Deep learning-based fine-tuned multi-vehicle tracking with classification correction

Single-camera multi-vehicle tracking is a preliminary step for traffic optimization. To perform vehicle tracking, a fine-tuned classification dataset was created from multiple sources. A ResNet classifier was trained, and its knowledge was utilized to build a localization dataset automatically without manual annotations. The dataset consists of videos generated from the MLK testbed. Three different object detection frameworks (SSD, Yolov3, and Yolov4) were evaluated with a pipeline that consists of the detector, the fine-tuned classifier and the tracker (DeepSort). Then the detector with the highest accuracy (Yolov3) was trained on the localization dataset. 90% of the ResNet classifier knowledge was distilled successfully in the final trained Yolov3 model. The tracker classification accuracy was further improved by proposing a correction methodology that considers both the camera's distance and limited data from the future frames.

Grantor dc:publisher

University of Tennessee at Chattanooga

Year dc:date.available

2022

Author dc:creator

• Alharin, Alnour

Contributors dc:contributor

• Sartipi, Mina
• Liang, Yu (Hugh); Wu, Dalei
• College of Engineering and Computer Science

• Computer vision
• Machine learning

Licence

• http://rightsstatements.org/vocab/InC/1.0/

Language dc:language

English, eng

Repository record dc:identifier

https://scholar.utc.edu/theses/698

OAI identifier oai:identifier

oai:scholar.utc.edu:theses-1869