Traffic Monitoring of Motorcycles during

companyscourgeAI and Robotics

Oct 19, 2013 (3 years and 9 months ago)

54 views

TRB 89
th

Annual Meeting

Traffic Monitoring of Motorcycles during
Special Events Using Video Detection


Dr.
Neeraj

K.
Kanhere

Dr. Stanley T.
Birchfield

Department of Electrical Engineering


Dr. Wayne A. Sarasua, P.E.

Sara
Khoeini

Department of Civil Engineering


College of Engineering and Science

Clemson University

TRB 89
th

Annual Meeting

Introduction


Data from
NHTSA FARS indicates disturbing
trends
in motorcycle safety



In 2006, motorcycle rider fatalities increased for


the ninth consecutive year.

During this period, fatalities more than doubled

Significantly outpaced motorcycle registration



Traffic data collection and motorcycles

In June 15, 2008 FHWA began requiring
mandatory reporting of motorcycle travel as part
of HPMS

Need VMT data as well as crash data to assess
motorcycle safety

In September, 2008, an HPMS report indicated
that the quality of MC data was questionable due
to the inability and inconsistency of current
traffic monitoring equipment.


TRB 89
th

Annual Meeting

TRB 89
th

Annual Meeting

Challenges with motorcycles

Three main reasons why motorcycles are difficult
to count:


light axle weight


low metal mass


narrow footprint


Historically, collection
of motorcycle data has
been a low
priority. Many commercially
available classification systems are generally
unable to accurately capture motorcycle traffic.
Emphasis in the past has been on detection.

Overview of this research

Significant amount of motorcycle traffic

Variety of formations


Chose a motorcycle rally Myrtle Beach, SC


TRB 89
th

Annual Meeting


Evaluate a computer vision based tracking system
that can count and classify motorcycles


TRB 89
th

Annual Meeting

Collecting Vehicle Class Volume Data

Different types of sensors can be used to gather these data:


Axle sensors


Presence sensors


Machine vision sensors



Several manufacturers indicate their devices can
detect/classify motorcycles



motorcycle classification accuracy specifications not
available



we could not identify any validation studies


Motorcycle classification with traditional sensors

Issues with length based classification

Some cars are not much longer than the average motorcycle


European “city cars” are gaining popularity

Average motorcycle size is larger than ever before.


Cruisers have become very popular


Wheel base is within 10” of many subcompacts


Axle counters are especially prone to length base
classification errors


TRB 89
th

Annual Meeting

TRB 89
th

Annual Meeting

Loop detector


Amongst the most reliable traffic


Capable of collecting speed, volume, and classifications


Several configurations depending on application


Length based classification is most common



Adjusting detector
senstivity

may lead


to
crosstalk with trucks in nearby
lanes

Motorcycle detection and classification



Classification possible w/loop arrays



Electromagnetic profiling promising


Motorcycle Travel Symposium

Overhead and side
non
-
intrusive
devices

Active and passive infrared, radar, and acoustic devices


Capable of collecting speed, volume, and classifications


Length based classification is most common

Motorcycle detection and classification


Vehicle profiling is possible (e.g. vehicle contour)


Some specify >99% accuracy (scanning infrared)

TRB 89
th

Annual Meeting

Small footprint sensors


Magnetometers


Capable of collecting speed, volume, and classifications


Length based classification is most common


Motorcycle
detection and classification


is
most promising with an array of



probes
spaced at 3’ to 4’ intervals

Motorcycle detection and classification

TRB 89
th

Annual Meeting

Axle sensors


Most are intrusive (
piezo
). Some temporary (hose)


Capable of collecting speed, volume, and classifications


Several configurations depending on application


Length based and weight base classification possible



Weight base may be most


promising

Motorcycle detection and
classification

TRB 89
th

Annual Meeting

Machine Vision Sensors


Proven technology


Capable of collecting speed, volume, and classifications


Several commercially available systems


Uses virtual detection



Provides rich visual information for


manual inspection



No traffic disruption for installation



and maintenance




Covers wide area with a single camera

Benefits of video detection

Motorcycle Travel Symposium

Traditional Approach to Video Detection

Limitations of localized video detection


Errors caused by occlusions


Spill
-
over errors


Problems with length based classification


Cameras must be placed very high (to > 40’) to minimize error

Current systems use localized virtual detectors which can be prone

to errors when camera placement in not ideal.

Research on motorcycle video detection

Significant recent work on tracking but very little related to
motorcycle detection

Duan

et al. present on
-
road lane change assistant that can
identify motorcycles using AI including Support Vector
Machines


Detection rates over 90%

Chiu et al. uses an occlusion detection and segmentation
method using visual length and width and helmet detection.


95% recognition rate for a field study of 42 motorcycles


TRB 89
th

Annual Meeting

TRB 89
th

Annual Meeting


Clemson’s tracking
a
pproach

Tracking enables prediction of a vehicle’s location in

consecutive frames.








Clemson System demo

TRB 89
th

Annual Meeting

Algorithm Overview

TRB 89
th

Annual Meeting

Simple Calibration

TRB 89
th

Annual Meeting

Classification

TRB 89
th

Annual Meeting

Classified vehicles

TRB 89
th

Annual Meeting

Oops…

TRB 89
th

Annual Meeting

Field evaluation of Clemson system

First attempt at automated motorcycle data
collection at a bike rally

Literature indicated several manual efforts


Jamar

type counters


Post processing video

Sturgis has been used automated counters since
1990 but only to
collect total vehicle volumes


TRB 89
th

Annual Meeting

Camera details

Pan
-
Tilt
-
Zoom

Autofocus with
automatic exposure

640 x 480 resolution

30 frames per second


TRB 89
th

Annual Meeting

Data collected at 2 locations

TRB 89
th

Annual Meeting

Summary of Results

TRB 89
th

Annual Meeting

Approaching

Departing

Total

MC Actual Counts

805

684

1489

MC System Counts

784

714

1498

MC Percent of Difference

-
2.61

4.38

0.6

PC and HV Actual Counts

580

598

1178

PC and HV System Counts

593

582

1175

PC and HV Percent of Difference

2.24

-
2.67

-
0.25

Total Actual Counts

1385

1282

2667

Total System Counts

1377

1296

2673

Total Percent of Difference

-
0.57

1.09

0.22

Actual
Counts

System
Result

Dif
(Percents)

MC

726

681

-
6.19

PC and HV

333

321

-
3.60

Total

1059

1002

-
5.38

Myrtle Beach Site

Garden City Site

Garden city results (both directions)

Garden city results (both directions)

Garden City results
-

regression analysis



PC & HV

MC

All Vehicles

Slope

1.0009

0.9861

0.9925

R
-
Sq

1.0000

0.9998

1.0000

Myrtle Beach results

Myrtle Beach site video

TRB 89
th

Annual Meeting

Garden City site video

TRB 89
th

Annual Meeting

Two directions at once (speed calibrated)

TRB 89
th

Annual Meeting

Verifying speeds

TRB 89
th

Annual Meeting

TRB 89
th

Annual Meeting

Conclusion


Motorcycle classification within 6% of actual even
in extreme conditions:


Algorithm works in real time



Very high volumes of motorcycles


Tight formations (staggered and pairs)



Improve robustness to eliminate systematic errors


Evaluate night time/low light conditions


Augment
algorithim

with pattern
-
based descriptors

Future work

TRB 89
th

Annual Meeting

Thank you !

TRB 89
th

Annual Meeting

For more info please contact:


Dr. Stanley T.
Birchfield

Dr.
Neeraj

K.
Kanhere

Department
of Electrical Engineering

stb@clemson.edu

nkanher@clemson.edu


Dr. Wayne A. Sarasua, P.E.

Department of Civil Engineering


sarasua@clemson.edu