Method for efficient target detection from images robust to occlusion

• US 20110050940A1
• Filed: 09/01/2010
• Published: 03/03/2011
• Est. Priority Date: 09/01/2009
• Status: Active Grant

• Current Assignee: Fondazione Bruno Kessler
• Original Assignee: Fondazione Bruno Kessler
• Inventors: Messelodi, Stefano, Lanz, Oswald
• Granted Patent: US 8,436,913 B2
• Time in Patent Office: Days
• US Class Current: 348/222.1
• CPC Class Codes: G06T 2207/10016 Video; Image sequence
• CPC Class Codes: G06T 2207/30232 Surveillance
• CPC Class Codes: G06T 7/251 involving models
• CPC Class Codes: G06T 7/277 involving stochastic approaches, e.g. using Kalman filters

Abstract

The method for efficient target detection from images robust to occlusion disclosed by the present invention detects the presence and spatial location of a number of objects in images. It consists in (i) an off-line method to compile an intermediate representation of detection probability maps that are then used by (ii) an on-line method to construct a detection probability map suitable for detecting and localizing objects in a set of input images efficiently. The method explicitly handles occlusions among the objects to be detected and localized, and objects whose shape and configuration is provided externally, for example from an object tracker. The method according to the present invention can be applied to a variety of objects and applications by customizing the method'"'"'s input functions, namely the object representation, the geometric object model, its image projection method, and the feature matching function.

Claims

1. A method for efficient and occlusion robust target detection from images captured by cameras, said method comprising the steps of:

a) collecting information about geometric and calibration parameters regarding the capturing cameras;

b) choosing a suitable numerical representation for the object'"'"'s configuration to be recognized;

c) selecting a geometric object model that is parameterized with the chosen numerical representation;

d) implementing an image projection procedure that maps the object model onto an image according to a given state hypothesis;

e) implementing a likelihood function between the rendered object model and image features that peaks on those object states that fit an un-occluded instance of the object in a real image;

f) compiling, with the above calculated likelihood functions, the intermediate representations of detection probability maps, hereafter referred to as pixel support maps;

(off-line method)

g) processing, with the above calculated likelihood functions, a set of live images to provide the set up of detection probability maps from pixel support maps computed in step f) on a set of input images. (on-line method)

2. The method of claim 1, wherein said suitable numerical representation for the object'"'"'s configuration to be recognized, in case of a rigid and vertically symmetric object moving on a horizontal plane, is a two dimensional vector describing the objects coordinates on the horizontal plane.

3. The method of claim 1, wherein said likelihood function implemented at the previous step e) is a combination of elementary functions that operate on single feature pixels.

4. The method of claim 3 wherein said combination is a sum or a product of elementary functions.

5. The method of claim 1, wherein said step f) of compiling the intermediate representations of detection probability maps (off-line method) is performed as follows:

h) a finite number of suitable object states to be detected is chosen;

i) the state is first rendered using the image projection method and the camera'"'"'s calibration information;

j) each pixel activated by the above rendering is accessed in sequence and considered to be an active feature pixel with weight equal to 1;

k) the contribution of said activated pixel to the likelihood function calculated for the rendered state image, divided by the number of activated pixels in the rendered state image, is computed;

l) steps i) to k) are repeated for each state in the chosen set,

m) the set of all values computed at the previous step for a given pixel are stored in a list that is called the pixel'"'"'s support map wherein each entry is coupled with its originating target state;

n) the list of entries compiled at the previous step is sorted according to the distance of the corresponding state to the camera, with the closest one in front of the sorted list;

o) steps i) to n) are repeated for each camera.

6. The method of claim 5, wherein said step h) is performed superimposing a grid on the state space and choosing the centers of the grid cells.

7. The method of claim 5 characterized in that for PDL (People Detection and Localization) tasks with a regular state space grid, said support maps can be displayed as grey level images.

8. The method of claim 1, wherein step g) is performed as follows:

p) a pre-processing feature extraction step is applied to each image to be processed;

q) if a list of states of external objects are provided as external inputs, each state of external objects is given a positive weight, and the sum of weights associated to the same target must equal 1 in a way that the weighted states represent a probability distribution for each of the external objects;

r) if external inputs are provided, the list of object states provided externally is sorted according to their distance to the camera, with the closest one in front of the list. If no external input is provided, this list is considered to be empty;

s) for each active feature pixel—

like the pixels with non-zero value in the feature image obtained at previous step p)—

the corresponding pixel support map is accessed, and a unique list is created by merging the support map with the list of externally provided object states, and this list is sorted according to camera distance;

t) if present, each external object is associated a real number whose value is initialized to 0;

u) the entries of the joint list are accessed sequentially, from front to back. If an entry belongs to an external object and the considered feature pixel falls inside its rendered geometric model the real number is incremented with the value of the corresponding state weight. If, instead, the entry belongs to the support map the following is done;

The real values r_i of all the external objects indexed by “

i”

are used to compute a new value π

_i(1−

r_i) which reflects the probability of occlusion by an external object on that pixel at the distance of the considered support map state from the camera, this value is then multiplied with the value of the support map entry and accumulated in the entry of the detection probability map corresponding to the considered support map state.

v) Steps s) to u) are repeated for all the active feature support maps of the activated pixels, for all cameras.

9. The method of claim 8, wherein said pre-processing feature extraction of step p) comprise motion detection followed by edge detection and distance transform.

10. The method of claim 8, wherein the list of states of step p) is delivered by an object tracker or provided by manual insertion wherein, for each external object, an image projection procedure is implemented.

11. Apparatus for efficient and occlusion robust target detection comprising means for the implementation of the method according to claim 1.

12. Computer program comprising computer program code means adapted to perform all the steps of claim 1, when said program is run on a computer.

13. A computer readable medium having a program recorded thereon, said computer readable medium comprising computer program code means adapted to perform all the steps of claim 1, when said program is run on a computer.