University of MarylandComputer Vision Lab 1 A Perturbation Method for Evaluating Background Subtraction Algorithms Thanarat Horprasert, Kyungnam Kim, David.

Slides:



Advertisements
Similar presentations
Applications of one-class classification
Advertisements

Principles of Density Estimation
Road-Sign Detection and Recognition Based on Support Vector Machines Saturnino, Sergio et al. Yunjia Man ECG 782 Dr. Brendan.
Evaluating Classifiers
Caroline Rougier, Jean Meunier, Alain St-Arnaud, and Jacqueline Rousseau IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 21, NO. 5,
Change Detection C. Stauffer and W.E.L. Grimson, “Learning patterns of activity using real time tracking,” IEEE Trans. On PAMI, 22(8): , Aug 2000.
Evaluation of segmentation. Example Reference standard & segmentation.
© Tan,Steinbach, Kumar Introduction to Data Mining 4/18/ Other Classification Techniques 1.Nearest Neighbor Classifiers 2.Support Vector Machines.
Abandoned Object Detection for Public Surveillance Video Student: Wei-Hao Tung Advisor: Jia-Shung Wang Dept. of Computer Science National Tsing Hua University.
Qualifying Exam: Contour Grouping Vida Movahedi Supervisor: James Elder Supervisory Committee: Minas Spetsakis, Jeff Edmonds York University Summer 2009.
Foreground Background detection from video Foreground Background detection from video מאת : אבישג אנגרמן.
Robust Object Tracking via Sparsity-based Collaborative Model
Adviser : Ming-Yuan Shieh Student ID : M Student : Chung-Chieh Lien VIDEO OBJECT SEGMENTATION AND ITS SALIENT MOTION DETECTION USING ADAPTIVE BACKGROUND.
Optimization & Learning for Registration of Moving Dynamic Textures Junzhou Huang 1, Xiaolei Huang 2, Dimitris Metaxas 1 Rutgers University 1, Lehigh University.
Modeling Pixel Process with Scale Invariant Local Patterns for Background Subtraction in Complex Scenes (CVPR’10) Shengcai Liao, Guoying Zhao, Vili Kellokumpu,
International Conference on Image Analysis and Recognition (ICIAR’09). Halifax, Canada, 6-8 July Video Compression and Retrieval of Moving Object.
1 Learning to Detect Objects in Images via a Sparse, Part-Based Representation S. Agarwal, A. Awan and D. Roth IEEE Transactions on Pattern Analysis and.
Segmentation Divide the image into segments. Each segment:
Multiple Human Objects Tracking in Crowded Scenes Yao-Te Tsai, Huang-Chia Shih, and Chung-Lin Huang Dept. of EE, NTHU International Conference on Pattern.
Object Detection and Tracking Mike Knowles 11 th January 2005
1 Integration of Background Modeling and Object Tracking Yu-Ting Chen, Chu-Song Chen, Yi-Ping Hung IEEE ICME, 2006.
CSE 291 Final Project: Adaptive Multi-Spectral Differencing Andrew Cosand UCSD CVRR.
Effective Gaussian mixture learning for video background subtraction Dar-Shyang Lee, Member, IEEE.
Performance Evaluation in Computer Vision Kyungnam Kim Computer Vision Lab, University of Maryland, College Park.
Multi-camera Video Surveillance: Detection, Occlusion Handling, Tracking and Event Recognition Oytun Akman.
Jeremy Wyatt Thanks to Gavin Brown
Smart Traveller with Visual Translator for OCR and Face Recognition LYU0203 FYP.
Shadow Detection In Video Submitted by: Hisham Abu saleh.
Jacinto C. Nascimento, Member, IEEE, and Jorge S. Marques
Today Concepts underlying inferential statistics
Statistical Color Models (SCM) Kyungnam Kim. Contents Introduction Trivariate Gaussian model Chromaticity models –Fixed planar chromaticity models –Zhu.
VINCENT URIAS, CURTIS HASH Detection of Humans in Images Using Skin-tone Analysis and Face Detection.
Noise Estimation from a Single Image Ce Liu William T. FreemanRichard Szeliski Sing Bing Kang.
1 Activity and Motion Detection in Videos Longin Jan Latecki and Roland Miezianko, Temple University Dragoljub Pokrajac, Delaware State University Dover,
