Você está na página 1de 22

D.M. Gavrila, V.




Advanced Computer Vision - ACS 2015 Zafiu Ioan - Florin


The paper presents an efficient shape-based object detection

method based on Distance Transforms and describes its use for
real-time vision on-board vehicles. The method uses a template
hierarchy to capture the variety of object shapes; efficient
hierarchies can be generated offline for given shape distributions
using stochastic optimization techniques (i.e. simulated annealing).

With the help of different sensors, “smart” vehicles can provide the driver
relevant information about the surroundings or even perform simple vehicle
control tasks.

Advanced on-board vision systems are used to detect objects.

For example traffic sign detection and recognition. Another idea is to have
a system that detects pedestrians and by taking passive or active measures the
accident rate can be reduced.
DT - distance transform
Matching with DT involves using two binary images, a segmented template T(feature template) and a
segmented image I(feature image).
What is it trying to be achieved?
- The point is to determine the presence of a template in a given image
Matching using a DT
(a) original image (b) template (c) edge image (d) DT image

So far we have looked at matching single edge templates to an

image. In the real world however, objects tend to appear in many
different shapes:

● The viewpoint can change

● The object might actively change its shape(pedestrians)

A lot of templates need to be created. For real-time purposes, it

would be to many to match individually.
Matching using a template hierarchy
At a coarse level of search, when the image grid size of the search is large,
it would be inefficient to match each of the N objects separately, if they are
relatively similar to each other. Instead, one would group similar templates
together and represent them by a prototype template; matching would be done
with this prototype, rather than with the individual templates, resulting in a
speed-up. This grouping of templates is done at various levels, resulting in a
hierarchy, where the leaf level contains the N templates one needs to match
with and the intermediate levels contains the prototypes.
Coarse to fine hierarchical search
● Start at the root template and try to find a match in the image. Choose the distance threshold to
be large enough so that the match could potentially contain any of the child-nodes.
● If a match is found, descend down the tree, and try to match the next level of templates (by
focusing only on the area in the image that has been matched by the parent). Now use a smaller
distance threshold that is still large enough to possibly contain each of the child-templates.
● Repeat this process (usually using depth-first search) until one of the leafs matches.
● How much speed is it gain? Gavrila and Philomin say, “Up to three orders of magnitude”, but
depends on various factors

Pseudo-C-code for the original (sequential) implementation of the

computationally-intensive chamfer transform (with x-y kernel, forward
pass, image width W, image height H).
SIMD Pseudocode

To illustrate the proposed matching method, a series of on board (test vehicles)

and off-line experiments were conducted.

The method was applied to traffic signs and pedestrian detection.

A three level template hierarchy was used, traversed in a depth-first order.

On-board camera and display
Traffic signs
The application was used to detect circular and triangular signs.
Templates for circles and triangles with radii in range of 7-18 pixels were used.
A manually specified template tree was used, containing 36 templates.

Result - 80% to 95% recognition rate, based on visibility

A hierarchy for traffic sign shapes (hard-coded)
Traffic sign detection (and recognition)
1100 shapes were used, on 5 scales(70-102 pixels) resulting in a 5500
templates hierarchy.
The system running at 1-5 Hz.

Preliminary results showed 75%-80% detection rate when requiring false

positives to be 2 or less.
A hierarchy for pedestrian shapes (partial view)
Pedestrian detection results
False positives
● Speed is impressive, DT and Hierarchical Organization can improve the
speed dramatically.
● Detection results are only as good as the underlying edge-data.
● Ambiguity and False Positives are a serious issue. Oriented edges can
potentially help.
● It is questionable whether this system is ready for production.
○ Most Modern Smart Cars seem to rely mainly on active sensors (LIDAR & RADAR)
Gavrila, V. Philomin - http://remus.ulbsibiu.ro/teaching/courses/docs/acs/iccv99.pdf