Target Response Adaptation for Correlation Filter Tracking

Similar documents
Understanding the Spike Algorithm

Adaptive Modulation for Multiple Antenna Channels

Dynamic Optimization. Assignment 1. Sasanka Nagavalli January 29, 2013 Robotics Institute Carnegie Mellon University

Walsh Function Based Synthesis Method of PWM Pattern for Full-Bridge Inverter

problems palette of David Rock and Mary K. Porter 6. A local musician comes to your school to give a performance

Fast Code Detection Using High Speed Time Delay Neural Networks

A Comparison of Two Equivalent Real Formulations for Complex-Valued Linear Systems Part 2: Results

Decomposition Principles and Online Learning in Cross-Layer Optimization for Delay-Sensitive Applications

The Detection Algorithms Performance in BLAST Enhanced IEEE a WLAN Standard on Measured Channels. University of Bristol

Passive Filters. References: Barbow (pp ), Hayes & Horowitz (pp 32-60), Rizzoni (Chap. 6)

Graph Method for Solving Switched Capacitors Circuits

Research of Dispatching Method in Elevator Group Control System Based on Fuzzy Neural Network. Yufeng Dai a, Yun Du b

DIMENSIONAL SYNTHESIS FOR WIDE-BAND BAND- PASS FILTERS WITH QUARTER-WAVELENGTH RES- ONATORS

CS345a: Data Mining Jure Leskovec and Anand Rajaraman Stanford University

Generalized Incomplete Trojan-Type Designs with Unequal Cell Sizes

Chinese Remainder. Discrete Mathematics Andrei Bulatov

29. Network Functions for Circuits Containing Op Amps

Source Localization by TDOA with Random Sensor Position Errors - Part II: Mobile sensors

Learning Ensembles of Convolutional Neural Networks

NOVEL ITERATIVE TECHNIQUES FOR RADAR TARGET DISCRIMINATION

Rejection of PSK Interference in DS-SS/PSK System Using Adaptive Transversal Filter with Conditional Response Recalculation

Multiband Jamming Strategies with Minimum Rate Constraints

IEE Electronics Letters, vol 34, no 17, August 1998, pp ESTIMATING STARTING POINT OF CONDUCTION OF CMOS GATES

Optimal Periodic Training Signal for Frequency Offset Estimation in Frequency Selective Fading Channels

NMR Part II, Quadrature Detection

ALICE AND BOB GO TO DINNER: A VARIATION ON MÉNAGE

Location of Rescue Helicopters in South Tyrol

To: Professor Avitabile Date: February 4, 2003 From: Mechanical Student Subject: Experiment #1 Numerical Methods Using Excel

Integer Programming. P.H.S. Torr Lecture 5. Integer Programming

PRACTICAL, COMPUTATION EFFICIENT HIGH-ORDER NEURAL NETWORK FOR ROTATION AND SHIFT INVARIANT PATTERN RECOGNITION. Evgeny Artyomov and Orly Yadid-Pecht

Phasor Representation of Sinusoidal Signals

Recognition of Low-Resolution Face Images using Sparse Coding of Local Features

Two-Stage Detector for SC-FDMA Transmission over MIMO ISI Channels

Multi-sensor optimal information fusion Kalman filter with mobile agents in ring sensor networks

Network Theory. EC / EE / IN. for

A Robust Adaptive Carrier Frequency Offset Estimation Algorithm for OFDM

Optimal Allocation of Static VAr Compensator for Active Power Loss Reduction by Different Decision Variables

A Lower Bound for τ(n) of Any k-perfect Numbers

Unit 1. Current and Voltage U 1 VOLTAGE AND CURRENT. Circuit Basics KVL, KCL, Ohm's Law LED Outputs Buttons/Switch Inputs. Current / Voltage Analogy

Piecewise Linear Approximation of Generators Cost Functions Using Max-Affine Functions

Exploiting Dynamic Workload Variation in Low Energy Preemptive Task Scheduling

ANNUAL OF NAVIGATION 11/2006

UNIT 11 TWO-PERSON ZERO-SUM GAMES WITH SADDLE POINT

Comparison of Gradient descent method, Kalman Filtering and decoupled Kalman in training Neural Networks used for fingerprint-based positioning