Prakash Chockalingam Clemson University Non-Rigid Multi-Modal Object Tracking Using Gaussian Mixture Models Committee Members Dr Stan Birchfield (chair)
BraMBLe: The Bayesian Multiple-BLob Tracker By Michael Isard and John MacCormick Presented by Kristin Branson CSE 252C, Fall 2003.
Perception Introduction Pattern Recognition Image Formation
Background Subtraction for Temporally Irregular Dynamic Textures Gerald Dalley, Joshua Migdal, and W. Eric L. Grimson Workshop on Applications of Computer.
1 Lucia Maddalena and Alfredo Petrosino, Senior Member, IEEE A Self-Organizing Approach to Background Subtraction for Visual Surveillance Applications.
A Comparative Evaluation of Three Skin Color Detection Approaches Dennis Jensch, Daniel Mohr, Clausthal University Gabriel Zachmann, University of Bremen.
Ground Truth Free Evaluation of Segment Based Maps Rolf Lakaemper Temple University, Philadelphia,PA,USA.
Hierarchical Method for Foreground DetectionUsing Codebook Model Jing-Ming Guo, Yun-Fu Liu, Chih-Hsien Hsia, Min-Hsiung Shih, and Chih-Sheng Hsu IEEE TRANSACTIONS.
Stable Multi-Target Tracking in Real-Time Surveillance Video
D. M. J. Tax and R. P. W. Duin. Presented by Mihajlo Grbovic Support Vector Data Description.
Tracking and event recognition – the Etiseo experience Son Tran, Nagia Ghanem, David Harwood and Larry Davis UMIACS, University of Maryland.
Expectation-Maximization (EM) Case Studies
Computer and Robot Vision II Chapter 20 Accuracy Presented by: 傅楸善 & 王林農 指導教授 : 傅楸善 博士.
Boosted Particle Filter: Multitarget Detection and Tracking Fayin Li.
Introduction to Pattern Recognition (การรู้จํารูปแบบเบื้องต้น)
A Reliable Skin Detection Using Dempster-Shafer Theory of Evidence
GENDER AND AGE RECOGNITION FOR VIDEO ANALYTICS SOLUTION PRESENTED BY: SUBHASH REDDY JOLAPURAM.
Chapter 13 (Prototype Methods and Nearest-Neighbors )
Presented by: Idan Aharoni
Visual Tracking by Cluster Analysis Arthur Pece Department of Computer Science University of Copenhagen
Suspicious Behavior in Outdoor Video Analysis - Challenges & Complexities Air Force Institute of Technology/ROME Air Force Research Lab Unclassified IED.
The Nested Dirichlet Process Duke University Machine Learning Group Presented by Kai Ni Nov. 10, 2006 Paper by Abel Rodriguez, David B. Dunson, and Alan.
ICCV 2007 Optimization & Learning for Registration of Moving Dynamic Textures Junzhou Huang 1, Xiaolei Huang 2, Dimitris Metaxas 1 Rutgers University 1,
Learning and Removing Cast Shadows through a Multidistribution Approach Nicolas Martel-Brisson, Andre Zaccarin IEEE TRANSACTIONS ON PATTERN ANALYSIS AND.
Shadow Detection in Remotely Sensed Images Based on Self-Adaptive Feature Selection Jiahang Liu, Tao Fang, and Deren Li IEEE TRANSACTIONS ON GEOSCIENCE.
Motion Estimation of Moving Foreground Objects Pierre Ponce ee392j Winter March 10, 2004.
Ch 1. Introduction Pattern Recognition and Machine Learning, C. M. Bishop, Updated by J.-H. Eom (2 nd round revision) Summarized by K.-I.
Video object segmentation and its salient motion detection using adaptive background generation Kim, T.K.; Im, J.H.; Paik, J.K.;  Electronics Letters 
Traffic Sign Recognition Using Discriminative Local Features Andrzej Ruta, Yongmin Li, Xiaohui Liu School of Information Systems, Computing and Mathematics.
A Forest of Sensors: Using adaptive tracking to classify and monitor activities in a site Eric Grimson AI Lab, Massachusetts Institute of Technology
LECTURE 05: THRESHOLD DECODING
Image Segmentation Techniques
Eric Grimson, Chris Stauffer,
PRAKASH CHOCKALINGAM, NALIN PRADEEP, AND STAN BIRCHFIELD
Where did we stop? The Bayes decision rule guarantees an optimal classification… … But it requires the knowledge of P(ci|x) (or p(x|ci) and P(ci)) We.
Estimation of Skin Color Range Using Achromatic Features
Presentation transcript:

