Hybrid CPU-GPU Implementation of Tracking-Learning-Detection Algorithm

Ilker Gurcan
Department of Information Systems, The Middle East Technical University
The Middle East Technical University, 2014



   author={G{"U}RCAN, {.I}LKER},




Download Download (PDF)   View View   Source Source   



Tracking objects in a video stream is an important problem in robot learning (learning an object’s visual features from different perspectives as it moves, rotates, scales, and is subjected to some morphological changes such as erosion), defense, public security and many other various domains. In this thesis, we focus on a recently proposed tracking framework called TLD (Tracking-Learning-Detection). While having promising tracking results, the algorithm has high computational cost. The computational cost of the algorithm prevents running it at higher resolutions as well as running multiple instances of the algorithm to track multiple objects on CPU. In this thesis, we analyzed this framework with an aim to optimize it computationally on a CPU-GPU hybrid setting and developed a solution via using GP-GPU (General Purpose GPU) programming using Open-MP and CUDA. Our results show that 2.82 times speed-up at 480×270 resolution can be achieved. The speed-ups are higher at higher resolutions as expected in a massively parallel GPU platform, increasing to 10.25 times speed-up at 1920×1080 resolution. The resulting performance of the algorithm enables the algorithm to track multiple objects at higher frame rates in real-time and improving detection and tracking quality by allowing selection of configuration parameters requiring higher processing power.
No votes yet.
Please wait...

* * *

* * *

* * *

HGPU group © 2010-2022 hgpu.org

All rights belong to the respective authors

Contact us: