Detecting cars and people in the parking lot video

1
Detecting cars and people in the parking lot
video
18798 Image, Video and Multimedia

• Euiseok Hwang
• Sheethal Bhat
• Electrical and Computer Engineering Dept.
• Carnegie Mellon University
• May 4 , 2007

2
Outline

• Schematics.
• Static analysis.
• Geometrical segmentation.
• Features for parked car detection.
• Semi-adaptive background estimation.
• Dynamic analysis.
• Motion detection.
• Features for moving objects.
• Classification.
• Results.

3
Schematics
4
I. Static Analysis

• Geometrical segmentation.
• Features for parked car detection.
• Semi-adaptive background estimation.

5
Geometrical Segmentation

• Segment image into parked area and road area (2
  area masks).
• Divide parked area into 55 slots (55 parking
  slot masks).

6
Training the Parking Slots

• Feature extraction of each slot.
• Mean.
• Standard deviation.
• Mean and std dev are good features for
  separating parked slots from empty ones.

7
Parked Car Detection I

• Testing with only two features, mean and std
  by nearest neighbor classification.
• 98.63 accuracy

8
Error Analysis

• Errors are caused by objects moving in the
  parking areas.
• Due to the car and the person crossing the
  parking slots.

9
Movement Detection

• Detect movement in the parking slot.
• Mean square distance (MSD) between frames.
• Approximate the transition time from MSD.

10
Parked Car Detection II

• Update result with the transition information.
• Ignore double transitions in one sequence.
• 99.92 accuracy

11
Background Estimation

• For changed parking slot.
• Search transition frames.
• Average selectively, before and after the
  transition.
• Other regions.
• Average all frames without upper and lower 20
  intensities.

12
Background Subtraction Example
Background from averaging whole sequence
Semi-adaptive background
13
II. Dynamic Analysis.

• Motion detection.
• Features for moving objects.
• Classification.

14
Threshold the Difference

• The difference Image is noisy after
  thresholding.
• To minimize error we take the maximum of
  difference in 3 channels.

15
Filter the Image

• Simple Median Filter removes the noise
  components.
• Open areas.
• Object may not always resemble a recognizable
  shape.

16
Bounding the Object

• Need to get exact width and position of the
  object.
• Need to reduce the number of squares drawn around
  an object.

17
Find the best fit square

• Need closely standing objects to be as well
  separated as possible.
• Recognize the overlapping boxes around the cars
  are not different objects.

18
Find the bounding areas

• Algorithm to find the closest bounding areas for
  an object of any shape.
• Usually maximum 4 number of iterations to
  converge.
• Features are upper left co-ordinates, height,
  width.

19
Classification Algorithm

• The boundary requires special criteria in
  classifying.
• Make sure that enough number of samples for
  velocity.
• In the boundary calculate the Maximum width over
  frames.

20
III. Results

• Final result video.
• Detection errors

21
Detection Result
Morning Sequence 2
Afternoon Sequence 2
Click for media player
Click for media player
22
Detection Errors

• Afternoon Seq.
• Seq 1 1
• Seq 2 70
• Seq 4 3

• Morning Seq.
• Seq 1 6(1)
• Seq 2 30(17)
• Seq 4 24(5)