary drivable area segmentation, dashed and
straight lane line segmentation, and traffic object
detection. We utilize two YOLOv5s and YOLOP, a
multitask architecture, to create a complete panoptic
perception model and train the member models using
the public BDD100K, traffic sign dataset, and our
private dataset. To reduce the significant annotation
overhead of the segmentation tasks, we use semi-
supervised learning techniques and a different
annotation approach for lane line labeling. We also
present 2 lane violation detection algorithms and
temporal smoothing techniques for the segmentation
tasks. We managed to achieve remarkable results in all 3
of our tasks and showed the usability of our system
under real-world scenarios.
Keywords :
YOLOP; Lane Detection; Traffic Object Detection; Drivable Area Segmentation; Traffic Violation.