Multiple Road Users Detection And Tracking System In Urban Mixed Traffic Scenes
Video analytic technology in traffic control and monitoring is getting more attention in recent years. This is because video analytic technology can perform traffic surveillance to extract traffic information such as vehicle counting and classification from video sequence. Large amount of road user...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2020
|
| Subjects: | |
| Online Access: | http://eprints.usm.my/56068/1/Multiple%20Road%20Users%20Detection%20And%20Tracking%20System%20In%20Urban%20Mixed%20Traffic%20Scenes.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Video analytic technology in traffic control and monitoring is getting more attention in recent years. This is because video analytic technology can perform traffic surveillance to extract traffic information such as vehicle counting and classification from video sequence. Large amount of road user data can be generated from video and this data would benefit for traffic planner. This provides the impact of video analytic in traffic surveillance. However, multiple road users tracking in urban traffic remains challenging because of large variation of road user appearance. To overcome the problems of multiple object tracking in mixed urban traffic, which are mis-detection, frequent ID switches and mis-classification, a system known as City Tracker, which incorporates Maximum Likelihood Estimation (MLE), YOLOv3 and DeepSORT is proposed in mixed urban traffic. City Tracker predicts the potential bounding box coordinates from the result of YOLOv3 and DeepSORT, then matches with the latest actual bounding box to overcome the mis-detection and frequent identity switch. On the other hand, MLE provides trajectory-based classification to solve mis-classification. This solution is tested with Urban Tracker dataset based on detection and tracking performance. The performance evaluations show that implementation of City Tracker increases Multiple Object Tracking Accuracy (MOTA) from 0.3503 to 0.3793 (8.28%) and Multiple Object Tracking Precision (MOTP) from 0.6245 to 0.6442 (3.15%) which are calculated from Precision and Recall as the evaluation metrics. MLE improves Recall from 0.7032 to 0.7838 (11.46%) and Precision from 0.7214 to 0.8334 (15.53%) in classification performance, which is better than conventional YOLOv3 and DeepSORT that do not consider City Tracker. |
|---|