Online multi-object tracking using KCF-based single-object tracker with occlusion analysis
- PDF / 5,567,538 Bytes
- 15 Pages / 595.276 x 790.866 pts Page_size
- 87 Downloads / 160 Views
REGULAR PAPER
Online multi‑object tracking using KCF‑based single‑object tracker with occlusion analysis Honghong Yang1,2 · Sheng Gao2 · Xiaojun Wu1,2 · Yumei Zhang1,2 Received: 20 January 2020 / Accepted: 9 July 2020 / Published online: 6 August 2020 © Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract Most state-of-the-art multiple-object tracking (MOT) methods adopt the tracking-by-detection (TBD) paradigm, which is a two-step procedure including the detection module and the tracking module. In these methods, the tracking performance heavily depends on initial detections and data association. In this paper, we present an online MOT method by introducing a single-object tracking (SOT) based on correlation filter. Our contributions lie in twofold: (a) we use the KCF-based SOT in learning of discriminative target appearance relying on handcrafted and deep features and (b) we employ the predicted result to refine the detection mistakes in a new way. Furthermore, we introduce normalize APCE score as an occlusion indicator of tracklet confidence, and build a candidate target hypotheses set to improve the association performance. Both approaches are found beneficial to eliminate the track errors caused by the inability of association algorithm. The experimental results, both qualitative and quantitative on three benchmark datasets, demonstrate that our tracking algorithm achieves comparable or even better results than competitor approaches. Keywords Online multi-object tracking · Kernelized correlation filter · Two-step data association · Detection mistake · Track error
1 Introduction Multiple-object tracking aims to identify relevant objects and estimates their locations in a video sequence. It has various applications in academic and industrial field in recent years, such as surveillance, human computer interface and intelligent transportation. State-of-the-art MOT algorithms follow the trackingby-detection (TBD) framework, which takes the detections provided by an object detector as input and links the detections among different frames to get the final trajectories. Communicated by C. Xu. * Xiaojun Wu [email protected] Yumei Zhang [email protected] 1
Key Laboratory of Modern Teaching Technology, Ministry of Education, Shaanxi Normal University, Xi’an 710062, China
School of Computer Science, Shaanxi Normal University, Xi’an 710062, China
2
Therefore, TBD-based MOT can be classified into two steps: detection module and tracking module. Typically, detection results are essential in MOT but sometimes are not accurate enough for a good MOT. Existing MOT methods always pay their attention on tracking module, a process consisting of data association and model optimization. To improve MOT, many data association methods are proposed, such as Multi-Hypothesis Tracker (MHT) [1], Joint Probabilistic Data Association (JPDA) [2], Network Flow [3] and Learning methods [4]. Among these algorithms, two kinds of features are involved. First, the appearance, shape and location of one object are used and
Data Loading...
