Robust correlation filter tracking with deep semantic supervision. Issue 5 (19th March 2019)

• Record Type: Journal Article
• Authors: Wang, Wei; Chen, Zhaoming; Douadji, Lyes; Shi, Mingquan
• Is Part Of: IET image processing. Volume 13:Issue 5(2019)
• Page Start: 754
• Page End: 760
• Publication Date: 2019-03-19
• Subjects: particle filtering (numerical methods) -- learning (artificial intelligence) -- target tracking -- convolutional neural nets; robust correlation filter tracking -- high tracking performance -- tracking failure -- deep semantic supervision tracking framework -- redetection tracking mechanism -- particle filtering resampling -- CF tracker -- deep convolutional neural network -- tracking frames -- target occlusion -- handcrafted features -- real‐time performance -- OTB2013 benchmark datasets -- OTB2015 benchmark datasets; Image processing -- Periodicals; 621.36705
• DOI: 10.1049/iet-ipr.2018.5314
• Languages: English
• ISSNs: 1751-9659

Abstract

Abstract : Traditional correlation filter (CF) tracking has achieved high tracking performance and speed. However, it easily falls into tracking failures in some cases of target occlusion, deformation, rotation etc. Tracking failure also contaminates the CF model and makes it less discriminative. To tackle these problems, the authors propose a deep semantic supervision tracking framework. This framework integrates the advantages of multiple features and tracking methods into an evaluation and redetection tracking mechanism. In this work, customised deep convolutional neural network (CNN) with particle filtering (PF) resampling was employed to alleviate the contamination of the CF model and improve tracking performance. The authors also adopted a mixed decision mechanism for CF tracking results evaluation. Furthermore, based on the observation that most tracking frames can be easily tracked by a CF tracker using handcrafted features, authors' tracking method achieves real‐time performance. It should be noted that the proposed framework is flexible and extensible to improve other existing trackers. In authors' extensive experiments on large benchmark datasets including OTB2013 and OTB2015, the proposed tracker performed favourably compared to the state‐of‐the‐art methods.