Thrust Technical Review March 2012 Distributed Sensor Exploitation 6.1 Rapid Deployment of Smart Camera Networks Dr. B.S. Manjunath University of California, Santa Barbara manj@ece.ucsb.edu 805-893-7112 Distributed Sensor Exploitation 6.1 Rapid Deployment of Smart Camera Networks 2012 Thrust Area Technical Review UAV: Provides intelligence from overhead RESEARCH QUESTIONS: • How can we geo-calibrate a stationary network using mobile sensors and associated GPS and other information? • What are the issues in discovering the geometric/spatial and temporal connections and constraints within a camera network with overlapping and non-overlapping views? • What information needs to be computed at the smart sensor nodes for activity analysis without distributing the raw video? • When and how to fuse multi-modal information, e.g., mobile, airborne, and fixed cameras, for robust object tracking and activity analysis? MILITARY RELEVANCE/OPERATIONAL IMPACT: Rapid deployment of a network of smart cameras and UAVs can help monitor any potential area, especially in high risk and remote sites. Rapidly Deployed Smart Camera: Provides ground-based views and distributed image analysis TECHNICAL APPROACH: Geo-calibration of fixed cameras using GPS and other sensors on smart phones. Auto-calibration and network topology discovery using visual analysis Detecting and tracking objects in a camera network through fusion of multiple views and geometry. Analyzing events over non-overlapping camera views, with emphasis on distributed tracking and fusion of tracked events spatially and temporally. NAVAL S&T FOCUS AREAs ADDRESSED: Asymmetric & irregular warfare SCHEDULE: TASKS FY10 FY11 Camera network setup Auto-calibration and topology Object detection & 2D tracking Object tracking in 3D Event & behavior detection PERFORMER: University of California at Santa Barbara B. S. Manjunath, Professor Demos on campus testbed TTA: N/A TECH TRANSITION PATH: DYNAMIC WIKI/RESOURCE MANAGER (D&I) FY12 Top Level POA&M (Summary Task) 2012 Thrust Area Technical Review Principal Events & Activities FY2010 FY2011 FY2012 O N D J F M A M J J A S O N D J F M A M J J A S O N D J F M A M J J A S Camera Network Setup Auto Calibration & Toplology Geo-calibration Object Detection and 2D tracking Object Tracking in 3D Event/Behavior analysis Demonstration on DURIP Testbed Field deployment Peer Review presentation Final demonstration 3 Project Work Breakdown Structure 2012 Thrust Area Technical Review Detailed Technical Approach 2012 Thrust Area Technical Review • Objective – Development of a rapidly deployable network of smart camera sensors to gather mission-specific information for surveillance and human activity analysis • Challenges – Minimize human interaction – Wireless communication constraints: limited bandwidth, unpredictable latency, dropped packets, and time synchronization – Few examples to train the system 5 Detailed Technical Approach 2012 Thrust Area Technical Review Project Roadmap Events in FY 10 Setup of Outdoor Camera Network Quick Chessboard Calibration Preliminary network topology discovery Events in FY11 Network topology discovery Automatic calibration without patterns Beyond 2D tracking Continued multi-camera tracking with overlapping and non-overlapping views. Preliminary distributed processing with existing work Events in FY12 • Geo-calibration with mobile sensors • Multi-camera multi-object tracking in a wide-area network • Browsing and search in a distributed network 6 Detailed Technical Approach 2012 Thrust Area Technical Review • Project Introduction • Overview of the UCSB Campus Camera Network – Infrastructure funded by a ONR DURIP award – Implementation by students on this project • Research Accomplishments – – – – Distributed Geo-calibration Distributed tracking and fusion Topology discovery and network data summarization Browsing and search in a distributed network • Summary and Conclusions 7 Research Emphasis 2012 Thrust Area Technical Review Geo-calibration Topology discovery Wide-area activity analysis Network summarization Camera placement N+1 camera calibration Object searching Distributed object tracking UCSB Camera Network 2012 Thrust Area Technical Review • Examined challenges through manual network setup • Developing optimal camera placement algorithm based on GPS trajectories 11 Project Technical Assesment 2012 Thrust Area Technical Review Recent Progress New Geo-calibration formulation with mobile sensors Distributed Object Tracking on a Camera Network Video Summarization in a Multi-camera setting Distributed browsing and search Collection of ground based and UAV (quadcopter) data for activity analysis 12 Geo-calibration 2012 Thrust Area Technical Review How do you calibrate cameras with non-overlapping views? 