We propose several improvements of an existing baseline short-term visual tracking algorithm.
The baseline tracker applies a dynamic graph representation to track the target. The target local parts are used as nodes in the graph, while the connections between neighboring parts represent the graph edges. This flexible
model %representation of the target structure
proves useful in the presence of extensive target visual changes throughout the sequence.
A recent benchmark has shown that the tracker compares favorably in performance with other state-of-the-art trackers, with a notable weakness in cases of input sequences with high variance in scene and object lighting. We have performed an in-depth analysis of the tracker and propose a list of improvements.
With respect to an unstable component in the tracker implementation of the foreground/background image segmentation, we propose an improvement which boosts the accuracy in cases of rapid illumination change of the target.
We also propose a dynamic adjustment of the aforementioned segmentation with respect to the size of the resulting foreground, which improves tracking reliability and reduces the number of tracking failures.
The implemented improvements are analyzed on the VOT2015 benchmark. Fixing the unstable component yields improvements in cases of rapid illumination change and reduces failure rate, while the dynamic segmentation adjustment improves tracking accuracy and robustness in the vast majority of cases, barring rapid illumination change.