This document describes how to use MapReduce to process video from CCTV cameras to identify faces and their locations. The process involves: 1. The mappers take chunks of video as input, detect faces, extract features, and output face data with identifiers. 2. A shuffler redistributes the face data to reducers based on the identifiers. 3. The reducers calculate distances between the extracted features and a database of known faces. Faces that match are output along with their locations from the video frames. This allows identifying 20 people across 470,000 cameras in Beijing.

1. MAP REDUCE
Syed Zaid Irshad
15007114015
MS-SE

2. Scenario
■ City: Beijing, China
■ Population: 11.51 million
■ CCTV Cameras: 470,000 (24.5 person/camera)
■ EstimatedVideo generation:
– Frames per second: 30
– Video Length: 1 hour
– Storage: 6091.2TB.
■ Task: Find 20 peoples and their location

3. How to do this with single Camera

4. Face Detection
■ Training
– Integral Image
– AdaBoost
■ Real time Detection
– Cascade classifier

5. Face Detection
■ Find features
– Haar feature
■ Similar Properties
– The eyes region is darker than the upper-cheeks.
– The nose bridge region is brighter than the eyes.
■ Values
– FeatureValue = ∑ (pixels in black area) - ∑ (pixels in white area)
Integral Image

6. Face Detection
■ Problem
– Too many features (24x24 image contain 160000 features)
■ We have to select a subset of relevant features – which are informative - to model a
face
■ Constructs a “strong” classifier as a linear combination of weighted simple “weak”
classifiers
■ Experiments showed that a 200-feature classifier achieves: 95% detection rate
AdaBoost

7. Face Detection
■ Properties of Strong Classifier
– Number of layers in cascade (strong classifiers)
– Number of features of each strong classifier
– Threshold of each strong classifier
Cascade Classifier

9. Face Recognition
Geometric
■ Now the detected face image is preprocessed (resize)
■ Calculate distance between all points
– Euclidian distance

10. Face Recognition
Database Search
Found
Not Found
Face Feature
Assign an ID to results.

12. Scenario
■ City: Beijing, China
■ Population: 11.51 million
■ CCTV Cameras: 470,000
■ Estimated Video generation:
– Frames per second: 30
– Video Length: 1 hour
– Storage: 6091.2TB
■ Task: Find 20 peoples and their location
– 3.84MB/image

13. MapReduce
■ Scalability
■ Cost-effective solution
■ Flexibility
■ Fast
■ Parallel processing
■ Availability

14. MapReduce
■ Tasks to perform
– Face detect
– Face recognition
■ Feature extraction
■ Comparison

15. ■ Input stream format: MJPEG-2
■ Assign each section of input stream (64MB) to a device and do operation

16. Mapper
■ It will receive the chunk of video.
■ Detect faces
■ Identify face-points

17. Class : Mapper
Function : Map
Map(TEXT key ( filename ) , BytesWritable value ( imgdata ) , FaceInfoString, ValueBuffer)
getImgBinaryData_From_Value;
convertBinaryData_To_JavaImage;
InitializeOpenCV_Via_JNI Interface;
runOpenCV_HaarLikeFaceDetector;
foreach ( DetectedFace )
createFaceBuffer_FaceSize;
createValueBuffer_Size30;
copyFacePixels_To_FaceBuffer ;
runGeometricBasedValues_To_ValueBuffer;
copyFaceValues_To_ValueBuffer;
create_FaceInfoString ;
//Output
set_key_FaceInfoString ;
set_value_ValueBuffer ;
end_foreach

18. Shuffler
■ On the bases of face-points it will send them to reducer

19. Reducer
■ Calculate distance from face-points.
■ Compare it with input face-points distance

20. Class : Reducer
Function : reduce
Reduce(TEXT key ( FaceInfoString ) , TEXT value ( ValueBuffer ) , FaceInfoString, DetectedFace)
for ( ValueBuffer )
createDifferenceBuffer_Size30;
runEculidean_To_Valuebuffer;
copyResults_To_DifferenceBuffer;
compareDifferenceBuffer_With_Database;
if (compareResultIsTrue)
copyImageFromDataBase_To_DetectedFace;
outputInformer :
set_key_FaceInfoString;
set_value_DetectedFace;
end_if
end_for