Publication: Motorcycle detection and tracking on malaysia road using deep learning
Loading...
Date
2022-08-01
Authors
Abdul Halim, Amin Haikal
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
In general, the possibility of an accident involving motorcyclist users larger than in other vehicles. The sharp movement of motorcycle motion on the road increases the risk of an accident in the surroundings due to the body frame of the vehicle itself which having smaller and has less protection compared to the other vehicles. To minimize this problem, a solution has been done by using deep learning for made predictions in detecting the presence of a motorcycle. Also, toward Sustainability Development Goal11 (SDG 11), these projects can be a kick starter as a smart monitoring system of the country road which provides information to users on the movement and road condition. This system used YOLOv4 to detect a better interface time than another detector (ACFand FRCNN). For performing the project, a custom dataset was made using only 1 class which was a motorcycle. The custom-trained dataset has been training using YOLOv4 and tested using its dataset and surveillance video. The challenge in detecting motorcycles in the surveillance video was the poor quality of the video. These videos experienced pre-processing of the project which used image processing (CLAHE andSR) to improve the quality of the video. This video which
has been applied with and without image processing been used as a test subject for real-time detection using YOLOv4 and compared using the post-processing method that evaluated the averageprecision of each surveillance video. The average precision of custom trained data setusing its dataset was 96.5%. The surveillance video without image enhancement forVideo 1 was 55%, Video 2 was 70% and Video 3 was 10%. The surveillance video with image enhancement (CLAHE and SR) for Video 1 was 70%, Video 2 was 50% and Video 3 was 10%.