Performance of Optimized CSK, DFT, and LOT for Video-based Container Tracking System using Simulated Annealing.

This study considers the performance of three object tracking algorithms, namely Circulant Structure of Tracking-by-Detection with Kernels (CSK), Distribution Fields for Tracking (DFT), and Locally Orderless Tracking (LOT). These methods are used to track containers in five video recordings of container loading and unloading. At each video frame change, the three methods are implemented to detect and track moving containers. The comparison is aimed to evaluate the tracking performances of the three methods for this instance by means of seven performance indexes, i.e., Frame per Second (FPS), Sample Accuracy, Average IoU, Average IoG, Object Tracking Accuracy (OTA), Precision, and Recall. The experimental results are presented in terms of the average accuracy, IoU, IoG, OTA, Precision, and Recall of the sample. Based on the results, tracking methods using DFT are recommended over those using CSK or LOT, even though its average frame rate is slower compared with CSK. Optimization is applied to each method by using simulated annealing to find their optimal parameters. The results show that DFT and LOT perform best, while CSK is not able to track the containers. DFT especially yields better performance on four of the videos. [ABSTRACT FROM AUTHOR]