Evaluating The Trade-Off Between Accuracy And Real-Time Tracking Performance Of Discriminative Correlation Filter (DCF) Based Short-Term Visual Tracker

Tung, Pooi Lee (2018) Evaluating The Trade-Off Between Accuracy And Real-Time Tracking Performance Of Discriminative Correlation Filter (DCF) Based Short-Term Visual Tracker. Project Report. Universiti Teknikal Malaysia Melaka, Melaka, Malaysia. (Submitted)

Text (24 Pages) Evaluating The Trade-Off Between Accuracy And Real-Time Tracking Performance Of Discriminative Correlation Filter (DCF) Based Short-Term Visual Tracker.pdf - Submitted Version Download (456kB)

Abstract

Object tracking is an important component of many computer vision systems.In recent years,Discriminative Correlation Filter (DCF) based methods have significantly advanced in tracking.DCF have demonstrated excellent tracking accuracy,measured in Expected Average Overlap (EAO),for visual object tracking.Three methods that have been applied on selected DCF based tracker (Channel and Spatial Reliability DCF,CSR-DCF) to evaluate the trade-off between accuracy and real-time tracking performance. The first proposed speed-accuracy trade-off method is to reduce the tracker computation by disable the update of the channel reliability weights.The second proposed method is to further reduce the tracker computation by disable the update of spatial reliability map but still performing filter learning for each frame.Then,the third method is to even further reduce the computation by skipping the learning of the filter in every alternative frames.The performance of CSR-DCF based tracker is tweak-off by increasing the number of histogram oriented gradients (HoGs) features channel from 18 to 27 channels to increase the tracking accuracy with a bit reduce of speed reduced.Result of benchmarking shows that the third method,disable both channel and spatial reliability with alternative frame filter learning,is able to meet the 10.8FPS real-time requirement with EAO of 0.615.

Actions (login required)

View Item