Comparison of Visual Tracking Algorithms and Realization of Online Tracking System

Abstract

Visual tracking is one of the essential application in computer vision. Many visual tracking algorithms have been designed and realized in practical application. By employing these visual tracking algorithms, a lot of trackers can be used in target tracking in different conditions. In order to utilize these tracker more efficiently, we need to estimate the performance of them and apply them into proper environment respectively. Due to the various characteristics and principle of these algorithms, their performances in robustness evaluation should be analyzed specially for test sequences with different attributes. This project brief introduces and estimates the performance of the following nine trackers: CSK (Circulant Structure of Tracking-by-Detection with Kernels), CT（Real-time Compressive Tracking）, CXT (Context Tracker), MTT (Multi-Task Sparse Learning), L1APG (L1 Tracker Using Accelerated Proximal Gradient Approach), DFT (Distribution Fields for Tracking), IVT (Incremental Learning for Robust Visual Tracking), SCM (Robust Object Tracking via Sparsity-based Collaborative Model), LOT(Locally Orderless Tracking) with categorizing them as generative or discriminative trackers. The robustness test includes one-pass evaluation (OPE), temporal robustness evaluation (TRE) and spatial robustness evaluation (SRE), which estimate the trackers’ performance comprehensively. The evaluation on tracking speed is also given, so as to apply the high speed trackers into online tracking system. Ultimately, an online tracking system is developed with three different trackers, to cater the need of internal and external tracking attributes.