a real-time big data architecture for glasses detection using computer vision techniques

Post on 13-Apr-2017

491 Views

Category:

Career

1 Downloads

Preview:

Click to see full reader

TRANSCRIPT

A Real-Time Big Data Architecture For Glasses Detection Using Computer Vision Techniques

Alberto Fernández, Rubén Casado, Rubén Usamentiaga

1

OUTLINE

● Introduction● Algorithm for glasses detection● Big Data architecture for glasses detection● Conclusions● Future work

2Alberto Fernández - alberto.fernandez@fundacionctic.org

INTRODUCTION● The rise of Internet, IoT and Cloud Computing has

led to an impressive growth of data● Increasing information gathered by low-cost

information sensing devices● Domain-specific information collected by

organizations should be analyzed automatically

3

LOGS SENSORS CAMERAS MICROPHONESMOBILE

DEVICES

Alberto Fernández - alberto.fernandez@fundacionctic.org

INTRODUCTION● SQL-based DB perfect for storing and processing

structured data but not prepared for Big Data● Big Data is characterized by the 3Vs

○ Volume: the size of the data to be processed○ Velocity: frequency of the data generation, dynamic

aspects of the data and generating the results in RT.○ Variety: multimodal nature of data:

■ different data schemas of data source■ structured data (ontologies)■ unstructured data (sensors signals)

4Alberto Fernández - alberto.fernandez@fundacionctic.org

INTRODUCTION● Big Data architectures can be classified into:

○ Batch processing (not real time)■ Efficient way of processing high volumes of data

collected during a period of time. ■ Information collected into “batches” as a unit

○ Stream processing■ Continuous input > processing > output data ■ low-response time achieved at the expense of less

rigorous analysis of data○ Hybrid processing

■ Batch and stream processing results are required■ Results are merged, synchronized and composed

5Alberto Fernández - alberto.fernandez@fundacionctic.org

INTRODUCTION● Video/image generated by sensors and devices has

become the largest source○ Processing surveillance videos for information extraction

requires real-time stream processing○ The video data requires to get processed on time to extract

the full benefit of surveillance:■ warning in case of emergency■ something wrong is detected

6Alberto Fernández - alberto.fernandez@fundacionctic.org

INTRODUCTION● Big Data architecture for streaming processing of

large amounts of images is proposed:○ A real-time scalable system for automatic glasses

detection using video images.● Contributions

○ A scalable low-latency architecture for image analysis using Big Data technologies

○ Parallelization of a glasses detection strategy○ Parametrized to detect other face attributes

7Alberto Fernández - alberto.fernandez@fundacionctic.org

PROPOSED SYSTEM AND ARCHITECTURE

● Glasses detection on face images● Big data architecture for glasses detection on face

images

8Alberto Fernández - alberto.fernandez@fundacionctic.org

GLASSES DETECTION ON FACE IMAGES

● Image acquisition● Face detection● Preprocessing of detected face ● Get the feature sets● Classify features

9

Imageacquisition

Face detection

Pre- processing

Build features

Classifi- cation

Alberto Fernández - alberto.fernandez@fundacionctic.org

GLASSES DETECTION ON FACE IMAGES

● Image acquisition○ Read frame from input video○ Convert it to grayscale

10Alberto Fernández - alberto.fernandez@fundacionctic.org

GLASSES DETECTION ON FACE IMAGES● Face detection

○ Viola & Jones algorithm is used:■ robust (high detection rate:high TP and very low FP) ■ executed in real time

11Alberto Fernández - alberto.fernandez@fundacionctic.org

GLASSES DETECTION ON FACE IMAGES● Preprocessing of detected face in order to deal with:

○ pose, rotation, scale and inaccuracies of located face○ noisepiece is placed at the same level as the eyes both in

height and width

12Alberto Fernández - alberto.fernandez@fundacionctic.org

GLASSES DETECTION ON FACE IMAGES● Get the feature sets: Local Binary Pattern (LBP)

○ LBP is a well-known technique in computer vision■ LBP is a type of feature used for classification

○ LBP histogram (LBPH) is usually built for texture classification○ LBPH into mxn regions is calculated to get spatial information

