A Performance Comparison of the Single Object Tracking Algorithms ROLO and Tiny-ROLO

Sammanfattning: Object tracking in videos is an extension of ordinary object detection, to continuously predict and follow the objects’ path of motion. The Recurrent YOLO (ROLO) algorithm is a two-stage tracking algorithm that utilizes deep neural networks to track objects in image sequences. This study investigates how object detection in ROLO affects the accuracy of object tracking, by comparing ROLO with Tiny-ROLO. Tiny-ROLO is our suggested version of the tracker, where the object detection is made up of Tiny-YOLO instead of YOLO. The object tracking accuracy is quantified with Intersection over Union and prediction offset from the actual position. The algorithms were evaluated on datasets with cars and humans. Our analysis shows that ROLO has accurate and reliable tracking performance. Tiny-ROLO has worse accuracy. With certain conditions, the accuracy is the same, which can promote the use of Tiny-ROLO because of faster image processing.