design implementation of a wireless video surveillance system

Design & Implementation of a Wireless Design & Implementation of a Wireless Video Surveillance Systemy

Aakanksha Chowdhery*, Tan Zhang†, Victor Bahl*, Kyle Jamieson‡, Suman Banerjee†

*Microsoft Research ‡UCL/Princeton †UWisconsin Madison

Video Surveillance: Pervasive and UsefulVideo Surveillance: Pervasive and Useful

London & Beijing: 1 million cameras deployed

Intrusion detection – campus, airport, train stationCustomer analytics store toll booth parking garageCustomer analytics – store, toll booth, parking garageTraffic monitoring – cities, freeways

2

Wireless surveillance camerasWireless surveillance cameras

Easy to install

No need for wired backhaul to camera

Comprehensive coverage

How do we build a large‐scale wireless video

3

How do we build a large scale wireless video surveillance network?

Video Cameras overwhelm wireless capacity quicklydeo Ca e as o e e e ess capac ty qu c y

Wireless is a shared medium in scarce spectrum

Today’s wirelessSupports 20 cameras@1 Mbps streams

Wireless Coverage Area: 10,000 sq ft, q

Complete Surveillance coverage p gRequires 400 cameras@10 ft camera range 100 sq ft

Vigil: Wireless Video Surveillance SystemVigil: Wireless Video Surveillance System

Goals:Goals:

1. Maximize surveillance application accuracy1. Maximize surveillance application accuracy

2. Minimize wireless capacity usage

Techniques:Techniques:

Edge Redundancy Content-awareEdge Computing

Redundancy Suppression

Content-aware Traffic Scheduling

5

Vigil: Wireless Video Surveillance SystemVigil: Wireless Video Surveillance System

Goals:Goals:

1. Maximize surveillance application accuracy1. Maximize surveillance application accuracy

2. Minimize wireless capacity usage

Techniques:Techniques:

Edge Redundancy Content-awareEdge Computing

Redundancy Suppression

Content-aware Traffic Scheduling

6

Useful video content is sparse

Example: find a person of interest by face

Useful video content is sparse

Example: find a person of interest by face250 hours of video feed in busy office halls

1 f

2 faces, 3% >=3 faces, 1%

1 face, 16%

No faces, 80%

7

Less than 20% feed has people

Edge Computing NodeEdge Computing Node

Edge Computing Node detects queried objectsg p g q j

ControllerController

Edge ComputingNode

8

Vigil Architecture with Edge ComputingVigil Architecture with Edge Computing

F IDU l d!

ControllerQuery

Frame ID, Person Count

Upload!

ControllerQ y

9Frame ID, 

Person CountFrame ID, 

Person Count

Vigil only uploads relevant frames!

Vigil: Wireless Video Surveillance SystemVigil: Wireless Video Surveillance System

Goals:Goals:

1. Maximize surveillance application accuracy1. Maximize surveillance application accuracy

2. Minimize wireless capacity usage

Techniques:Techniques:

Edge Redundancy Content-awareEdge Computing

Redundancy Suppression

Content-aware Traffic Scheduling

10

Video Codecs compress frames temporallyVideo Codecs compress frames temporally

Motion suppressed by difference coding (H.264)

How can we compress frames from multiple cameras?

11

Compressing frames across camerasCompressing frames across cameras

Camera 1 Camera 2

Camera Cluster

How do we suppress redundant images between cameras in a cluster?

12

between cameras in a cluster?

Vigil’s Redundancy Suppression AlgorithmVigil s Redundancy Suppression Algorithm

Camera 1 Camera 2

Camera 2 hasFrame Index 1 2 3 4 5

Camera 1 0 0 0 0 0

Camera 2 has higher utility

Step 1: Frame Utility = Number of queried objects in frameCamera 2 3 3 3 3 3

p y q jStep 2: Select frames from camera 2

13

Vigil’s Redundancy Suppression AlgorithmVigil s Redundancy Suppression Algorithm

Camera 1 Camera 2

Camera 1 hasFrame Index 1 2 3 4 5

Camera 1 3 3 3 3 3

Camera 1 has higher utility

Step 1: Frame Utility = Number of queried objects in frameCamera 2 0 0 0 0 0

p y q jStep 2: Select frames from camera 1

14

Vigil’s Redundancy Suppression AlgorithmVigil s Redundancy Suppression Algorithm

Camera 1 Camera 2

Camera 1 hasFrame Index 1 2 3 4 5

Camera 1 3 3 3 3 3

Camera 1 has higher utility

Step 1: Frame Utility = Number of queried objects in frame

Camera 2 2 2 2 2 2

Step 1: Frame Utility Number of queried objects in frameStep 2: Re-identify objects from camera 2 in camera 1

15

Vigil’s Redundancy Suppression Algorithm: R id tifi tiRe-identification

Camera 1 Camera 2

Pixels of 2nd face  Pixels of 

Check if pixels

2nd face  

Frame Index 1 2 3 4 5

Camera 1 3 3 3 3 3

Check if pixels map to same 

physical location?