15 Geo-Calibration 2012 Thrust Area Technical Review • Geo-Calibration is the process of finding the projective geometry of a camera in a global coordinate system. • Traditional, image-based methods are in a local coordinate system and cannot be shared across different calibrations1 • Existing methods for integrating global coordinates involve GPS tagged objects in the scene or aligning with a map2 1. Zhang, “A flexible new technique for camera calibration,” PAMI 2000 2. Kaminsky et al. “Alignment of 3D Point Clouds to Overhead Images.” W on Internet Vision 2009. 16 Geo-Calibration Method 2012 Thrust Area Technical Review • Take pictures of the camera’s scene using smartphones and collect location and orientation metadata – Location sensor: GPS – Orientation sensor: compass + accelerometer or gyroscope • No objects need to be placed in the scene • Data can be collected rapidly 17 Geo-Calibration Method 2012 Thrust Area Technical Review • Geo-calibration: Given image(s) from a camera 1. Collect, using smartphone, calibration images with GPS and orientation metadata 2. Calibrate cameras in a local coordinate system using images 3. Transform from local to global GPS coordinates 4. Refine position and orientation estimates • Two approaches: centralized and distributed 18 Simulation Results 2012 Thrust Area Technical Review • 2 quarter circle rows of cameras every 5 degrees • Middle camera used as fixed camera • Random set of points around origin Estimated location from Distributed Method Real World Results 2012 Thrust Area Technical Review • Smartphone: Data collected using HTC Evo smartphone • Accurate Sensors: Data collected using CyberQuad UAV sensors and camcorder • Error Distance compared against GPS/Sensor from data Dataset Error Distance Batch Distributed Smartphone Location 2.83 m 4.21 m Orientation 0.2399 0.0017 Location 1.80 m 2.001 m Orientation 5.1490 0.0002 Accurate Sensors Smartphone Results 2012 Thrust Area Technical Review Smartphone Estimate Estimation from distributed method 17 34 1 2 16 22 3 35 4 24 20 5 25 19 6 17 27 Estimation from centralized method H e ig h t( m e te r s ) 15 14 16 9 10 1 .5 0 − . 5 − 1 − 1 . 5 − 2 − 5 31 14 11 12 28 34 35 15 30 31 13 5 1 0 32 1 5 1 0 Y [N o r th − S o u th ] ( m e te r s ) 5 29 33 0 0 26 18 7 8 23 21 X [E a s t− W e s t] ( m e te r s ) 1 5 23 Accurate Sensor Results 2012 Thrust Area Technical Review UAV Estimate Estimation from distributed method Estimation from centralized method 6 5 8 9 4 6 5 7 1 0 H e ig h t ( m e te r s ) 8 .7 1 9 4 1 2 9 3 8 0 − .7 1 0 7 6 1 2 5 3 4 4 5 3 6 7 X [E a s t− W e s t] ( m e te r s ) 2 8 9 1 1 0 0 Y [N o r th − S o u th ] ( m e te r s ) Geo-calibration: MLE 2012 Thrust Area Technical Review • New formulation as a maximum likelihood • Maximize the orientation and position probability estimates of the fixed camera given: – Smartphone orientation position and location – Pairwise relationship based on the image measurements 25 Consensus Framework 2012 Thrust Area Technical Review • This estimate can be solved using a consensus algorithm • In a consensus algorithm, each smartphone has a state of the fixed camera’s orientation and location. The states are iteratively updated to a consensus state. 29 Object Tracking and Search 2012 Thrust Area Technical Review In a smart camera network setup, • Distributed multi-camera object tracking – Given a target, how do the network of cameras with overlapping views robustly track the target with collaboration despite limited network bandwidth? • Efficient object searching in a large network – How to search for objects of interest efficiently without significant visual processing at search time? 33 Browsing and Searching 2012 Thrust Area Technical Review Problem: • A camera network deployed over a large area • No live streaming of any video • Local camera nodes have storage to archive video and limited processing power for simple video analysis ? • How a human image analyst at a distance central node interacts with the remote cameras? 