**8. Conclusion and discussion**

The evaluation demonstrated that the system is capable of autonomously detecting, counting and localization of objects with an accuracy of about 15–20 cm. It was proven that an optimal value for MA (20 cm) has to be a bit higher than the accuracy of the position system and that objects with the distance of 20 cm (MA) in each axis can still be distinguished. Also the coherence of the parameters MA and VFOV on the performance of the search and the detection errors was demonstrated. A smaller VFOV with a smaller MA leads to more double detections, while a too high MA leads to misses of nearby objects. As a general rule, too high VFOV leads to misses because some areas are not searched properly. In this context the acceptance tolerance, which was set to 25 cm in setup 2, is a parameter, which comes into effect. A waypoint is already marked as reached if the current position of the quadrocopter is within this tolerance. This can result in an incomplete cover of the search area and it explains why the BFS misses some targets at the side of the search area.

The best parameter for VFOV was 30 cm × 45 cm. This setting together with the best value for MA showed no detection error even in a challenging room with eight objects.

Furthermore, the evaluation proved that the DFS performed better than the BFS. The reason for that is the fact that smaller waypoint steps are less accurate than less big ones because of the set point jumps and the jump effect as well as the control and sensor system. These result simplified mean that also for a flying robot such as a quadrotor using the underlying on‐board sensors less turns and commands are better. This could already be demonstrated in previous experiments [21]. The reason for that can be found in the dynamic of the quadrotor as an aerial vehicle with very little friction (air) and the on‐board optical sensors, which are especially affected by the behaviour of the system. Rotations, which mainly occur after set point changes, are a source of error for the position determination.

Although the system was proven capable of performing autonomous and challenging search, count and localization missions, the evaluation of the system did not show a very high accuracy according to the determined positions and the fact that optical sensors were used, which generally can reach higher accuracies. There are multiple sources for accuracy errors, which start from the manually measured and placed target positions in a region of several centime‐ tres. The next major source of error is the starting error, which means the wrong position measured by the optical flow during lift off and the wrong initial position and orientation or placement error of the quadrocopter on the starting position. An initial orientation error for yaw of only 1° leads to a position error of 5 cm after 3 m. It is most likely that the initial yaw orientation error was sometimes in the range of a few degrees. These are good explanations for the high systematical error, which can be seen in the data. A proof of this fact is given by a closer look at some raw data. They demonstrate that the accuracy for the closer object is much better than for the farer object, even if the closer object is detected later in some cases. The best explanation for this is an initial yaw orientation error or missing alignment.

In general, it can be concluded that for this setup proof of high accuracy is challenging and the accuracy of the system might be better than the data show, but at the same time this is not the presented work.

Other sources of error are wrong calibration values for the relative position of the detected object Po (Formula (2))) and simplifications of Formula (2), an incorrectly measured height, a wrong scaling factor for the optical flow and bad lighting and surface conditions, which lead to position errors measured by the optical flow sensor.

The current orientation of the quadrocopter is not considered in the computation of the position *P*T. This was intended because the effect of an orientation error should be excluded from the evaluation. In some cases this led to double detection errors.
