Determining Radius

1
Development of an Automated Intersection
Accident Detection System
by
Yunlong Zhang, Ph.D., P.E. Texas AM University
January 9, 2005
2

• Incident and Congestion
• Incidents are major cause of the congestion
• 50-60 of delay on urban freeways
• Will cost US 75 billion in lost productivity
  and 8.4 billion gallons of wasted fuel in the
  year 2005

Vehicle accidents play a significant role in
causing non-recurring congestion
3

• Accidents on Urban Streets
• A large percent of traffic accidents on urban
  streets occur at or near intersections
• more than 2.8 million intersection-related
  crashes occurred in the U.S.approximately 44 of
  all crashes reported in 2000
• over 8,500 fatalities (23)
• nearly 1 million injuries (more than 48)

4

• History of Automated Accident Detection
• Freeway incident detection can be dated back to
  1960s
• Accident detection research on urban streets has
  been lagging
• Algorithms used pattern recognition, fuzzy set
  and neural networks and video image processing
  techniques
• Other than video-based system, detection
  algorithms are indirect
• Most systems require expensive instrumentation
  and complicated algorithms
• Many systems have reported unsatisfactory
  performance

5

• Research Approach
• Audio Based System
• Digital signal processing
• Direct detection

6
Normal Sound Signal
7
Crash Sound Signal
8
Detection system Components
Digital signal
Feature vector
Reduced feature vector
9
Data Collection / Data Processing

• Signal recorded using Sony TCD-D8 Digital Audio
  Tape recorder.
• Crash signals obtained from Texas Transportation
  Institute (TTI) crash test facility.
• Database created using crash sounds, normal
  sounds, and special sounds (brake, pile-driving,
  etc.)
• Normal sound-62
• A three-second signal has 66176 samples.

10
Synthesizing the signals
11
Feature Extraction methods

• Discrete Wavelet Transform (DWT)
• Mother wavelets Haar, Daubechies, symlets,
  coiflets
• Fast Fourier transform (FFT)
• Discrete Cosine Transform (DCT)
• Real Cepstral Transform (RCT)
• Mel-frequency Cepstral Transform (MCT)

12
Feature Optimization

• The feature vector obtained using one of the
  transform method is optimized using Fishers
  Linear Discriminant Analysis (LDA).
• The output of the LDA is an optimal linear
  combination weight matrix.
• The number of samples in a
• three-second signal is reduced using feature
  extraction method and further optimized thereby
  reduced to very few coefficients.

13
Statistical Classification

• Three types of classifiers were investigated
  for this system
• Classifier must be trained using the data for
  which the user knows the correct
• Classification
• Nearest mean and maximum likelihood
• classifiers are parametric classifiers, and
  Nearest neighbor is a non-parametric
  classifier

F2
14
Testing method

• Testing method leaves one signal under
  investigation as the testing data and the rest of
  the data set as the training data.
• In the automated detection system, features are
  extracted from the signals in the database whose
  classifications are known using one of the
  transform methods and leaving one signal to be
  investigated.

crash
brake
pile drive
normal
15
Lab Testing to Determine Best Algorithms

• Comparison of transform-based feature extraction
  methods
• Comparison of mother wavelets of DWT for
  different classifiers
• Comparison of Statistical classifiers and system
  sensitivities

16
Maximum Likelihood Classification Accuracies for
Various feature extraction methods
17
Overall Classification Accuracies with DWT for
two-class system
18
Algorithm Performance Comparison

• Five different feature extraction methods
  analyzed RCT,MCT,FFT,DCT and DWT
• RCT, MCT, DWT performed best, and taking
  the computation time into consideration DWT is
  chosen.
• Different mother wavelets Haar, Symlet,
  Coiflets, Daubechies4, Daubechies5 are analyzed
• HAAR performed among the best and easy to
  implement
• Three different classifiers Nearest mean,
  Nearest neighbor
• and Maximum likelihood are investigated
• MAXIMUM LIKELIHOOD performed best
• An overall accuracy of about 98 is obtained
  using the algorithms selected selected.

19
Real-time Detection System Development
Audio Signal
Microphone
Feature Extraction
Feature Optimization
Extigy Sound card
Statistical Classification
Matlab Simulink
Label Crash or Non-Crash
20
Real-Time Testing Approach

• Continues 3 second signal segmentation
• Two testing methods
• Tested at intersection for normal conditions
• Playing back recorded signals (with crashes) in
• lab environment (Emulating real-world scenarios)

21
Real-Time Testing Data

• Crash signals from Kentucky Transportation
  Cabinet (KTC)s Auto Incident Recording
  System (AIRS)
• Normal Sounds collected in Jackson and
  Starkville
• Database contains various signal types
• Crash sound
• Brake/ Pile drive sound
• Bird/Siren sound
• Normal sound

22
Real-Time Testing Crash
23
Real-time System Testing Results
Two-Class System Testing Results
-
Overall accuracy
Crash

Non
crash


26

1

0.96

5

76

0.94


24
Real-time System Testing Results
Multi-Class System Testing Results
Brake/Pile
Bird Chirping

Overall
Crash

Normal

driving

/Siren

accuracy

25

0

2

0

0.93

0

61

0

1

0.98

1

2

7

0

0.70
0

2

0

7

0.78


25
Real-time Performance

• Fast Detection Detecting Crashes in a few
  seconds (real-time)
• High Detection rate
• Low False Alarm rate

26
Future Work

• Implementation at selected Intersections
• Real-world testing and fine-tuning
• Integration of Audio/Video components
• C-coding for multi-intersection system
• Integration with other incident management
  components

27
(No Transcript)