An Adaptive Windowing Prediction Algorithm for Vehicle Speed Estimation

1
An Adaptive Windowing Prediction Algorithm for
Vehicle Speed Estimation

• Tun-Wen Pai and Wen-Jung Juang
• Dept. of CS, National Taiwan Ocean University,
  Taiwan
• Lee-Lyi Wang
• Dept. of EE, Tung-Nan Institute of
  Technology,Taiwan

2
Outlines

• Motivation
• System Configuration and Assumptions
• System Algorithms
• Modules in Phase I
• Modules in Phase II
• Simulation Results and Conclusions
• Future Works

3
Motivation
4
Motivation
5
Motivation
6
Basic Assumptions

• 1.The speed is finite and its value is
  nonnegative.
• 2.The speed and direction cannot change suddenly.
• 3.Drive in the permitted direction on the road.
• 4.The size is realized from a know distribution
• 5.The height of CCD and the distance between the
  CCD and Base line on screen is known.

7
System Configuration
8
Modules in Phase I

• To acquire video image sequences
• From road-side CCTV
• 320X240 pixels
• Fifteen frames per second
• Perform CCD noise margin calibration
• Tolerant parameter em

9
Modules in Phase I (cont.)

• Moving Object Detection
• gi (x,y) fi (x,y)-fi-1 (x,y) ?em AND
• fi1 (x,y)-fi (x,y) ?em

10
Moving Object Detection Example ( i-1th frame
fi-1 )
11
Moving Object Detection Example ( ith frame fi )
12
Moving Object Detection Example ( i1th frame
fi1 )
13
Moving Object Detection Example ( gi )
14
Morphological operations

• Dilation
• Erosion
• Structuring Element 3x5 square element

n
15
Morphological example
16
BCC and Region Growing

• Block size more then 5x5 we see it as seed
• Region growing criteria
• 1. different not more then 2(em )
• 2. pixel add to region at least one of his
  8-neighboring pixel previously included in the
  region.
• Bounding Box size larger than 5x5 we see it as
  candidate vehicle objects.

17
BCC and Region Growing Example
18
BCC and Region Growing Example(cont.)
19
Modules in Phase II

• AWP Block Matching
• Inverse Perspective Transformation
• Statistical Analysis

20
Adaptive Windowing Prediction (AWP)

• Block Matching Algorithm (BMA)
• Mean Absolute Difference (MAD)
• Uni-direction problem
• Uni-model error surface

21
Uni-model error surface
22
AWP algorithm
23
AWP algorithm (cont.)
24
AWP algorithm summarized

• 1.Use full search found initial moving distance
• 2.MAD for predefined adaptive checking point set
• 3.Find a temporary minimal MAD point
• 4.Checking the 4-neighbor , if all MAD measured
  then stop, else calculate un-measured neighbors
• 5.If new MAD value greater then old MAD then stop
• else replace the temporary minimal MAD point
• 6.Go to step3,until AWP found minimal MAD or
  until moving object vanished from screen
• 7.Use look-up table to obtain speed, and provide
  pixel distance to next coming frame

25
Full Search BMA Initial Moving Distance
Adaptive Checking Point set MAD(V)
Temporary Minimal MAD p(i,j)min(MAD(q))q V
4-neighboring points MAD measurement
No
Minimal MAD ?
Yes
Measuring Moving Distance
Look-up Table
Next coming frame
26
Partial Look-up table
Distance
Position
27
Inverse Perspective Transformation
28
Inverse Perspective Transformation (cont.)
29
Statistical Analysis

• Obtain a set of sequential speed data for each
  observed vehicle,we can get a mean speed for each
  time slot.
• So we can obtain it form www or wap

30
Simulation Examples

• The mean value of tolerant margin for CCD is
  2.45(with gray-level scale0,255)
• There are vehicle objects detected from frame
  number 450 to 600
• The average speed is about 30km/hr
• Only 8.6 of the computations(in terms of check
  points) are required in AWP algorithm with
  respect to the Full Search algorithm

31
Examples
Original Video
Detected Vehicles
32
Future Work

• Difference Weather Condition
• Size distribution
• CCD height and base line distance
• Morphological structure element