Step 1: Frame Utility = Number of queried objects in frame

Camera 2 2 2 2 2 2

Step 1: Frame Utility Number of queried objects in frameStep 2: Re-identify objects from camera 2 in camera 1

16

Vigil’s Redundancy Suppression AlgorithmVigil s Redundancy Suppression Algorithm

Camera 1 Camera 2

Upload onlyFrame Index 1 2 3 4 5

Camera 1 3 3 3 3 3

Upload only change frames

Step 1: Frame Utility = Number of queried objects in frameCamera 2 2 2 2 2 2

p y q jStep 2: Re-identify objects from camera 2 in camera 1Step 3: Upload frames from camera 1 17

Vigil’s Redundancy Suppression AlgorithmVigil s Redundancy Suppression Algorithm

Step 3: Upload only change frames

Frame Utility= Number of faces detected

1 1 2 2 3

Upload! Upload! Upload!Upload! Upload! Upload!

18

Vigil increases surveillance accuracy h i l it t twhen wireless capacity saturatesSingle cluster, 3 cameras, Medium activity of people

80%

100% Without Vigil Vigil

60%A

40%Accuracy

0%

20%

0%50 kbps 100 kbps 200 kbps 300 kbps

Available Wireless Capacity (kbps)

19100% accuracy = Upload all change framesVigil provides 1.7x accuracy for same wireless capacity!

Vigil increases surveillance accuracy when wireless capacity saturateswhen wireless capacity saturates

100%Single cluster, 3 cameras, High activity of people

80%

100%Without Vigil With Vigil

60%

40%Accuracy

20%

0%50 kbps 100 kbps 200 kbps 300 kbps

Available Wireless Capacity (kbps)

20

Available Wireless Capacity (kbps)

Vigil provides ≈1.7x accuracy for same wireless capacity!

Vigil: Wireless Video Surveillance SystemVigil: Wireless Video Surveillance System

Goals:Goals:

1. Maximize surveillance application accuracy1. Maximize surveillance application accuracy

2. Minimize wireless capacity usage

Techniques:Techniques:

Edge Redundancy Content-awareEdge Computing

Redundancy Suppression

Content-aware Traffic Scheduling

I ill li ti h21

Increases surveillance application accuracy when available wireless capacity is limited in camera clusters

Vigil: Wireless Video Surveillance SystemVigil: Wireless Video Surveillance System

Goals:Goals:

1. Maximize surveillance application accuracy1. Maximize surveillance application accuracy

2. Minimize wireless capacity usage

Techniques:Techniques:

Edge Redundancy Content-awareEdge Computing

Redundancy Suppression

Content-aware Traffic Scheduling

P i iti l t ith bj t22

Prioritizes camera clusters with objects relevant to query

Hybrid surveillance-access networkHybrid surveillance access network

Reuse bandwidth savings to provide Wi-Fi access

F ID

g p& recoup the surveillance network cost

Controller

Frame ID, Person Count

Controller WiFi traffic

WiFi AP

WiFi traffic

WiFi t ffi

23Frame ID, 

Person CountFrame ID, 

Person Count

WiFi traffic

System DeploymentSystem Deployment

Whitespaces in MSR UWisconsin; Wi-Fi in UCLWhitespaces in MSR, UWisconsin; Wi Fi in UCL- Using a 802.11 baseband protocol- Integrated frequency translator to operate at UHF bandg q y p

24

System Deployment at Microsoft ResearchSystem Deployment at Microsoft Research

25

Related workRelated work

CloudletsCloudletsVM-based Cloudlets (Pervasive Computing ’09)Dynamic offloading of mobile apps – Maui (MobiSys’10),Gabriel (MobiSys’14)

Cloud-based video surveillance systemsWired - IBM’s Smart Surveillance System (’05)Wireless Dropcam ships video to cloudWireless - Dropcam ships video to cloud

Video Compression algorithmsMPEG-4 H 264 eliminate redundancy across framesMPEG-4, H.264 eliminate redundancy across framesImage similarity - Re-identification, Perceptual hashing

Vision analytic algorithms on MobileVision analytic algorithms on MobileHarr Cascade based face detection – Glimpse (MobiSys’14)SIFT based object detection- CarSafe (MobiSys’13)

26

ConclusionsConclusions

Vigil’s edge computing provides at least 5x reduction in frames to be uploaded

Vigil’s redundancy suppression provides 1.7xVigil s redundancy suppression provides 1.7x surveillance accuracy when available wireless capacity is limitedcapacity is limited

Vigil’s content a are traffic sched ling ploadsVigil’s content-aware traffic scheduling uploads 25% more objects relevant to the user’s query

27