Abstract

This thesis explores two topics in video object tracking: (1) performance evaluation of tracking techniques, and (2) tracker drift detection and correction. Tracking performance evaluation consists into comparing a set of trackers' performance measures and ranking these trackers based on those measures. This is often done by computing performance averages over a video sequence and then over the entire test video dataset, consequently resulting in an important loss of statistical information of performance between frames of a video sequence and between the video sequences themselves. This work proposes two methods to evaluate trackers with respect to each other. The first method applies the median absolute deviation (MAD) to effectively analyze the similarities between trackers and iteratively ranks them into groups of similar performances. The second method gains inspiration from the use of robust error norms in anisotropic diffusion for image denoising to perform grouping and ranking of trackers. A total of 20 trackers are scored and ranked across four different benchmarks, and experimental results show that using our scoring evaluation is more robust than using the average over averages.
In the second topic, we explore methods to the detection and correction of tracker drift. Drift detection refers to methods that detect if a tracker is about to drift or has drifted away while following a target object. Drift detection triggers a drift correction mechanism which updates the tracker's rectangular output bounding box. Most drift detection and correction algorithms are called while the target model is updating and are, thus, tracker-dependent. This work proposes a tracker-independent drift detection and correction method. For drift detection, we use a combination of saliency and objectness features to evaluate the likelihood an object exists inside a tracker's output. Once drift is detected, we run a region proposal network to reinitialize the bounding box output around the target object. Our implementation applied on two state-of-the-art trackers show that our method improves overall tracker performance measures when tested on three benchmarks.