34 Browsing and Searching 2012 Thrust Area Technical Review • Envision the following application scenarios: – A user instantiates the interaction with the network by specifying regions on the image plane (cameras, time intervals) of interest. E.g., “FIND object instances related to region A FROM camera 1 OR region B FROM camera 4 between time 9:30am and 9:35am” – With the results from the previous scenario, the user could then identify one specific object of interest to initiate further searching for the same or related objects. E.g., “FIND all objects related to the object instance at region C FROM camera 1 at time 9:32:41.3am”. 35 Conventional Approaches 2012 Thrust Area Technical Review • Try to provide human users with an high-level interface, such as dynamic global scene visualization and aim to detect/track all observed objects across the entire camera network • To deal with appearance variations across views, much prior work focused on finding the best matching criterion (Javed ICCV’03,Javed CVPR’05, Farenzena CVPR’10, Zhen CVPR’11, Rios-Cabrera CVPR’11.) • What if we cannot have reliable object detection and tracking? Any possible strategy without pair-wise matching or global trajectory finding? 36 Proposed System 2012 Thrust Area Technical Review • Instead of trying to find global trajectories for every object visible in the network, model camera observation directly with a graph model • Act as an intermediate agent between distributed cameras and human image analysts and provide recommendations to the user with a concise and representative set of video snapshots captured by the camera network • Help the image analysts to browse, search and identify objects of interest by giving a canonical overview of the entire set of visual observations in the network 37 Proposed System 2012 Thrust Area Technical Review 38 Real-time detection & tracking 2012 Thrust Area Technical Review • Real time object detection with background modeling and tracking with mean-shift algorithm • An observation record is generated for each detected and tracked object by the camera and sent to the central node over the network 39 Modeling camera observations 2012 Thrust Area Technical Review • Given a user query, we need to find video frames with following properties: – Centrality, representative ones which are closely related to the query and many other observations and hence considered important. – Diversity, covering as many distinct groups as possible • A graph G(V, W) to model relationship among camera observations and perform unified graph ranking for different queries – Individual camera observations (i.e., frames with detected objects) form the vertices set V – Weight matrix W defines the strength of connectivity between camera observations – G(V, W) built at the central server incrementally as the new records are received in real time from the cameras. 40 Spatial Temporal Topology 2012 Thrust Area Technical Review • Divide image plane into 8x6 blocks • Model the time delay for an object to travel between any two blocks across cameras with a Gaussian model with known mean and variance 41 Demonstration – Test bed 2012 Thrust Area Technical Review • • • 11 camera nodes (Cisco wireless-G WVC2300) observing bike path Each camera streams video to a dedicated computer to simulate a smart camera node Approx. 600 meters in width/length 42 Browsing Demonstration 2012 Thrust Area Technical Review • • Top 10 ranked frames (Decreasing order: left to right, top to down) A total of 10 distinct objects satisfying the criterion. All of them have been identified (labeled in yellow). The 8th ranked frame is a “false positive” (it has not passed the queried regions within the specified time interval). 