Evaluation of Real-Time Single-Object Tracking Algorithms in a Non-Stationary Robotic Agent

Abstract: Visual object tracking is a fairly easy task for humans but a challenging problem in computer vision and thereby in humanoid implementation. Most of the existing object tracking evaluations are performed with prerecorded video footage, often with a stationary camera. This is not representative of a humanoid platform. The aim of the present thesis was to evaluate different object tracking algorithms’ suitability for being implemented in a humanoid by testing the algorithms’ performance in real-time using a non-stationary robotic agent. The results reflect a general trade-off between accuracy and computational cost. Kernelised correlation filters are depicted as a suitable choice for single-object tracking systems with limited computational power. Deep learning tracking algorithms is argued to be the better choice for systems with sufficient computational power.