Robust pedestrian tracking and recognition from FLIR video: a unified approach via sparse coding.

Xin Li,Rui Guo,Chao Chen

doi:10.3390/s140611245

Abstract

Sparse coding is an emerging method that has been successfully applied to both robust object tracking and recognition in the vision literature. In this paper, we propose to explore a sparse coding-based approach toward joint object tracking-and-recognition and explore its potential in the analysis of forward-looking infrared (FLIR) video to support nighttime machine vision systems. A key technical contribution of this work is to unify existing sparse coding-based approaches toward tracking and recognition under the same framework, so that they can benefit from each other in a closed-loop. On the one hand, tracking the same object through temporal frames allows us to achieve improved recognition performance through dynamical updating of template/dictionary and combining multiple recognition results; on the other hand, the recognition of individual objects facilitates the tracking of multiple objects (i.e., walking pedestrians), especially in the presence of occlusion within a crowded environment. We report experimental results on both the CASIAPedestrian Database and our own collected FLIR video database to demonstrate the effectiveness of the proposed joint tracking-and-recognition approach.

Highlights

The capability of recognizing a person at a distance in nighttime environments, which we call remote and night biometrics, has gained increasingly more attention in recent years
We report our experimental results with two forward-looking infrared (FLIR) pedestrian databases: one is collected by CASIA (Dataset C in the CASIA Gait Database, Publicly available at http://www.cbsr.ia.ac.cn/english/Databases.asp), and the other is collected at the WVU Erickson
We studied a unified approach toward robust pedestrian tracking and recognition from FLIR video via sparse coding

Summary

Introduction

The capability of recognizing a person at a distance in nighttime environments, which we call remote and night biometrics, has gained increasingly more attention in recent years. Object tracking and recognition are two basic building blocks in almost all video-based biometrics systems, including forward-looking infrared (FLIR)-based ones. The literature of object detection/tracking, face recognition and visual surveillance is huge; for recent advances, please refer to [1,2,3] and their references; pedestrian detection and tracking from FLIR video has been studied in [4,5,6,7]. The sparsity constraint is enforced about the total number of nonzero coefficients in a, which gives rise to the following constrained optimization problem: min ||a||0 subject to ||x − Aa|| ≤ a (1). Various algorithms have been developed in recent years to solve this class of l1 -minimization problems (for a recent review, please refer to [28] and its references). Within the scope of this paper, we opt to review two of them; namely, object tracking and object recognition

Objectives

Results

Conclusion