**5. Conclusions**

The parameters that define the behavior of the particle estimator were described in paragraph 3. Although an analysis of the estimator's response can be made based on each of these parameters; here we will perform only one, for the number of particles of the estimator.

For this purpose, we rely on the experiment corresponding to the tracking of the six-rotor UAV, because it was simultaneously tracked by a positioning system. Therefore, taking as true the observation obtained from the positioning system, it was possible to determine the relative error corresponding to the particle estimator.

To observe the dynamic response and the internal state of the particle estimator, the relative error (upper graph) and the number of valid particles (lower graph) as a function of time were plotted on **Figure 10**.

This figure show the behavior of the estimator, as a function of the number of particles and in different colors, for values of 30 (red), 50 (magenta), 70 (blue) and 90 (green), the time between marks is the acquisition video time, which is 0.050 seconds.

As can be seen always at least a number of 6 frames are needed to begin reach the lock state, and the number of particles parameter does not affect this number to get this state. But when the number of the particles of the estimator it increase, decreases its relative error, increase the computer time process, and presents greater stability to maintain the lock.

Another parameter that improved the particle estimator response is the dispersion multiplying factor of the state variables. This multiplying factor is increased when the number of valid particles decrease, covering a wider area of search in the

**Figure 10.** *Dynamic response of the particle estimator.*

image. With this search strategy the estimator maintains its stability in the response to the lock with the objective, in a few frames.

After several execution of the algorithm in the experiment b); and varying different initialization parameters of the estimator, the values that produce an acceptable behavior are: number of particles 70 or higher, number of characteristic points of the SURF algorithm 70 or higher, threshold value of probability of similarity 90% or higher.

With the considerations in the initialization parameters of the particle estimator mentioned above, the state lock could be maintained with no more than two frames unlocked. And another observed feature is that the estimator could track the target even when the object is mostly occluded on its surface, for example when the UAV is behind trees (see **Figure 8**).

As a proposed improvement to enhancement to achieve better detection of the target when it is mainly occluded, or when the image is heavily contaminated with noise; it is to reformulate the main strategy. This strategy consists of decomposing the reference image into *NxN* sub-images, in sequence and with their corresponding identification. And for each sub-image a particle estimator is applied, having *NxN* particle estimators, each one looking for a part of the reference image. From there and starting with the first particle, we look for the particles with more probability, and that are located in the correct sequence; the average of these particles will give us the more likely position of the target to be follow. In case that no particles are found in the correct positions we can obtain the most probability position of the object to be followed as the position of the particle with the largest probability, or as the integration of the information provided by the *NxN* estimators.

#### **Acknowledgements**

Special thanks to the Post Graduate staff of the National Technological University – Buenos Aires Regional Faculty. And the staff of the Laboratories of the

Institute of Scientific and Technical Research for Defense: Thermal Imaging Laboratory for providing the images of the experience, and to the Digital Techniques Laboratory for carrying out the flight with the unmanned aerial vehicle.
