Method for setting a tridimensional shape detection classifier and method for tridimensional shape detection using said shape detection classifier

1. A method for setting a tridimensional shape detection classifier for detecting tridimensional shapes from depth images in which each pixel represents a depth distance from a source to a scene, each depth image being dividable into one or more patches (I p , I n ) of given dimensions, the classifier comprising a forest of at least a binary tree for obtaining the class probability of a given shape comprising nodes associated with a distance function (f) that taking at least a pixel position in a patch calculate a pixel distance, the method comprising:

obtaining one or more positive patches (I p ) of the given dimensions from a depth image, said one or more positive patches containing the shape to be detected,

and obtaining one or more negative patches (I n ) of the given dimensions from a depth image, said one or more negative patches not containing the shape to be detected,

the method comprises:

for each obtained positive or negative patch

traversing the binary trees starting from their root node using the distance function at each node to decide to continue to one of the child nodes of the next level until a leaf node of each binary tree is reached;

calculating in each reached leaf node the distance function for the patch using candidate groups of parameters, each candidate group of parameters comprising:

at least a pixel position (u, v) in a patch,

a depth clipping window (k) in the patch and

a pixel distance threshold (θ);

by comparing the result of the distance function using the at least a pixel position and the depth clipping window with the pixel distance threshold of each candidate group, and computing its statistics;

and when more than a predefined number of positive or negative patches are applied to a leaf node of the classifier:

selecting for that leaf node the candidate group of parameters that best separate the positive and negative patches into two groups for calculating the class probability of the shape in that leaf node using the distance function;

creating a new level of the binary tree from that leaf node comprising two newly created leave nodes, thus that leaf node becoming a node, and

passing the statistics from that leaf node that has become a node to the newly created leaf nodes.

2. The method according to claim 1 , wherein each candidate group of parameters comprises

at least a random pixel position (u, v) in a patch,

a random depth clipping window (k) in the patch and

a random pixel distance threshold (θ).

3. The method according to claim 1 , wherein at least part of the one or more positive patches and one or more negative patches are obtained from the same depth image.

4. The method according to claim 3 , wherein the obtained one or more negative patches are the ones in each depth image with highest positive class probability according to the statistics at the reached leaf nodes of the binary trees.

5. The method according to claim 1 , wherein the distance function calculates the relation between the depths represented by two pixels of the patch located in the random pixel position (u, v) from the center (x) pixel of the patch, normalized with the depth evaluated in the center pixel, each depth being upper and lower limited by the random depth distance clipping window centered in the value of depth represented by the center pixel of the patch, as per the formula:

f

⁡

(

I

,

x

)

=

max

⁡

(

min

⁡

(

I

⁡

(

x

+

u

I

⁡

(

x

)

)

,

I

⁡

(

x

)

+

κ

)

,

I

⁡

(

x

)

-

κ

)

-

max

⁡

(

min

⁡

(

I

⁡

(

x

+

v

I

⁡

(

x

)

)

,

I

⁡

(

x

)

+

κ

)

,

I

⁡

(

x

)

-

κ

)

.

6. A method for shape detection from a depth image using the shape detection classifier of claim 1 , comprising the steps of:

dividing the depth image into patches of given dimensions;

traversing the binary trees by applying the distance function at each visited node using the associated pixel displacement (u,v) and maximum pixel distance (k) and comparing the result of the distance function with the associated pixel distance threshold (θ) and;

obtaining the probability (p) of a specific shape leveraging on the statistics of the leaf nodes reached in each tree.

7. The method according to claim 6 , further comprising:

averaging the probabilities of a specific shape obtained from the different leaf nodes of binary trees of the forest according to:

p

⁡

(

c

❘

I

,

x

)

=

1

M

⁢

∑

m

=

1

M

⁢

⁢

p

l

m

⁡

(

c

❘

I

,

x

)

Where m=1 . . M represents the different leaf nodes of the binary trees.

8. The method according to claim 6 , further comprising the steps of:

casting a vote for a target shape per each patch whenever the probability of said shape is higher than the probability of not being said shape or being another shape, and

estimating a probability density for the target shape using said votes.

9. The method according to claim 6 , further comprising taking into account temporal consistency by recursively updating the probability density with votes aggregated from past time instants as follows:

p ′( c|I t )=α p ( c|I t )+(1−α) p ′( c|I t-1 ).

10. The method according to claim 6 , further comprising detecting the pixel location of the target shape as the pixel with maximum probability.

11. The method according to claim 6 , wherein the probability is thresholded by locally integrating the probability measure through a circular surface of radius inversely proportional to the depth and centered at the global maximum

V

=

∑

x

∈

𝒮

⁢

⁢

p

′

⁡

(

c

❘

I

t

⁡

(

x

)

)

.

12. The method according to claim 6 , wherein the shape is a hand gesture.

13. A data processing apparatus comprising means for carrying out the method of claim 1 .

14. A computer program stored on a non-transitory computer-readable medium and adapted to perform the method of claim 1 .