Robust Multivehicle Tracking With Wasserstein Association Metric in Surveillance Videos

Yanjie Zeng,Yuheng Li,Xinsha Fu,Jiawei Zhu,Lei Gao,Haifeng Li

doi:10.1109/access.2020.2978539

Abstract

Vehicle tracking based on surveillance videos is of great significance in the highway traffic monitoring field. In real-world vehicle-tracking applications, partial occlusion and objects with similarly appearing distractors pose significant challenges. For addressing the above issues, we propose a robust multivehicle tracking with Wasserstein association metric (MTWAM) method. In MTWAM, we analyze the advantage of the 1-Wasserstein distance (WD-1) on partial occlusion and employ the WD-1 as the similarity criterion to measure the similarity between tracklets and detections. Moreover, for distinguishing different objects with a similar appearance, we improve the feature presentation of vehicles by developing target-specific feature sparse coding (TSSC). To demonstrate the validity of this method, we present a quantitative evaluation of both the UA-DETRAC dataset and our vehicle highway surveillance videos dataset (VecHSV). In both cases, our method achieves state-of-the-art performances.

Highlights

Monitoring systems play an important role in the daily management of highways
Based on the above analysis, we propose a two-stage multivehicle tracking with Wasserstein association metric (MTWAM) method to track vehicles in highway surveillance videos
The contributions of this paper are as follows: 1) We propose a novel vehicle-tracking method called MTWAM for surveillance videos, which combines rich semantic and fine-grained features that are robust against appearance changes and possess sufficient discriminative power for similarity distractors

Summary

Introduction

Vehicle tracking based on surveillance videos, for which the goal is to provide a continuous trajectory to each target, is the main component of a monitoring system [1]–[4]. Gies, such as the Kalman filter [6]–[8] and the particle filter [9]–[11] These approaches must assume a dynamic model a priori and have trouble distinguishing objects close to other targets. The vast majority of recent methods are based on the tracking-by-detection approach, which builds trajectories via associated detections. These methods typically consist of the following components: a detection method and detection association method based on a similarity measure.

Objectives

Methods

Findings

Conclusion