Intensity- modulated radiotherapy. Beyond fluence map op.miza.on

arxiv: v1 [astro-ph.im] 24 Apr 2015

Full-duplex Relaying for D2D Communication in mmwave based 5G Networks

An Effective Approach for Distribution System Power Flow Solution

Joint Scheduling and Resource Allocation in Uplink OFDM Systems for Broadband Wireless Access Networks

MIMO-OFDM Systems. Team Telecommunication and Computer Networks, FSSM, University Cadi Ayyad, P.O. Box 2390, Marrakech, Morocco.

Block Diagonal Precoding Based Power Allocation for Coordinated Multi-Point Transmission

2522 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO. 6, JUNE TDOA Based Positioning in the Presence of Unknown Clock Skew

Figure 1. DC-DC Boost Converter

Enhanced Artificial Neural Networks Using Complex Numbers

Test 2. ECON3161, Game Theory. Tuesday, November 6 th

Throughput Maximization by Adaptive Threshold Adjustment for AMC Systems

The Spectrum Sharing in Cognitive Radio Networks Based on Competitive Price Game

Harmonic Modeling of Inrush Current in Core Type Power Transformers using Hartley Transform

Optimal Decentralized Kalman Filter

How to Scale Up the Spectral Efficiency of Multi-way Massive MIMO Relaying?

Performance Study of OFDMA vs. OFDM/SDMA

Subcarrier allocation for OFDMA wireless channels using lagrangian relaxation methods

Medium Term Load Forecasting for Jordan Electric Power System Using Particle Swarm Optimization Algorithm Based on Least Square Regression Methods

RECENT years have witnessed a great interest in cognitive

Cod and climate: effect of the North Atlantic Oscillation on recruitment in the North Atlantic

Maximizing Lifetime of Sensor-Target Surveillance in Wireless Sensor Networks

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 62, NO. 8, AUGUST

Introduction. (Received 08 January 2009; accepted 10 March 2009)

Joint Data and Power Transfer Optimization for Energy Harvesting Mobile Wireless Networks

Joint Adaptive Modulation and Power Allocation in Cognitive Radio Networks

NETWORK 2001 Transportation Planning Under Multiple Objectives

AN EFFICIENT ITERATIVE DFT-BASED CHANNEL ESTIMATION FOR MIMO-OFDM SYSTEMS ON MULTIPATH CHANNELS

Research Article Semidefinite Relaxation Algorithm for Multisource Localization Using TDOA Measurements with Range Constraints

Outlier-Tolerant Kalman Filter of State Vectors in Linear Stochastic System

1 GSW Multipath Channel Models

Optimal Placement of PMU and RTU by Hybrid Genetic Algorithm and Simulated Annealing for Multiarea Power System State Estimation

DETERMINATION OF WIND SPEED PROFILE PARAMETERS IN THE SURFACE LAYER USING A MINI-SODAR

A MODIFIED DIFFERENTIAL EVOLUTION ALGORITHM IN SPARSE LINEAR ANTENNA ARRAY SYNTHESIS

NATIONAL RADIO ASTRONOMY OBSERVATORY Green Bank, West Virginia SPECTRAL PROCESSOR MEMO NO. 25. MEMORANDUM February 13, 1985

EE 508 Lecture 6. Degrees of Freedom The Approximation Problem

Network Reconfiguration in Distribution Systems Using a Modified TS Algorithm

RESOURCE CONTROL FOR HYBRID CODE AND TIME DIVISION SCHEDULING

Joint Rate-Routing Control for Fair and Efficient Data Gathering in Wireless sensor Networks

Planning of Relay Station Locations in IEEE (WiMAX) Networks

Spectrum Management and Power Allocation in. MIMO Cognitive Networks

Machine Learning in Production Systems Design Using Genetic Algorithms

Vectorless Analysis of Supply Noise Induced Delay Variation

A Novel Optimization of the Distance Source Routing (DSR) Protocol for the Mobile Ad Hoc Networks (MANET)

