ITS at the Rail Crossing: Needs and Opportunities

1
ITS at the Rail Crossing Needs and Opportunities

• T. E. Cohn and Z. Kim
• Visual Detection Lab
• UC Berkeley
• TO-4138 and RTA 65Y350

2
Outline

• Rail Crossing Collisions
• New Technology
• Embedded Pavement Signals
• LED Warning Lights
• Advanced Signals, Signs
• Data Needs
• Driver Behavior at the crossing
• Driver Behavior from the Train
• The LVDAS
• LVDAS Data Processing (Z. Kim)

3
Project Team

• Scott Johnston
• Kent Christiansen
• Dan Greenhouse
• Tieuvi Nguyen
• Henry Truong
• Zhiyong An
• Zuwhan Kim
• Delphine Cody

• Ted Cohn
• Jim Misener

4
Rail Crossing Collisions

• One type of intersection collision
• Train ? Vehicle
• Destructive consequences
• Loss of life and also property
• Costly Delays rail is increasingly crowded
• Ceiling on rail speed
• Higher speed threatens rail infrastructure,
  equipment
• Current speed ceiling limits rail efficiency

5
Solution genres

• Brute force solutions
• Grade separation (tunnel or bridge)
• Very costly
• Retire crossing can be costly, politically
  difficult
• Partial solutions
• Upgrade signals/gates
• Passive to active signal
• Active signal to multi-signal cantilever
• Signal to gate
• Simple gate to four quadrant gate
• Gate to barrier

ascending cost
6
ITS Solutions

• Train tracking for intelligent signaling
• Active signage Countdown second train coming
• Wireless signals for in-vehicle warning
• in-vehicle signing in school buses
• Infrastructure to vehicle communication
• Modulated signals
• In-pavement warnings
• Visual barrier at crossing
• Improved signs and signals
• Diamond warning
• Moving, not flashing signal

7
In-pavement signals
½ in. exposed above pavement

• Deploy three / lane
• COTS ignite simultaneously
• Our (tentative) improvement
• Inside-Out ignition pattern
• Cut tens of milliseconds off of RT average for a
  bus rear-end collision warning

8
(No Transcript)
9
(No Transcript)
10
(No Transcript)
11
(No Transcript)
12
(No Transcript)
13
(No Transcript)
14
(No Transcript)
15
(No Transcript)
16
(No Transcript)
17
(No Transcript)
18
(No Transcript)
19
(No Transcript)
20
(No Transcript)
21
(No Transcript)
22
In-Pavement SignalsCross-walk application
pair
individual
23
Installation is planned for one of
three crossings in Kern Co. along the
BNSF right-of-way to Bakersfield
24
Project purpose TO-4138

• Study means of improving the visual effect of
  in-pavement signals
• Speed of response
• As proxy for intelligibility of signal
• Test deployment for such improved signals
• Measure driver behavior
• Vehicle monitors necessary
• Radar could work, but is a target in Kern Co.
• Studying magnetometer-based

25
Improved Signs
26
IMPROVED SIGNALS
27
IMPROVED SIGNALS
28
We will study crossing activity from orthogonal
perspectives

• From the ground, at a given crossing
• From the train, as it passes many

29
Collision Factors

• For as long as rail crossing collisions have
  occurred, the question of what causes them has
  been raised.
• Classes of answer
• Equipment (signs, signals, gates)
• Road Parameters (angle with track crown no.
  lanes speed limit)
• Track parameters (number tracks, speed)
• Visual sightlines, weather, day/night
• Victim (age, alcohol, speed, etc)

30
Analyses of collision records

• Multi-factor analyses have been performed
• Some insights thus gained appeal to intuition
• Upgraded signals are better
• Main purpose is to identify crossings at which to
  spend scarce upgrade dollars
• Secondary purpose to identify collision causes
• To inform intervention strategies
• But collisions are rare

31
Database of precursor events

• Since collisions are rare
• We choose to study near-misses, event that might
  have become collisions
• These should be much more common
• A database of near-misses will allow tests of
  intervention strategies such as in pavement
  lights
• To build the database, we need to record events

32
The Locomotive Video Data Acquisition System
(LVDAS)
This consists of a three cameras (one
facing forwards, and to each side. A video
switcher moves their images to a computer. The
computer stores data for one round trip
along with time marks, GPS coordinates and train
speed. While overnighting in the Oakland yard,
data are downloaded to a landside server. There
is storage for about 6 trips.
33
Video System Schematic(ENSCO, Inc.)
When, where, speed
Locomotive
34
Oakland Yard


Caltrans HQ
35


36
Automated Railroad Grade Crossing Violation
DetectionRTA- 65Y350

• Zu Kim and Ted E. Cohn
• California PATH
• University of California, Berkeley

37
Introduction

• Railroad crossing accidents
• 2,895 accidents (306 deaths) in 2000 (DOT)
• It is important to understand the factors
  underlying grade crossing crashes, and examine
  potential solutions
• Use video data to examine RXing violations
• Install video cameras in front of a train
• Study recorded accident scenes and near misses
  (violations)

38
Data Challenge

• San Joaquin Rail Corridor
• About 280 miles from Bakersfield to Emeryville
• About 700 crossings
• MPEG (320x240, 30 fps) GPS (each sec.)
• Requires automatic data reduction
• More than 10 hours of video data per week
• Small of useful scenes

39
Data Reduction

• By GPS
• GPS coordinates of crossings are known
• Collect scenes only near the crossings
• By computer vision
• Automatic detection of crossing violation
• Moving object detection from moving camera
• Fast detection (pseudo-realtime)
• Needs state-of-art computer vision technology

40
Example Image
41
Algorithm
Image1
Image2
Corner Detection Matching
Ego Motion Estimation
Moving Object Detection
Results fromPreviousVideo Frames
Consistency Check
Moving Objects
42
Corner Detection Matching
43
Motion Detection

• Ego-motion (camera motion) estimation
• Not a static scene (moving objects exist)
• Corner matching is not perfect
• Minimize the number of mismatches / moving
  corners
• Find consistent inconsistency
• Corners moving inconsistently w.r.t. camera
  motions for consecutive frames

44
Result
45
Conclusion Future Work

• Proposed fast (pseudo real-time) robust
  algorithm for railroad grade crossing violation
  analysis
• Need more thorough evaluation
• Generate a learning data set by showing video
  clips to human
• Refine algorithm with the learning data
• Compare with the human detection result