University of MarylandComputer Vision Lab 1 A Perturbation Method for Evaluating Background Subtraction Algorithms Thanarat Horprasert, Kyungnam Kim, David Harwood, Larry Davis Computer Vision Lab, UMIACS, Univ.of Maryland at College Park Oct 12, 2003

University of MarylandComputer Vision Lab 2 Contents l Introduction to Background Subtraction (BGS) l BGS Algorithms l Classical ROC Analysis l Perturbation Detection Rate Analysis l Experimental Results l Conclusion and Future Work l Introduction to Background Subtraction (BGS) l BGS Algorithms l Classical ROC Analysis l Perturbation Detection Rate Analysis l Conclusion and Future Work l Experimental Results

University of MarylandComputer Vision Lab 3 Introduction to Background Subtraction (BGS) l The capability of extracting moving objects from a video sequence captured using a static camera is a typical first step in visual surveillance. l The idea of background subtraction is to subtract or difference the current image from a reference background model.

University of MarylandComputer Vision Lab 4 Introduction to Background Subtraction (BGS)

University of MarylandComputer Vision Lab 5 BGS Algorithms l Unimodal distribution  The simplest background model assumes that the intensity values of a pixel can be modeled by a unimodal distribution, like a Gaussian distribution, N(μ,σ 2 ) [Wren et al.(1997), Horprasert et al.(1999)]. l Mixture of Gaussians (MOG)  The generalized MOG has been used to model complex, non- static multiple backgrounds [Stauffer & Grimson (1999), Harville (2002)].  Modified/advanced versions are widely used among the research community. (Disadvantages) A few Gaussians cannot accurately model background having fast variations. Depending on the learning rate, it faces trade-off problems.

University of MarylandComputer Vision Lab 6 BGS Algorithms (cont.) l Non-parametric technique  Estimating the probability density function at each pixel from many samples using Kernel density estimation technique [Elgammal et al. (2000)].  It is able to adapt very quickly to changes in the background process and to detect targets with high sensitivity.  Cannot be used when long-time periods are needed to sufficiently sample the background due mostly to memory constraints. l Region- or frame based approach  Pixel-based techniques assume that the time series of observation is independent at each pixel.  High-level approach by segmenting an image into regions or by refining low-level classification obtained at the pixel level [Toyama (1999), Harville (2002), Cristani et al. (2002)].

University of MarylandComputer Vision Lab 7 BGS Algorithms (cont.) l Codebook-based technique (new)  We adopt a quantization and clustering technique motivated by Kohonen to construct the background model from long observation sequences, without making parametric assumptions.  For each pixel, a codebook (CB) consists of one or more codewords. Mixed backgrounds can be modeled by multiple codewords.  Samples at each pixel are clustered into the set of codewords based on a color distortion metric together with a brightness ratio.

University of MarylandComputer Vision Lab 8 BGS Algorithms (cont.) l Four algorithms in evaluation NameBackground subtraction algorithm CBcodebook-based technique in the paper MOG mixture of Gaussians by Stauffer & Grimson (1999) KER and KER.RGB* non-parametric method using Kernels by Elgammal et al. (2000). UNI unimodal background modeling by Horprasert et al.(1999). * The algorithm accepts normalized colors (KER) and RGB colors (KER.RGB) as inputs

University of MarylandComputer Vision Lab 9 Classical ROC Analysis l Performance evaluation is required in terms of how well the algorithm detects the targets with less false alarms. l ROC (Receiver Operating Characteristic) Analysis.  Applied when there are known background(BG) and foreground(FG) distributions.  Requires (hand-segmented) ground truth for analysis.  Evaluation is centralized around the tradeoff of ‘miss detection rate’ and ‘false alarm rate’.

University of MarylandComputer Vision Lab 10 Classical ROC Analysis (cont.) l True Negative: when BG is classified correctly as the BG. l True Positive: when FG is classified correctly as the FG. l False Negative: when FG is classified incorrectly as the BG. l False Positive: when BG is classified incorrectly as the FG.

University of MarylandComputer Vision Lab 11 Classical ROC Analysis (cont.) l Detection errors can be classified into 2 types:  False alarm rate (FA-rate) = FP / (FP+TN)  Miss detection rate (MD-rate) = FN / (FN + TP) RROC curve algo.1 algo.2 better

University of MarylandComputer Vision Lab 12 Classical ROC Analysis (cont.) l Limitations of ROC:  In typical video surveillance applications, we are usually given a BG scene for a fixed camera, but do not or cannot know what might possibly move in the scene as FG objects.  Requires manual groundtruth evaluation.  Measures the errors for detecting a particular FG against a particular BG. There are as many ROC curves as possible different FG targets.

