By combining object detection algorithms with conventional tracking algorithms, this work aims to increase the performance of object detectors. The detected boxes are compared to the tracker boxes using the Hungarian algorithm to determine which boxes are not associated with a tracker. When a box is not associated, the detector box is used to initialize a tracker and generate a new box when there is no detection in the spatial region in sequential frames. Every box should be moving and is not allowed to touch the edge of the frame. The object detectors used in this framework are SSD Inception, Yolo, and Retinanet. The object trackers used are MOSSE, KCF, and CSRT. SSD Inception has low performance, with the framework it achieves a doubling in IoU, from 0.14 to 0.29. CSRT shows the most significant increase in performance across the board.
Click here for the Paper and the Poster
Geart van der Ploeg