IEEE Robotics & Automation Magazine - September 2017 - 132

Captured Frame

(b)

(c)

(d)

(e)

(f)

(g)

(h)

(i)

(j)

Recognition
Result

(a)

Figure 7. The qualitative experimental results in a semistructured scenario in which both (a)-(e) the camera viewpoint and (f)-(j) the
object location and orientation were continuously changed. All of the objects were successfully recognized even when their orientation,
location, or the camera's viewpoint was changed.

orientation or location in the scene or the cameras changed
their viewpoint.
With the purpose of validating the obtained qualitative
results, a quantitative evaluation was carried out. In this case,
the true positive rate (TPR) and false positive rate (FPR)
measurements were used [80], i.e., the proportion of correctly classified positives (TPR) and the proportion of incorrectly
classified negatives (FPR). So, a good performance was
obtained when the TPR was close to 1 while the FPR was
near 0. The performance analysis of the proposed approach
resulted in a 0.780 TPR and a 0.001 FPR, which indicates
good performance and that the system was successful in the
object detection/recognition task.
Experiment 2: Real Scenarios
In this experiment, the objects to be detected and recognized were placed on a desk. Two unstructured environments were used, composed of everyday objects of different
nature and features, such as textured books, pens, and
clocks. In this context, the objects to be detected and recognized included a red ball, a toy car, a yellow ball, a green
bulb, a stapler, and a wooden generalized cylinder. These
objects were located at different positions and/or orientations within the considered scenario, resulting in partial
occlusion in some cases. As in the previous case, a human
was continuously interacting with the target objects but not
with the robot system, so that the motion cue triggered a
visual attention focus. However, the other two visual cues
were required to distinguish between the target objects and
other moving elements in the scene, such as the person. For
this reason, the three cues had the same weight in the object
recognition process.
In the first experiment, three different objects were used: a
toy car, a stapler, and a wooden generalized cylinder. Some of
the one-trial results are shown in Figure 8. Note that, despite
132

IEEE ROBOTICS & AUTOMATION MAGAZINE

SEPTEMBER 2017

the nature of the environment and that of the objects themselves, all the targets were properly detected even in the case
of the toy car, which had a great color similarity with its background. Figure 8(i) also shows an example of the simultaneous recognition of two objects in the same image, in which
the car and the stapler were correctly detected and recognized. In a similar way, the developed approach adequately
focused its attention on the target object (i.e., the generalized
cylinder), although several objects were added to the scene
(the toy car and the stapler), as shown in Figure 8(j).
In the subsequent experiment, the visual system was
aimed at detecting and recognizing four objects (a red ball, a
green bulb, a yellow ball, and a wooden generalized cylinder)
while a person was interacting with the objects in the scene,
changing their positions on the desk. As a consequence, the
motion cue again played a main role in the object recognition
process. Some of the obtained one-trial results are presented
in Figure 9. In this case, unlike in previous examples, the
binary image is shown, highlighting the detected objects,
especially when they are partially occluded or color similarity
with the background is considerably high.
Once more, a quantitative analysis validates the prior mentioned qualitative results. In this case, the detection and recognition resulted in a 0.898 TPR and a 0.0009 FPR,
confirming the quality of the approach performance when
real-life scenarios are considered.
Experiment 3: Image Repository
For the third validation experiment, we compare the performance of our approach with state-of-the-art methods by
using a public image repository. Given that the ability to recognize objects is crucial for many applications, a wide range
of public image repositories are available. These data sets
allow researchers to evaluate their approaches with a large
number of objects and under different conditions as well as

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2017