Desensitized Kalman Filtering with Analytical Gain

Multi-focus Image Fusion Using Spatial Frequency and Genetic Algorithm

Coordinating the Motions of Multiple Robots with Specified Trajectories

A Novel GNSS Weak Signal Acquisition Using Wavelet Denoising Method

Iterative Tuning Strategy for Setting Phase Splits in Traffic Signal Control

ECE315 / ECE515 Lecture 5 Date:

Side-Match Vector Quantizers Using Neural Network Based Variance Predictor for Image Coding

A NSGA-II algorithm to solve a bi-objective optimization of the redundancy allocation problem for series-parallel systems

Distributed and Optimal Reduced Primal-Dual Algorithm for Uplink OFDM Resource Allocation

Opportunistic Beamforming for Finite Horizon Multicast

CDMA Uplink Power Control as a Noncooperative Game

Digital Transmission

Transcription:

Target Response Adaptaton for Correlaton Flter Tracng Adel Bb, Matthas Mueller, and Bernard Ghanem Image and Vdeo Understandng Laboratory IVUL, Kng Abdullah Unversty of Scence and Technology KAUST, Saud Araba {adel.bb,matthas.mueller.,bernard.ghanem}@aust.edu.sa In ths supplementary materal, we followup our dscusson on the soluton to the followng problem: mn w,y Aw y + λ w + λ y y 0 The problem s convex quadratc and a statonary pont s necessary and suffcent for global optmalty. The followng sectons wll dscuss how Problem s solved n the prmal doman, dual doman, and for both sngle and multple templates along wth the formula used to generate the response map, whose maxmum value determnes the current detecton. Lastly, we dscuss a one way of ncorporatng SRDCF wth our proposed target adaptve framewor. Soluton to Problem n the Prmal Doman. Usng a Sngle Template Problem can be rewrtten n terms of z wth z T w T y T : fz A I z + λ I 0 z + λ 0 I z y 0, where w R n, A R n n, y R n, y 0 R n, z R n, then: A z fz T A A T I 0 0 0 0 z + λ A I z + λ 0 0 z λ 0 I y I 0 0 A z fz T A + λ I A T 0 z λ A + λ I y I 0 F 0 dagâ â + λ dagâ F H 0 0 0 F dagâ dag + λ 0 F H z λ y I 0 dagâ â + λ dagâ ŵ 0 dagâ dag + λ ŷ λ F H y 0 ŵ dagâ â ŷ λ + λ dagâ 0 y dagâ dag + λ 0 F H

Adel Bb, Matthas Mueller, and Bernard Ghanem Note that the nverse lemma states: B N B NC V B NC V NC V C C VB NC V C VB NC V NC + C 3 B NC V dagâ â + λ dagâ dag + λ dagâ B NC V + λ dag λ â â + λ + λ B NC V NC â dag λ â â + λ + λ Snce: ŵ ŵ λ B NC V NC F H y o 4 λ â ŷo λ â â + λ + λ ŵ λ â ŷ o λ â â + λ + λ 5 Detecton Formula wth a Sngle Template As for the detecton formula n the prmal doman, we consder a new test sample u. For detecton, we construct all the crcular shfts of u n matrx U. Therefore, the response map on the test sample s: Tu Uw FdagûF H w ˆT u û ŵ ˆTu û ŵ 6. Usng Multple Templates Followng s the dervaton of the soluton for the multple template case n prmal doman,.e. when à Rn n, s the total number of templates, and à T A T A T... A T and Ĩ T I I... I. fz à Ĩ z + λ I 0 z + λ 0 I z y 0 fz ÃT à ÃT Ĩ I 0 0 0 0 z + λ z + λ ĨT à ĨT Ĩ 0 0 z λ 0 I y I 0 0 ÃT à + λ fz I ÃT Ĩ 0 z λ y ĨT à + λ I I 0 }{{} Γ