13

LBP

LBP

Alberto Fernández - alberto.fernandez@fundacionctic.org

GLASSES DETECTION ON FACE IMAGES● Classify features

○ Support Vector Machine (SVM) is applied to classify the feature sets

○ SVMs are a useful technique for data classification■ have been proven useful in many pattern recognition

tasks i.e. face recognition

14

GLASSESNO GLASSES

Alberto Fernández - alberto.fernandez@fundacionctic.org

LBP histogram

BIG DATA ARCHITECTURE● Big Data architecture is proposed● Parallelize the different steps of the glasses

detection workflows.○ Topology implemented with a streaming technology

Apache Storm○ Storm is a distributed real-time computation system

released as open source by Twitter● Parametrized to detect other face

attributes

15Alberto Fernández - alberto.fernandez@fundacionctic.org

BIG DATA ARCHITECTUREArchitecture in Storm: two elements● Spouts read information from the source and emit

the data as K-V tuples○ Reads from a message broker (RabbitMQ, Kafka) or

streaming API● Bolts process information coming from the spouts

or other bolts.● Storm defines topologies connecting bolts and

spouts to process in streamingSPOUT represented as

BOLTrepresented as

16Alberto Fernández - alberto.fernandez@fundacionctic.org

BIG DATA ARCHITECTURE

17Alberto Fernández - alberto.fernandez@fundacionctic.org

BIG DATA ARCHITECTURE

VideoSpout:● Split the video streaming into a sequence of images

(frames).○ This Spout uses a shuffle grouping

■ Frames are randomly distributed across the next bolts■ Each bolt is guaranteed to get an equal number of

frames

18Alberto Fernández - alberto.fernandez@fundacionctic.org

BIG DATA ARCHITECTURE

V&Bolt:● Viola & Jones algorithm is applied for each frame● The output of this algorithm is estimated positions

of detected faces as rectangles

19Alberto Fernández - alberto.fernandez@fundacionctic.org

BIG DATA ARCHITECTURE

NormalizationBolt:● From each rectangle, it calculates the region

around the eyes● Returns this region to next step

20Alberto Fernández - alberto.fernandez@fundacionctic.org

BIG DATA ARCHITECTURE

LBPHBolt:● LBP operator is applied to the normalized region● A spatially enhanced histogram is constructed● These features are used in the next step

21Alberto Fernández - alberto.fernandez@fundacionctic.org

BIG DATA ARCHITECTURE

SVMBolt:● Support Vector Machine (SVM) is applied on the

classification step.● Glasses/no glasses classification will be finally

obtained

22Alberto Fernández - alberto.fernandez@fundacionctic.org

BIG DATA ARCHITECTURE● Closed/open eyes classifier using the same architecture

23Alberto Fernández - alberto.fernandez@fundacionctic.org

BIG DATA ARCHITECTURE● Smile classifier using another normalization bolt

24Alberto Fernández - alberto.fernandez@fundacionctic.org

BIG DATA ARCHITECTURE● Gender classifier using another normalization bolt

25Alberto Fernández - alberto.fernandez@fundacionctic.org

BIG DATA ARCHITECTURE

● Type of glasses using another classification bolt

26

SPORT GLASSES

SAFETY GLASSES

SUNGLASSES

GOOGLE GLASSES

Alberto Fernández - alberto.fernandez@fundacionctic.org

CONCLUSIONS

● Real-time Big Data architecture○ Collect, maintain and analyze massive volumes of

images○ It can be modified in order to detect other attributes:

■ smile, gender, age or face recognition classifiers

27Alberto Fernández - alberto.fernandez@fundacionctic.org

FUTURE WORK

● Deep Learning○ Deep Learning algorithms in our pipeline detection

architecture will be proposed

28Alberto Fernández - alberto.fernandez@fundacionctic.org

29Alberto Fernández - alberto.fernandez@fundacionctic.org

Thanks for your attention

Alberto Fernándezalberto.fernandez@fundacionctic.org

A Real-Time Big Data Architecture For Glasses Detection Using Computer Vision Techniques

Alberto Fernández, Rubén Casado, Rubén Usamentiaga

top related