Publication: On-road motorcycle detection and tracking using deep learning
| datacite.subject.fos | oecd::Engineering and technology::Electrical engineering, Electronic engineering, Information engineering | |
| dc.contributor.author | Zainon, Ahmad Zakwan | |
| dc.date.accessioned | 2024-02-13T09:31:40Z | |
| dc.date.available | 2024-02-13T09:31:40Z | |
| dc.date.issued | 2020-08-01 | |
| dc.description.abstract | Road conditions are becoming more difficult, particularly in an urban area such as Penang. The rising use of motorcycles also raises the number of motorcycle crashes and associated with a fatal injury. The existing system tracks traffic breaches mainly through CCTV videos where traffic managers have to peer into the frame where the traffic breach happens. For intelligent traffic monitoring using vision-based approach, motorcycles are often difficult to detect because of the variation in colour, scale, form,and trajectories. Hence, a better vision-based on-road motorcycle detection system is needed to improve the detection of motorcycles for traffic surveillance analysis. This research is aimed to build a DNN-based on-road motorcycle detection and tracking system which suitable for Penang traffic road conditions. This motorcycle detection and tracking system include a motorcycle detection algorithm, video enhancement algorithm, and motorcycle tracking. First, the motorcycle detection algorithm is used. YOLOv3 with pretrain COCO dataset is used to detect the motorcycles. The motorcycle detection algorithm achieves Average Precision (AP) of 4.54% for video 1, 12.09% for video 2, 1.39% for video 3 and 5.01% for video 4. Then, video enhancement techniques, which are interlace filter, denoise filter, and super-resolution are used to further improve the performance. Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network (ESPCN) model is used for super-resolution. By using deinterlacing filters, the performance of motorcycle achieves AP of 10.88% for video 1, 12.58% or video 2, 1.29% for video 3 and 1.92% for video 4. By using deinterlacing and denoising filters, the motorcycle detection achieves AP of 9.78% for video 1, 12.12% for video 2, 0.95% for video 3 and 2.04% for video 4. By using super-resolution, the motorcycle detection achieves AP of 20.97% for video 1, 14.36% for video 2, 1.52% for video 3, 8.21% for video 4. For the algorithm for motorcycle tracking, the Euclidian distance is used to determine the same motorcycle in each video frame by two essential parameters, MaxDisappeared and MaxDistance which can be used to presume the middle of the current frame and the previous frame to vanish after number of frames and distance parameters. | |
| dc.identifier.uri | https://erepo.usm.my/handle/123456789/18334 | |
| dc.language.iso | en | |
| dc.title | On-road motorcycle detection and tracking using deep learning | |
| dc.type | Resource Types::text::report | |
| dspace.entity.type | Publication | |
| oairecerif.author.affiliation | Universiti Sains Malaysia |