1. Human tracking and counting using the KINECT range sensor based on Adaboost and Kalman Filter
ISVC 2013

2. Problem
Human tracking
Avoid occlusion

3. Human Detection Observations:
There is an empty space in the front and back of head
The right side of right shoulder and the left side of left shoulder are also empty
There is a height difference between the head and the two shoulders
The first step of our tracking is to detect humans for each frame.
How to describe the spatial information of 3D HASP

4. Human Detection
Those criteria can be formulated as the difference of two pixel areas in the depth map
Haar-like feature
Adaboost is introduced to construct a strong classifiers from those weak criteria

5. Human Detection by Adaboost
Framework

6. Spatial feature
Processing window
20 redefined sub-windows

7. Spatial feature
Four Haar-like features

8. Depth integral image
The sum of rectangle pixel values from the top-left corner to a pixel in depth image
To speed up the computation of Haar-like features
All pixel intensity values of D:

9. Adaboost algorithm
Construct a strong classifier by a weighted linear combination of weak classifiers

10. Our Classifier Challenge Solutions
Human can stand and face all directions with many postures
Solutions
Combine a horizontal strong classifier and a vertical strong classifier

11. Horizontal Strong Classifier
Formulation

12. Vertical Strong Classifier
Formulation

13. Training
Took many depth maps of each object by rotating a certain degree
720 positive images negative images

14. Results Testing on three datasets:
Dataset 1: only one human object standing in different directions
Dataset 2: Two human objects
Dataset 3: three or more human objects

15. Results (Dataset 1)

16. Results (Dataset 2)

17. Results (Dataset 3)

18. Choice of window sizes
If the window size is smaller, the human standing with a big Head and Shoulder Profile (HASP) can not be detected. If the window size is bigger,
two humans standing closely can not be detected, which also decrease the detection rate.

19. Limitation
Fails if detected humans are standing two very close to each other
Improve tracking accuracy by incorporating Kalman Filter, since the closing time is short in real tracking application.

20. Conclusion
We construct a real-time human detection based the depth image from Kinect sensor
Head and Shoulder Profile described by some Haar-like features is incorporated into Adaboost algorithm to detect human objects.
Detection time for each image is about 33 ms.