Performance Evaluation Of GPU Accelerated Real-Time Discriminative Correlation Filter Based Short-Term Visual Object Tracking

Teng, Pin Cheng (2018) Performance Evaluation Of GPU Accelerated Real-Time Discriminative Correlation Filter Based Short-Term Visual Object Tracking. Project Report. UTeM, Melaka, Malaysia. (Submitted)

Text (24 Pages) Performance Evaluation Of GPU Accelerated Real-Time Discriminative Correlation Filter Based Short-Term Visual Object Tracking.pdf - Submitted Version Download (1MB)

Abstract

Computer vision has brought a lot of benefits and performance in various applications thanks to the recent advancement of technology. Discriminative Correlation Filter based visual object tracking method has shown state-of-the art tracking performance and robustness. Since DCF based visual object tracker uses convolutional feature for model representation requires huge computing resources and has caused the real-time performance to drop significantly. Due to this reason, real-time application such as DCF-based drone visual object tracking cannot be realised. After the implementation of selected DCF-based tracker (Channel and Spatial Reliability DCF, CSR-DCF) using OpenCV library in embedded systems, the hardware performance of selected DCF-based tracker is profiled and analysed by using Intel VTune Amplifier. The computational heavy image processing operation is then offloaded from CPU host to GPU device in attempting to accelerate the CSR-DCF tracking performance in OpenCV using OpenCL framework. Although some of the image processing operation such as colour conversion from RGB color space to HSV color space is accelerated by a significant amount from 10.467ms to 3.03ms, the benefit of offloading cannot be enjoyed due to the data transfer overhead between the CPU host and GPU device. Thus, to accelerate the CSR-DCF tracker, OpenVX framework provided by Intel OpenVINO toolkit is used. OpenVX acceleration has shown significant speed up of 27.14% for image processing operation such as image resize which is used in the first part of the CSR-DCF tracker.

Actions (login required)

View Item