Identifying landmark cues with LIDAR laser scanner data taken from multiple viewpoints

In this paper, we report on our ongoing efforts to build a cue identifier for mobile robot navigation using a simple one-plane LIDAR laser scanner and machine learning techniques. We used simulated scans of environmental cues to which we applied various levels of Gaussian distortion to test a number of models the effectiveness of training and the response to noise in input data. We concluded that in contrast to back propagation neural networks, SVM-based models are very well suited for classifying cues, even with substantial Gaussian noise, while still preserving efficiency of training even with relatively large data sets. Unfortunately, models trained with data representing just one stationary point of view of a cue are inaccurate when tested on data representing different points of view of the cue. Although the models are resilient to noisy data coming from the vicinity of the original point of view used in training, data that originates in a point of view shifted forward or backward (as would be the case with a mobile robot) proved much more difficult to classify correctly. In the research reported here, we used an expanded set of synthetic training data representing three view points corresponding to three positions in robot movement in relation to the location of the cues. We show that by using the expanded data the accuracy of cue classification is dramatically increased for test data coming from any of the points.