University of MarylandComputer Vision Lab 13 Perturbation Detection Rate Analysis l Perturbation Detection Rate (PDR) analysis measures the sensitivity of the algorithm in detecting low contrast targets against background as a function of contrast l Without knowledge of the actual FG distribution. l Assumption:  The shape of the FG distribution is locally similar to that of the BG distribution.  However, FG distribution of small contrast will be a shifted or perturbed version of the BG distribution.

University of MarylandComputer Vision Lab 14 PDR Analysis (cont.) Given the parameters to achieve a certain fixed FA-rate, the analysis is performed by shifting or perturbing the entire BG distributions by vectors in uniformly random directions of RGB space with fixed magnitude , computing an average detection rate as a function of contract .

University of MarylandComputer Vision Lab 15 PDR Analysis (cont.) l Procedure to produce a PDR graph: 1.Train N training (empty) frames, adjusting parameters to achieve a target FA-rate (practically.01% to 1% ). 2.Pass through those N frames again to obtain a test FG. For each frame, perturb a random sample of M pixel values ( R i, G i, B i ) by a magnitude  in uniformly random directions. ( R’ i, G’ i, B’ i ) = ( R i, G i, B i) ) + ( dR, dG, dB ) 3.Test the BGS algorithms on these perturbed FG pixels and compute the detection rate for the  4.By varying the FG contrast , obtain a monotone increasing PDR graph of detection rates. ( R i, G i, B i) ) ( R’ i, G’ i, B’ i ) 3D color sphere with radius  

University of MarylandComputer Vision Lab 16 Experimental Results l Configuration  Training frame: 100 empty consecutive frames from each video.  For each frame, 1000 points are randomly selected for perturbation  During testing, no updating of the BG model is allowed.  KER and KER.RGB: a sample size 50 (frames) represents the BG.  MOG: the max # of Gaussians is 4 for stationary BGs and 10 for moving backgrounds. The learning rate  is fixed and T is adjusted to give the desired FA-rate.  The FA-rate for each video is determined by (1) Video quality, (2) whether it is indoor or outdoor, and (3) good real FG detection results for most algorithms.

University of MarylandComputer Vision Lab 17 Experimental Results l Useful for choosing particular algorithm’s parameter values for use in a given application.  Shows trade-off between different parameters. RROC curve of CB algorithm 

University of MarylandComputer Vision Lab 18 Experimental Results l Indoor office video (mostly stationary BG) l MOG and KER.RGB don’t separately model brightness and color. l MOG does not model covariance (caused by variation in brightness) better worse

University of MarylandComputer Vision Lab 19 Experimental Results l Outdoor woods video (containing moving BG) l All algorithms perform somewhat worse. l UNI does not perform well as in the indoor case (not designed for outdoors).

University of MarylandComputer Vision Lab 20 Experimental Results l Sensitive detection in a real example l A red sweater against a reddish colored wall with difference (at the missing spots). l The graphs shows a large difference in detection rate.  CBMOG

University of MarylandComputer Vision Lab 21 Experimental Results l A window containing mostly moving BG l The FA-rate obtained only within the window. l Sample size of KER: 270 l Reduced sensitivity of all algorithms

University of MarylandComputer Vision Lab 22 Conclusion and Future Work l Summary of PDR (as alternative to classical ROC analysis):  does not require FG targets in videos or knowledge of actual FG distributions  assume that the FG has a distribution similar in form to BG, but shifted or perturbed.  applied to 4 representative algorithms on 4 videos, showing understandable results  useful for qualitative comparison of different algorithms as well as comparison of choice of parameters for a particular algorithms.

University of MarylandComputer Vision Lab 23 Conclusion and Future Work l Limitation:  Does not model motion blur of moving FG objects  In the case of mixed (moving) BG, the simulated FG distribution will be mixed (as plants or flags moving in the FG).  FG objects often have shading and reflection effects on BG. They are important for choosing a proper, practical false alarm rate.

University of MarylandComputer Vision Lab 24 Conclusion and Future Work l Future work  Extended to measure local detection rates throughout the frame of the scene or varying over time.  localized parameter estimation  PDR on the videos containing FG already.  Area of non-detection (PDR- II ?): measure the size of the area covered by the decision surface of the BG model at a certain false alarm rate.