continuous automatic identification and classification of ... · case 2: video classification of...

10
8/22/2019 1 1 Continuous Automatic Identification and Classification of Trash on Street using Smartphone I3 Conference on Aug. 19, 2019 Seon Ho Kim, Ph.D. Integrated Media Systems Center Viterbi School of Engineering University of Southern California [email protected] 2

Upload: others

Post on 03-Jun-2020

4 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Continuous Automatic Identification and Classification of ... · Case 2: Video Classification of Street Cleanliness 8/22/2019 10 10 ... Case 4: Road Damage Recognition and Classification

8/22/2019

1

8/22/2019 11

Continuous Automatic Identification and Classification of Trash on Street using Smartphone

I3 Conference on Aug. 19, 2019

Seon Ho Kim, Ph.D.

Integrated Media Systems Center

Viterbi School of Engineering

University of Southern California

[email protected]

8/22/2019 22

Page 2: Continuous Automatic Identification and Classification of ... · Case 2: Video Classification of Street Cleanliness 8/22/2019 10 10 ... Case 4: Road Damage Recognition and Classification

8/22/2019

2

8/22/2019 33

So many photos of Los Angeles downtown.How can we efficiently manage and learn from them?

8/22/2019 44

Motivation

• More and more visual data are generated

– More amount than human can physically watch machine watches

• Still, visual data collection is expensive and in adhoc manner

– Size of data, orchestration of collection

• Utilization of visual data for city, especially smart city

– Automation, efficiency, cost effectiveness

• Preparation for AI and machine learning is needed

Page 3: Continuous Automatic Identification and Classification of ... · Case 2: Video Classification of Street Cleanliness 8/22/2019 10 10 ... Case 4: Road Damage Recognition and Classification

8/22/2019

3

8/22/2019 55

Working with City of Los Angeles

• Beginning…– LA Mayor Eric Garcetti's Clean Streets initiative: monitor street

cleanliness

• Currently, data are manually collected and evaluated: inefficient, costly automate!

• In collaboration with the Sanitation Department of LA, automatically detect the cleanliness of streets as well as any special objects in need of removal.

8/22/2019 66

Case 1: Image Classification of Street Cleanliness

Images collected by the Sanitation Department, City of Los Angeles

Page 4: Continuous Automatic Identification and Classification of ... · Case 2: Video Classification of Street Cleanliness 8/22/2019 10 10 ... Case 4: Road Damage Recognition and Classification

8/22/2019

4

8/22/2019 77

Image Based Classifier

Published in IEEE Big Multimedia, 2018

• Achieved 80 – 90% accuracy (depending on class): stable and practical

• The more images in an area, the higher the accuracy becomes: promising

8/22/2019 88

Integration with I3(Intelligent IoT Integrator)

Seller: MyLA311 Mobile App (images) or Sanitation Dept. Truck (videos)

Broker: Classifier provided by IMSC

Buyer: City of Los Angeles or any Web Applications

Images need to be manually collected any way of automatic collection?

Page 5: Continuous Automatic Identification and Classification of ... · Case 2: Video Classification of Street Cleanliness 8/22/2019 10 10 ... Case 4: Road Damage Recognition and Classification

8/22/2019

5

8/22/2019 99

• 700+ garbage collection trucks are being operated

• Through four cameras (three attached on the sides on the truck and one on the front, each with a resolution of 704*480 and a frame rate of 29 fps), a truck operating for five hours (7 AM – 12 PM) generates videos whose size is around 19 GB.

• We extract one keyframe (i.e., an image) per second from videos.

Case 2: Video Classification of Street Cleanliness

8/22/2019 1010

Video Frame Examples and classification

Bulky Item Illegal Dumping

Encampment Clean

0

10

20

30

40

50

60

70

80

90

100

Bulky Item Clean Encampment Extra Vegetation Illegal Dumping

Mea

sure

men

t Sc

ore

Precision

Recall

F1 score

Video data transfer is too expensive any way to be affordable?

• Achieved comparable accuracy using the model learned from images

• For videos, higher performance is needed to be practical

Page 6: Continuous Automatic Identification and Classification of ... · Case 2: Video Classification of Street Cleanliness 8/22/2019 10 10 ... Case 4: Road Damage Recognition and Classification

8/22/2019

6

8/22/2019 1111

Case 3: Mobile Realtime Classification

Implemented and tested to simulate real time trash detection and classification with video camera on garbage collection truck: feasible!

