• PDF / 1,682,536 Bytes
• 16 Pages / 439.37 x 666.142 pts Page_size
• 78 Downloads / 259 Views

DOWNLOAD

REPORT

---

Abstract. We address an anomaly detection setting in which training sequences are unavailable and anomalies are scored independently of temporal ordering. Current algorithms in anomaly detection are based on the classical density estimation approach of learning high-dimensional models and finding low-probability events. These algorithms are sensitive to the order in which anomalies appear and require either training data or early context assumptions that do not hold for longer, more complex videos. By defining anomalies as examples that can be distinguished from other examples in the same video, our definition inspires a shift in approaches from classical density estimation to simple discriminative learning. Our contributions include a novel framework for anomaly detection that is (1) independent of temporal ordering of anomalies, and (2) unsupervised, requiring no separate training sequences. We show that our algorithm can achieve state-of-the-art results even when we adjust the setting by removing training sequences from standard datasets. Keywords: Anomaly detection · Discriminative text · Surveillance · Temporal invariance

1

· Unsupervised · Con-

Introduction

Anomaly detection is an especially challenging problem because, while its applications are prevalent, it remains ill-defined. Where there have been attempts at definitions, they are often informal and vary across communities and applications. In this paper, we define and propose a solution for a largely neglected subproblem within anomaly detection, where two constraints exist: (1) no additional training sequences are available; (2) the order in which anomalies occur should not affect the algorithm’s performance on each instance (Fig. 1). This is an especially challenging setting because we cannot build a model in advance and find deviations from it; much like clustering or outlier detection, the context is defined by the video itself. This setting is prominent in application fields such as robotics, medicine, entertainment, and data mining. For instance: Electronic supplementary material The online version of this chapter (doi:10. 1007/978-3-319-46454-1 21) contains supplementary material, which is available to authorized users. c Springer International Publishing AG 2016  B. Leibe et al. (Eds.): ECCV 2016, Part V, LNCS 9909, pp. 334–349, 2016. DOI: 10.1007/978-3-319-46454-1 21

A Discriminative Framework for Anomaly Detection in Large Videos

335

– First-time data. A robotics team wants to create a robust set of algorithms. They teleoperate a robot performing a new task or operating in a new environment. The team would like to find out what special cases the robot may have to handle on the perception side, so they ask for a list of the most anomalous instances according to the robot’s sensor data relative to that day’s conditions and performance. – Personalized results: context semantically defined as coming only from the test set. (a) A father wants to find the most interesting parts of the 4-h home video of his family’s Christmas. (b) A healthcare profession