Detecting tracking errors via forecasting

نویسندگان

  • Obaidullah Khalid
  • Andrea Cavallaro
  • Bernhard Rinner
چکیده

We propose a framework that detects the failures of a tracker using its output only (Figure 1). The framework is based on a state-background discrimination approach that generates a track quality score, which quantifies the ability of the tracker to remain on target. We define a background region around the target and split it into four sub-regions, each with the same size as the state. We then determine the distributions of the state and each of the smaller background regions using colour distribution fields (DF) [5]. A DF represents a smoothed histogram of the image region composed of several layers.We compare the state and background distributions to quantify the similarity between the two regions to produce the track quality score. However, the raw values of the track quality score [3] may have variable ranges, hence limiting its use to specific sequences or trackers only. To address this limitation, we model the track quality score as time series and employ a forecasting model to detect tracking errors. Let I = {It} t=1 be an image sequence and xt be the estimated state at time t = 1, ...,T . Let St be the region in It defined by xt . Using motion information ~ν∆t1 from a past short temporal window ∆t1 and the target state xt−1 we select the background region Bt in It (Figure 2). We split Bt into four smaller equally sized regions, ba t , each with the same width and height of St . We then determine the distribution for St , d ′′ St , and each of the smaller background regions ba t , d ′′ ba t , using colour DFs [5]. The tracking quality score yt is determined by quantifying the similarity between the distributions of Bt and St using the L1 distance, where low (high) values of yt indicate similarity (dissimilarity) between the two regions. We detect tracking errors by employing time series analysis to model Y = {yt} t=1, a univariate discrete time series, for forecasting. We use the Auto Regressive Moving Average (ARMA) model [1] which is built using past data and forecasts employing both the past and present data. The difference between the forecast and the original returns a re-scaled signal, which highlights only significant changes. We build the forecasting model using data within a past temporal window ∆t2 and then forecast future values ŷt+l using the forecasting model and its estimated parameters, Ψ, over the forecast length l ≥ 1 at time t. The forecasting error |ẽt+l | = yt+l − ŷt+l is employed to determine time instants when a tracking error occurs. Since values of ŷt+l are dependent on past values of yt , between t−∆t2 and t, |ẽt+l | temporally smooths yt . Significant changes (tracking errors) in the value of yt are reproduced by |ẽt+l | and detected for |ẽt+l | ≥ τ1, where τ1 is an experimentally derived threshold. We use a sparse features based tracker [4], to train the proposed approach Detecting Tracking Errors via Forecasting (DTEF) on 20 sequences from dataset D1 and then test DTEF on 20 sequences from the Object Tracking Benchmark (OTB) dataset. Using precision (P), recall (R), F-score (F) and false positive rate (FPR), we compare DTEF with two variations of the proposed approach: NAIVE and RAW; one state-ofthe-art (SOA) for tracker error detection [3]: CovF; and two SOA features employed for video tracking [2]: RgbHist and RLHist. Results on the

برای دانلود متن کامل این مقاله و بیش از 32 میلیون مقاله دیگر ابتدا ثبت نام کنید

ثبت نام

اگر عضو سایت هستید لطفا وارد حساب کاربری خود شوید

منابع مشابه

KHALID, CAVALLARO, RINNER: DETECTING TRACKING ERRORS VIA FORECASTING 1 Detecting tracking errors via forecasting

We propose a tracker-independent framework to determine time instants when a video tracker fails. The framework is divided into two steps. First, we determine tracking quality by comparing the distributions of the tracker state and a region around the state. We generate the distributions using Distribution Fields and compute a tracking quality score by comparing the distributions using the L1 d...

متن کامل

Automatic Detection and Correction of Multi-class Classification Errors Using System Whole-part Relationships

Real-world dynamic systems such as physical and atmosphereocean systems often exhibit a hierarchical system-subsystem structure. However, the paradigm of making this hierarchical/modular structure and the rich properties they encode a “first-class citizen” of machine learning algorithms is largely absent from the literature. Furthermore, traditional data mining approaches focus on designing new...

متن کامل

Event Detection and Localization Using Sensor Networks

Although sensor network research was initially driven by military applications such as enemy tracking and surveillance in battle field, they have been deployed in several civilian and commercial applications such as habitat monitoring, environmental observations and forecasting, health monitoring etc. In this paper, we study and propose algorithms for detecting fire and localizing the same usin...

متن کامل

Effect of Curved Path Monopulse Radar Platform’s Grazing Angle on Height of Floated Targets Detection

Monopulse radarsare one of the most accurate tracking radars used to guide various platforms. Detection and tracking of surface targets with these radars are performed in order to point strike targets. Sea clutter presents challenges for detecting floated targets and in addition to affecting detection height, it can cause errors in monopulse angle finding. In this paper, the airborne monopulse ...

متن کامل

Activity Tracking for Pilot Error Detection from Flight Data

This paper presents an application of activity tracking for pilot error detection from flight data. It describes the Crew Activity Tracking System (CATS), in-flight data collected from the NASA Langley Boeing 757 Airborne Research Integrated Experiment System aircraft, and a model of B757 flight crew activities. It then presents an example of CATS detecting actual in-flight crew errors.

متن کامل

ذخیره در منابع من


  با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید

برای دانلود متن کامل این مقاله و بیش از 32 میلیون مقاله دیگر ابتدا ثبت نام کنید

ثبت نام

اگر عضو سایت هستید لطفا وارد حساب کاربری خود شوید

عنوان ژورنال:

دوره   شماره 

صفحات  -

تاریخ انتشار 2016