IEEE Robotics & Automation Magazine - June 2011 - 29

pieces of furniture in 3-D point clouds
representing indoor environments.
To carry out the experiments, we
first downloaded the CAD models of
a chair, a table, and a sideboard
(colored as blue, red, and yellow in
Figure 1). These objects were obtained from different Web catalogs.
Following the procedure given in the
"Training Objects from Web Catalogs" section, we placed the objects in
front of the simulated laser and
rotated them around the up axis,
obtaining 16 scans of each object that
were segmented using our algorithm.
A final common vocabulary of parts
was created with k ¼ 40. These values provide a good balance between
performance and computational costs
(see Figure 6).
To obtain point clouds from real
scenes, we used a Hokuyo laser
mounted on a pan tilt on a robot
approximately 1:5 m above the floor
(see Figure 1). Service robots working in human environments typically have their 3-D sensor at that
height such that its perspective is
similar to that of a person. Both of Figure 5. Two examples of the whole process (see the legend in Figure 3).
the robots we used had such a setup,
and since they both ran ROS
(www.ros.org ), changing the robot was transparent to our detector is shown in Figure 3. The two tables were correctly
algorithm. However, different robots or sensors at different detected, and the system was able to detect five chairs out
heights can be used if the scanning of the training data is of seven, although including one false positive. An addiadapted accordingly. Finally, the point clouds were seg- tional experiment was carried out in a seminar room conmented using the same algorithm as for the training data. taining a big table and 15 chairs around it. The resulting
Each obtained part activated the best two vocabulary classifications are also shown in Figure 3.
entries.
Data acquisition using our laser takes 10 s. Smoothing,
segmenting, and extracting the features takes another 7:12 s
Recognition of Objects in Real Environments
on average using Cþþ code on a single core. The time
The first experiment was carried out in an office of our floor needed to select the final hypotheses for an object center
at Technische Universit€at M€
unchen. The corresponding seg- mainly depends on the total number of parts in the scene
mented 3-D point cloud is shown in Figure 3. This office and the number of clusters in the vocabulary (since each of
contained two tables, different chairs, and a sideboard. We the parts should activate the closest clusters). It also depends
would like to point out that the objects located in the office on the number of final selected hypotheses because they are
were similar to but not the same as the 3-D models we down- checked for inconsistencies (see "Model Fitting and Verloaded from the Web. In particular, the table in the middle ification" section). Using our current configuration and withof the chairs was round and we did not have any round table out any optimization, our approach took 0:20 s for the office
in our training set, and the chairs around the table were dif- and 0:24 s for the seminar using MATLAB on an Intel Core i5.
ferent from the training models. This setting aimed to demTo quantitatively evaluate the performance of our aponstrate that our approach is well suited for detecting proach, we generated recall versus (1 - precision) plots for
categories of objects instead of concrete instances only.
the two previous scenes. Different false-positive values were
We set the thresholds for the final hypotheses to 1:0 for obtained by increasing the threshold for the final hypotheses
tables, 0:6 for chairs, and 1:4 for sideboards. These values selection. A true positive is obtained when the distance
were selected as desired working points in classification between the detected center and the original center is less
performance plots (Figure 7). The result of applying our than half of the width of the object. The test set included two
JUNE 2011

IEEE ROBOTICS & AUTOMATION MAGAZINE

http://www.ros.org

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - June 2011