View Based Navigation Exploiting Temporal Information
213
Fig. 6. a) Runtime for PM (y axis) versus SMW (x axis). For the timing we used
medium resolution (180, 50) and blur, IDF and the window size 17. SMW is about
10 times faster than PM. b) Runtime and mean error for SMW using IDF or CC.
Point represents data for window sizes from 6 through to 20. IDF is much faster but
generates larger errors. For CC there is an optimal compromise for the error versus
runtime trade-off at window size around 10.
4 Discussion
In this paper we have shown that we can both increase the accuracy of navigation, and reduce the computational cost, by including temporal sequence information in the stored route from training. By focusing on the areas where the
algorithm fails which are typically caused by partial occlusion due to vegetation,
we showed that performance could be further improved by using correlationbased image comparisons as well as edge enhancement. While changing resolution had a relatively minor effect in terms of improving the best combination
of window length and processing parameters, it is clear that there will be an
optimal resolution for a given environment, at the cost of increased algorithmic
complexity. Likewise, finding the optimal window size is likely to be environment
specific and, given that in some situations, accuracy gains are marginal, it might
be that the choice is determined by computing time.
For sequence based navigation models, the length of the comparison window
is important and not all memory window lengths work. Indeed, we saw that
windows that are too small can fail fairly spectacularly. Examining this issue
reveals the weakness of this method. If the algorithm gets ‘lost’ due to an erroneous match, the window can get stuck at that point and never find the correct
matching point for subsequent test points. The smaller the window size and the
more cluttered the environment, then the higher chance of the window getting
’stuck’ and sending the agent in the wrong direction. This is linked to the issue
of at what point to start the algorithm and both points could be solved in a sim-
213
Fig. 6. a) Runtime for PM (y axis) versus SMW (x axis). For the timing we used
medium resolution (180, 50) and blur, IDF and the window size 17. SMW is about
10 times faster than PM. b) Runtime and mean error for SMW using IDF or CC.
Point represents data for window sizes from 6 through to 20. IDF is much faster but
generates larger errors. For CC there is an optimal compromise for the error versus
runtime trade-off at window size around 10.
4 Discussion
In this paper we have shown that we can both increase the accuracy of navigation, and reduce the computational cost, by including temporal sequence information in the stored route from training. By focusing on the areas where the
algorithm fails which are typically caused by partial occlusion due to vegetation,
we showed that performance could be further improved by using correlationbased image comparisons as well as edge enhancement. While changing resolution had a relatively minor effect in terms of improving the best combination
of window length and processing parameters, it is clear that there will be an
optimal resolution for a given environment, at the cost of increased algorithmic
complexity. Likewise, finding the optimal window size is likely to be environment
specific and, given that in some situations, accuracy gains are marginal, it might
be that the choice is determined by computing time.
For sequence based navigation models, the length of the comparison window
is important and not all memory window lengths work. Indeed, we saw that
windows that are too small can fail fairly spectacularly. Examining this issue
reveals the weakness of this method. If the algorithm gets ‘lost’ due to an erroneous match, the window can get stuck at that point and never find the correct
matching point for subsequent test points. The smaller the window size and the
more cluttered the environment, then the higher chance of the window getting
’stuck’ and sending the agent in the wrong direction. This is linked to the issue
of at what point to start the algorithm and both points could be solved in a sim-