Evaluation of visual tracking algorithms for embedded devices

Abstract

Today’s embedded platforms enable executing difficult tasks such as visual tracking. However, such resource-constrained systems are still facing challenges regarding the performance and accuracy in executing these tasks. This paper presents the evaluation of 5 open-source visual tracking implementations available from the contributions branch of the Open Computer Vision (OpenCV) library. This evaluation is performed based on the performance and accuracy of these implementations when embedded in a Raspberry Pi. The algorithms evaluated are On-Line Boosting, Multiple Instance Learning (MIL), Median Flow, Tracking-Learning-Detection (TLD), and Kernelized Correlation Filters (KCF). Even if commercial implementations of these algorithms perform better than their open-source version, the popularity of OpenCV motivates this evaluation. Tests are based on a benchmark of 100 video streams from which the tracking implementations should follow moving objects. The algorithms are evaluated for accuracy using averaged Jaccard indices and for performance by measuring their frame rate. We want to find an open-source implementation that performs well on these two criteria when tested on an embedded platform. Results show Median Flow being the fastest but its accuracy is the lowest. We therefore recommend KCF as it is the second fastest and the most accurate.