Conference Name: International Conference on Communication and Computer Research (ICCR) 2023
Abstract: In the realm of computer vision, real-time object detection plays a pivotal role in a plethora of applications, from autonomous vehicles to smart surveillance systems. This paper presents an in-depth exploration of utilizing the YOLOv4-tiny model for the task of object detection, specifically focusing on road surface imagery classification into five distinct classes namely plain, crack, pothole, black ice, and obstacle. The focus on road surface imagery introduces challenges related to varying roads and weather conditions making the anomaly detection both important and challenging. By leveraging YOLOv4-tiny's lightweight architecture, we demonstrate its effectiveness in achieving road surface anomaly detection accuracy on par with larger models. On top of the great detection accuracy already achieved, we propose an enhancement of the existing model that has the potential to gain commendable compactness, making it an ideal candidate for deployment on edge devices, including resource-constrained platforms such as mobile phones. We limit the results presented in this paper to the great detection results achieved for the mentioned five classes. We outline the framework of the potential enhancement of our model towards being embedded in mobile devices end of the paper.