Target Response Adaptaton for Correlaton Flter Tracng 3 Now, note the followng: à T à AT A, ĨT Ĩ IT I I, and à T Ĩ AT, and that ĨT à A. It s clear that ÃT à and ĨT Ĩ are crculant whch s a sum of crculant matrces. Therefore, the matrx Γ s bloc wse crculant such that Γ R n n. Smlar to the sngle template case, we have: w ỹ λ dagâ â + λ dagâ 0 dagâ dag + λ F H y 0 7 Usng the nverse lemma n Eq 3 on Γ, we get: B NC V dag â â + λ dag â dag â + λ 8 B NC V NC dag â + λ â aˆ + + λ λ â â ŵ λ ˆ a ŷ o + λ â aˆ + + λ λ â â 9 0 It s to be noted that when, then Eq 0 reduces to the sngle template case n Eq 5, snce â â â â for. Detecton Formula wth Multple Templates Here, the detecton formula s smlar to Eq 6. Soluton to Problem n the Dual Doman. Usng a Sngle Template The optmzaton problem becomes: fz A I z + λ I 0 z + λ 0 I z y 0 For smpler notaton, let G A I, also let E I 0, and D 0 I fz Gz + λ Ez + λ Dz y 0 λ z T d y 0 + λ Ez + Gz

4 Adel Bb, Matthas Mueller, and Bernard Ghanem z fz λ z T d y 0 d + λ E T Ez + G T Gz 0 λ z T d y 0 d λ E T E + G T G z z λ λ E T E + G T G z T d y 0 d Let α λ z T d y 0. Then z λ E T E + G T G. By substtutng the dual varables α nto Eq, we obtan: fα λ λ E T E + G T G D T α. Let K α T DK d y 0 + λ EK D T α + GK D T α Then, the new dual objectve s gven by: 3 fα λ DK D T α y 0 + λ EK D T α + GK D T α 4 By settng the gradent to zero, the soluton to Problem 4 s obtaned by solvng the followng lnear system: λ DK D T DK D T + λ DK E T EK D T + DK G T GK D T α λ DK D T y 0 DK λ D T D + λ E T E + G T G K D T α λ DK D T y 0 }{{} 5 Ψ A D T A + λ I A T A T A + λ I A T A T A + λ I A T A I A + λ I A I }{{} D T A α λ D T A + λ I A T D T y A I 0 6 By usng the nverse lemma and the dagonalzaton propertes of crculant matrces smlar to nvertng the crculant dagonal matrx n the sngle template Ψ

Target Response Adaptaton for Correlaton Flter Tracng 5 case, one can show the followng: A K T A + λ I A T A I K λ I λ A T 7 λ A λ AA T + I K ΨK λ I λ A T I 0 λ λ AA T λ λ A T + + λ I λ A λ AA T + I λ I + λ A T +λ A λ λ A T + λ A T AA T λ λ AA T A + +λ λ λ A λ AA T AA T + +λ λ λ AA T + + λ I 8 Therefore, by substtutng Eq 8 nto Eq 6, we get the lnear system: λ λ AA T AA T + + λ AA T + + λ I α λ AA T + I y 0 9 λ λ The prevous problem can be effcently dagonalzed and solved as follows: λ F λ dagâ â â â + + λ dagâ â λ + + λ I ˆα λ F dagâ â λ + λ I ŷo λ ˆα λ â â + λ ŷo λ â λ â â â + +λ λ â â + + λ λ λ â â + λ ŷ o ˆα λ â λ â â â + +λ λ â â + + λ 0 Detecton Formula wth a Sngle Template Tu Uw UEK D T α U I 0 λ I λ A T 0 λ A λ AA T + I UA T α α λ λ Fdagû â ˆα ˆTu λ û â ˆα