43 Mapping trajectories 2012 Thrust Area Technical Review 9:44:07 - 9:44:08 9:43:13 - 9:43:16 Missed 9:44:10 - 9:44:13 9:43:07 - 9:43:08 Searching Demonstration 2012 Thrust Area Technical Review • Results when searching for object P6 (2nd -11th ranked frame) Forecasted Key Events 2012 Thrust Area Technical Review UAV: Provides intelligence from overhead Events in FY 10 Setup of Outdoor Camera Network Quick Chessboard Calibration Preliminary network topology discovery Events in FY11 Network topology discovery Automatic calibration without patterns Beyond 2D tracking Continued multi-camera tracking with overlapping and non-overlapping views. Preliminary distributed processing with existing work Rapidly Deployed Smart Camera: Provides groundbased views and distributed image analysis Soldier image courtesy of US Marines. Map image courtesy of Google Earth. Events in FY12 • Geo-calibration with mobile sensors • Multi-camera multi-object tracking in a wide-area network • Browsing and search in a distributed network 46 Project Technical Risk Assessment 2012 Thrust Area Technical Review No. Risk Area and Description Risk Rating Risk Reduction Actions Due Date (Date Completed) 1) Improve antenna and camera placements 2) Robust algorithms 1) 12/2010 (L,M,H) General 1 Challenging datasets due to wireless communication constraints L 2) 09/2012 Calibration and Topology 2 Lack of distinctive environmental features for autocalibration M 1) Weak scene calibration 2) Explore mobile cameras 1) 05/2011 2) 05/2011 3 Current topology approach may not scale to large separations M 1) Explore tracking-based approach 2) Explore mobile sensors 1) 11/2010 2) 6/2012 M Explicit occlusion handling 09/2011 Object Detection/Tracking 4 Algorithm fails in severe occlusion Event/Behavior Analysis 5 Complexity of the algorithms M Distribute the algorithm to multiple computing nodes 03/2012 6 UAV data exploration: FAA and university regulations H Collect data in remote areas (nonurban) 6/2012 47 Summary 2012 Thrust Area Technical Review • On target with proposed schedule • • • • • Setup of campus-wide camera network Quick chessboard calibration Camera network topology discovery Wide-area human mobility patterns Single camera activity discovery • Ongoing • Geo-Calibration: automatic camera calibration and camera placement • Online learning and multi-camera fusion for distributed tracking • Multi-camera activity discovery 48 Publications/Reports 2012 Thrust Area Technical Review – – – – – – – – – – T. Kuo, Z. Ni, S. Sunderrajan, and B.S. Manjunath. Map Me: Camera Geo-calibration using Mobile Devices, 2012 (under preparation) T. Kuo, S. Sunderrajan, B.S. Manjunath. “Trajectory Matching for Camera Calibration and Synchronization ”, 2012 (under preparation) Jiejun Xu, Vishwakarma Singh, Ziyu Guan, B.S. Manjunath, “Unified Hypergraph for Image Ranking in a Multimodal Context”, Intl. Conf. on Acoustics, Speech and Signal Processing(ICASSP), Kyoto, Japan, March 2012 S. Sunderrajan, Z. Ni, B. S. Manjunath, “Robust Object Tracking in a Distributed Camera Network,” Technical Report, Nov 2011 (to be submitted). Z. Ni, J.-J Xu, B. S. Manjunath, “Object Browsing and Searching in A Camera Network using Graph Models”, Technical Report, Nov 2011. C. De Leo, B.S. Manjunath. Multicamera Video Summarization from Optimal Reconstruction, Tenth International Workshop on Visual Surveillance, Nov. 2010. J. Xu, Z. Ni, C. De Leo, T. Kuo, B.S. Manjunath, “Spatial-temporal understanding of Urban Scenes through Large Camera Network”, ACM Multimedia Workshop on Multimodal Pervasive Video Analysis, Oct. 2010. Z. Ni, S. Sunderrajan, A. Rahimi, B. S. Manjunath. "Distributed Particle Filter Tracking With Online Multiple Instance Learning In a Camera Sensor Network,” IEEE Int. Conf. on Image Processing, Sep. 2010. Z. Ni, S. Sunderrajan, A. Rahimi, B.S. Manjunath. "Particle Filter Tracking With Online Multiple Instance Learning", Int. Conf. Pattern Recognition, Aug. 2010. T. Kuo, Z. Ni, C. De Leo, B.S. Manjunath. "Design and Implementation of a Wide Area, Large-Scale Camera Network", The First IEEE Workshop on Camera Networks, Jun. 2010. 49 Students/Acknowledgement 2012 Thrust Area Technical Review • Zefeng Ni (graduated Fall 2011) – Distributed Tracking, activity analysis • Thomas Kuo (expected to graduate in 2012) – Calibration/Geo-calibration • Carter De Leo – Video summarization in a camera network, Topology discovery • JieJun Xu (expected to graduate in 2012) – Searching and browsing in large image/video datasets • Thanks to ONR/Martin Kruger for the support and encouragement! 50