Download Event detection - ifp.illinois.edu

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts
no text concepts found
Transcript
Unusual Event Detection
via Multi-camera Video Mining
Hanning Zhou ([email protected]) and Don Kimber ([email protected] )
Step 1: Temporal segmentation
Introduction
 Goal: detecting unusual events from a large
amount of multiple stream video.
 Challenges: multi-stream video
lack of labeled data
 Approach: semi-supervised learning
Applications
static scene
……………………
……
…
…
…………………
Segment k
 Online
 Detect unusual events alarm in surveillance system
 Offline
 Detect highlights from sport videos
 Analyze business process
 Other data streams
 audio, text and multimodal data streams
…
…
Segment k + 1
Step 2: Feature extraction
 Feature:
 size and location of the motion blobs
 clustered into GMMs
 Advantage:
 higher spatial resolution than motion histogram
[Zhang ‘05] [Zhong ‘04]
Key Idea
 New problem: spatial alignment
 Collaborative mining of multiple streams
 Solved with approximate KL-divergence
[Goldberger ‘03]
 Sensor network is prevailing
 Events from different sensors are related
 Two-Stage Training
………
Step 3: Training a Model for usual event
 Unusual events are rare
 Manual labeling is intractable
 1st Stage: Bootstrap


Previous Work
 General Event Detection
 specific event with well-defined model
 supervised statistical learning (DTW, HMM, factor graph)
 Unusual Event Detection
 unsupervised [Zhong & Shi ‘04]
 semi-supervised [Zhang et al. ‘05]
 All above are from single stream
Experimental Results
 Trained on one week’s video recorded in FXPAL
 Tested on another week’s video
 usual events: pass by, pick up print outs
 unusual events: distribute mails, open multiple drawers
to look for stationery, open the cabinet, multiple people
Clustering: keep the large clusters
User feedback: exam the small clusters
 2nd Stage: Train CHMM for usual events
 Inference in CHMM is efficient
O(T(CN)^2) vs O(TN^(2C))
 statistic model to handle variant durations, noisy
observation and asynchrony
 The hidden state depends on 3D location of the objects
 The observations are 2D projection of the objects
 CHMM as a loose stereo
 Dependent chains
The 3D location inferred from
different cameras are RELATED
Exact stereo does not work well, because:
gaps between the views
wide baseline, few correspondences
Examples of usual events
Step 4: Detecting usual event
Examples of unusual events
 Evaluating the likelihood with forward-backward
algorithm [Brand ‘97]
ROC curve of HMM vs CHMM on Terrascope data
 Quantitative experiments on Terrascope data
 48 segments in 4 scenarios
 usual events:
group meeting, natural video sequence
 unusal events:
group exit, intruder, theft, suspicious behavior
Mailroom camera setup
Terrascope dataset camera setup
Courtesy of Christopher Jaynes
Related documents