Continuous Classification

Using Smartphone

Report result with high confidence only

Report result with high confidence only

8/22/2019 1212

Edge Computing

• Train machine learning models on the server (initial model)

• Distribute the models to edge devices (e.g., smartphones, smart cameras)

• Inference happens on edge devices using CPU on edge

• Report selected results to involved agencies

• Improve models iteratively

Framework to train, distribute and adapt models

To appear in IEEE BigMM, Sep. 2019

Page 7: Continuous Automatic Identification and Classification of ... · Case 2: Video Classification of Street Cleanliness 8/22/2019 10 10 ... Case 4: Road Damage Recognition and Classification

8/22/2019

7

8/22/2019 1313

Challenges

• Edge devices exhibit different capabilities (Hardware)

– CPU/GPU, Memory, Bandwidth

• May have power limitations (Mobile)

• May have limited availability/connectivity (Bandwidth)

• Need to consider: the model and binary size, application speed and model loading speed, performance

• Most of the existing frameworks train a single model on the server side and distribute it to edge devices

8/22/2019 1414

• Any other useful application of collected data?

• The back-camera attached on truck cab be a visual source for road damage monitoring.

– All LA streets can be monitored once a week!

Case 4: Road Damage Recognition and Classification

Real Images from LASAN

Page 8: Continuous Automatic Identification and Classification of ... · Case 2: Video Classification of Street Cleanliness 8/22/2019 10 10 ... Case 4: Road Damage Recognition and Classification

8/22/2019

8

8/22/2019 1515

IEEE Big Data Challenge Cup Competition (Dec. 2018)Classification of Road Damages

Defined by Japan Road Association [*]

[*] Maintenance and Repair Guide Book of the Pavement 2013, 1st ed. Tokyo, Japan: Japan Road Association, 04 2017.

Road damage detection from LA garbage trucks can be feasible and practical too.We won Bronze Prize (3rd place) out of 59 teams from 15 countries

Using images recorded in Japan using smartphones

8/22/2019 1616

Case 5: Identify Surface Material Beneath Graffiti

• Cleaning crew need to know what material beneath graffiti is

Metal Telephone Pole Wood Telephone Pole

Class

Wall

Brick Wall

Concrete Wall

Wood Wall

Stucco Wall

Sound Wall

Door

Metal Door

Wooden Door

Barricade Fence

Metal Fence

Wooden Fence

Fire HydrantPedestrian Furniture

Pole

Metal Pole

Wooden Pole

Granite Pole

Mailbox Sidewalk

walkway

Curb

StairsTraffic

Control BoxSign

Street Signs

Traffic Signs

Tree Windows

Page 9: Continuous Automatic Identification and Classification of ... · Case 2: Video Classification of Street Cleanliness 8/22/2019 10 10 ... Case 4: Road Damage Recognition and Classification

8/22/2019

9

8/22/2019 1717

Classes and Classification Result

To appear in IEEE Image Processing, Sep. 2019

• Fundamentally challenging problem due to multistep learning

• As a research result, achieved an excellent result (50-60% accuracy)

• But, maybe not enough to be practical yet: need to enhance accuracy further

8/22/2019 1818

Visual Analysis for Smart City

Trash

Graffiti

Parking

Road Damage

Thermal Images

(Fire)

Edge Computing

I3 as IoT Platform

Visual Data Analysis

Action

Camera as sensor

Page 10: Continuous Automatic Identification and Classification of ... · Case 2: Video Classification of Street Cleanliness 8/22/2019 10 10 ... Case 4: Road Damage Recognition and Classification

8/22/2019

10

8/22/2019 1919

Discussion• Next steps? How to use the techniques for LA?• Continue our research

– Image analysis needs more data: trash, graffiti, road damage– Sharing through I3

• Open datasets– Make available to the public: trash and graffiti

• Real world test– Mobile app (MyLA311), Smart camera on garbage collection truck

• Funding, Industry partners– Thank Oracle for providing generous cloud resources

8/22/2019 2020

Q & A

Seon Ho [email protected]