¿Tiene seguro actualmente?
OFRECEN LAS COMPAÑÍAS ASEGURADORAS
The apparent increase of the computational effort that would suppose the utilization of the presented approach could be hard to justify within the field of filter based SLAM, which tries to keep reduced computational costs. But the cost increase is bounded and could be further reduced. For our Cs sequence set, made of a total of 7120 frames in all sequences, only 38.22 % (2721 frames) presented field of view overlap with the Cf camera. While this overlap ratio supposed an overhead of processing almost 40% more images, the exploration area was reduced with the search of the corespondence ROI. It is also interesting how the newly proposed approach made less effort per feature to be initialized in terms of number of frames requiring it to be tracked, compensating the larger number of features used.
TABLE 5.2:STATISTICS OF FEATURES USED AND TRACKING DUE DELAYED INITIALIZATION FOR ORIGINAL DI-DMONOCULAR AND FOR MULTIPLE VIEW APPROACH
Metric DI-D Monocular Multiple view
Features initialized (total) 1487 1549 Features initialized (avg.) 297.4 309.8
Average tracking period 24.6 10.4
TABLE 5.2 shows the features used on each feature initialization approach, and the tracking
effort required (measured in number of frames where the feature is tracked) until the initialization of the features. For the experimental set, the multiple view approach uses about 4% more features, but the time required to initialize them is smaller. This is because many features that are being tracked are instantly initialized through the multiple view method once they lay in the overlapped field of view. This is advantageous because it allows introducing features known to be strong (enough to be tracked) directly, avoiding the computational costs of tracking them, offsetting the additional costs introduced by by the multiple view approach.
Furthermore, in real-time applications employing this technique, the Cf sensor could be upgraded to an intelligent sensor, i.e., presenting processing capabilities. This approach
would integrate the image processing in the Cf sensor, allowing parallel processing of SURF features, and sending only extracted features, minimizing communications delay. This processing step could be done while the robotic camera Cs makes the general EKF- SLAM process, and thus it would be possible to have the SURF landmarks information after the EKF update, in time for the possible inclusion of new features.
5.7 Conclusions
In this chapter a new approach for feature initialization in SLAM has been proposed and discussed. A multiple view depth estimation procedure is proposed for feature initialization
under a collaborative sensing assumption. The approach is based on the DI-D technique discussed in Chapter 3 (Munguia and Grau, 2012) and expanded in Chapter 4 (Guerra et al., 2013), being heavily focused towards human-robot interaction frameworks, under the form of collaborative explorations of the environment. The human collaboration has been introduced through a monocular sensor with total freedom of movement and approximately known pose, which is a set of assumptions generally satisfied in collaborative SAR robotics. As the different monocular sensors move freely, sometimes their fields of view will be concurrent: both cameras observing the same elements of the environment, producing non-constant multiple view measurements of them. As the relative pose between the cameras and the calibration matrices of each one of them are known, the fundamental matrix of a stereo system could be found. Even though this would allow the utilization of stereo-based rectification to ease the correspondence problem, it was deemed inconvenient for the approach, and descriptor-based feature matching was considered a better option. Utilization of non-constant multiple view depth estimation allows improving the performance of two specific aspects in the local scale EKF-SLAM framework. Firstly, the requirement of an initial metric scale introduced through synthetic features can be removed, substituted by the initialization of a set of features with collaborative depth estimation. This depth estimation has proven to have a multiple advantages: the number of features introduced initially is not limited to four coplanar points; and the use of a larger number of features presenting diverse depth values makes the metric scale propagation smoother. Secondly, the introduction of later landmarks through multiple view depth estimation enables utilization of far distance features with real depth estimation, instead of the heuristically assigned value used in previous works, and the initialization of frontal landmarks when the camera Cs moves forward and other singular trajectories. These changes have produced a locally strong and robust SLAM approach, thus enabling its future utilization on larger scale SLAM, as commented on section 2.7.3. Using the proposed approach in an SLAM framework considering loop closure and large map management would further reduce the drift of the estimated trajectory, thanks to the covariance reduction produced by loop closure.
As the viability of the proposed approach has been demonstrated, research could focus on maximizing the advantages obtained from the HRI, while studying in depth the costs of the proposed technique. In terms of exploiting the HRI, the multiple view depth estimation could be introduced the measurement and update step of the EKF SLAM. This would probably require a general overhaul of the prediction and observation models currently used, but it should improve the accuracy of the approach. In line with this overhaul, the use of non-constant stereo allows to reinitialize a metric scale whenever the field of view overlaps, permitting the introduction of submapping techniques and other methods related with large map management to and achieve larger trajectories, including loop closing.
The proposed technique could be also expanded, with some techniques taken from the collaborative SLAM field, to deal with more Cf agents, e.g.: a group of different humans could explore an environment accompanied by a robot mapping their surroundings with data from the sensors deployed on the humans. While this approach would require much more computational power and an insightful architecture, it would be of great interest due its resemblance to hypothetical real cases where not a human alone, but a team, would explore new zones with robotic assistance.
Equation Chapter 6 Section 1