Large Margin Object Tracking with Circulant Feature Maps

Authors:

Mengmeng Wang， Yong Liu， Zeyi Huang

Institutions:

Mengmeng Wang and Yong Liu are from the Institute of Cyber-Systems and Control， Zhejiang University， China； Zeyi Huang is from Exacloud Limited， Zhejiang， China。

This paper is accepted by CVPR2017

The results of LMCF（HOG+CN） and DeepLMCF（CNN）：

Introduction

Motivation

:

①Framework： Structured output SVM based tracking algorithms have shown favorable performance while limited by the time-consuming candidate sampling and complex optimization。

②Forward Tracking： uncontrolled while decisive

③Model update： significant while time-consuming

Ø

Abstract:

Structured output support vector machine （SVM） based tracking algorithms have shown favorable performance recently。 Nonetheless， the time-consuming candidate sampling and complex optimization limit their real-time applications。

In this paper， we propose a novel large margin object tracking method which absorbs the

strong discriminative ability from structured output SVM and speeds up by the correlation filter algorithm significantly。 The experimental results demonstrate that the proposed tracker performs superiorly against several state-of-the-art algorithms on the challenging benchmark sequences while runs at speed in excess of 80 frames per second。

Ø

Contributions:

The main contributions of our work can be summarized as follows：

–We propose a novel structured SVM based tracking method which takes dense circular samples into account in both training and detection processes。 A bridge was built up to link our problem with CF， which speeds up the optimization process significantly。

–We explore a multimodal target detection technique to prevent the model drift problem introduced by similar objects or background noise。

–We establish a model update strategy to avoid model corruption by the high-confidence selection from tracking results。

We implement two versions of the proposed tracker with the representations from both conventional hand-crafted and deep convolution neural networks （CNNs） based features to validate the strong compatibility of the algorithm。

Method：

1。 Framework： Structured output SVM

Input： ？∈？

Output： ？={（？，ℎ）|？∈{0，…，？−1， ℎ∈？−1}}

All the cyclic shifts of the image patch centered around the target are considered as the training samples （？，？_（？，ℎ） ）。

Optimization problem：

2。 Forward Tracking： Multimodal

3。 Model update： High-confidence Update

Experiments：

OTB-2015 （LMCF， 80FPS）

More results can be found in our paper。

Conclusions：

The proposed LMCF tracker absorbs the strong discriminative ability from structured output SVM and speeds up by the correlation filter algorithm significantly。 In order to prevent model drift introduced by similar objects or background noise， a multimodal target detection technique is proposed to ensure the correct detection。 Moreover， we establish a high-confidence model update strategy to avoid the model corruption problem。

Download of the tracking results

：

Download password： 7q1g