6 Adel Bb, Matthas Mueller, and Bernard Ghanem. Usng Multple Templates For multple templates, the only dfference s that G T Therefore, the new dual objectve s gven as follows: where A T A T... A T I I... I D K λ D T D + λ E T E + G T G K D T α λ D K D T y 0 K AT A + λ I A T A I F 0 K dagâ â + λ I F dagâ 0 F H 0 dagâ I 0 F H 3 4 To fnd K, we use the nverse lemma n Eq 3 agan. Frst, we need to fnd: B NC V: B NC V dagâ â + λ I dagâ dagâ dag â â + λ I dag â dag â dag â â + λ dag â dag â â + λ â â â Then, usng the nverse lemma 3 and smlar trcs as used for tranng the flter n the sngle template case n the prmal doman, the nverse of K s gven as follows: K F 0 F Λ H 0 0 F 0 F H where dag Λ dag â â+λ â â â â â+λ â â â dag dag â â+λ â â â â + â â+λ â â 5

Target Response Adaptaton for Correlaton Flter Tracng 7 And snce: Ψ λ D T D + λ E T E + G T G AT A + λ I A T A + λ I F 0 0 F Ω F H 0 0 F H where Ω dagâ â + λ I dagâ dagâ + λ I Then, K Ψ K dag ΛΩ dag0 F 0 F ΛΩΛ H 0 0 F 0 F H dag dag λ â â â+λ â â â â+ +λ â â+ λ +λ â â+λ â â dag ΛΩΛ dag0 dag dag λ â â â+λ â â â â+ +λ â â+ λ +λ â â+λ â â dag dag â â+λ â â â â â+λ â â â dag dag â â+λ â â â â + â â+λ â â Υ Υ Υ Υ

8 Adel Bb, Matthas Mueller, and Bernard Ghanem Note that to to compute D K Ψ K D T, only the last bloc Υ s relevant. So, D K Ψ K D T FdagΥ F H, where λ Υ â + â â â â â â â + λ â â â â â â + λ â â + +λ â â â â + λ â â + + λ â â â â + λ â â + 6 Wth some mathematcal manpulaton Eq 6 s exactly equvalent to Eq 7 n the paper. As for the rght hand sde of Eq, one need to fnd: D K D T Fdag â â â â + λ â â + Therefore, soluton to Problem s computed as follows: FdagΥ F H α λ F â â â â + λ â â + F H 7 F H y o 8 ˆα λ dag Υ â â â â + λ ŷ 0 â ŷ0 + â 9 Detecton Formula wth Multple Templates Tu Uw UE K D T û α F â ˆα â â + λ â â ˆTu û â ˆα â â + λ â â 30

Target Response Adaptaton for Correlaton Flter Tracng 9 3 Integratng SRDCF Instead of re-dervng a closed form soluton to SRDCF s formulaton, one can use alternatng optmzaton, where the flter w and the target response y are updated n an teratve fashon eepng one of them fxed at any gven tme. The target adapted SRDCF objectve becomes: mn w,y Aw y + λ Xw + λ y y 0, 3 where X s the weght mas over the flter. Usng alternatng optmzaton, t can be mnmzed by solvng the next two optmzaton problems at teraton j. w j+ arg mn Aw y j + λ Xw 3 w y j+ arg mn Aw j+ y + λ y y 0 33 y Problem 3 can be solved usng the standard SRDCF soluton, whle Problem 33 has a closed form soluton as follows: y j+ + λ Aw j+ + λ y o 34 To compute y j+ effcently n the Fourer doman, we can use the followng strategy: y j+ Fdagâ F H w j+ + λ y o + λ 35 Fdagâ ŵ j+ + λ y o + λ 36 ŷ dagâ ŵ j+ + λ ŷ + λ o 37 â ŵ j+ + λ ŷ + λ o 38 Acnowledgments. Research n ths publcaton was supported by the Kng Abdullah Unversty of Scence and Technology KAUST Offce of Sponsored Research.

0 Adel Bb, Matthas Mueller, and Bernard Ghanem References. Bb, A., Mueller, M., Ghanem, B.: Target response adaptaton for correlaton flter tracng. In: European Conference on Computer Vson, ECCV, 06.. Danelljan, M., Hager, G., Shahbaz Khan, F., Felsberg, M.: Learnng spatally regularzed correlaton flters for vsual tracng. In: IEEE Internatonal Conference on Computer Vson, ICCV, 05.