Experimental comparison of AdaBoost algorithms applied on leg detection with different range sensor setups

Abstract

When tracking people or other moving objects with a mobile robot, detection is the first and most critical step. At first most researchers focused on the tracking algorithms, but recently AdaBoost (supervised machine learning technique) was used for people legs detection in 2D range data. The results are promising, but it is unclear if the obtained classifier could be used on the data from another sensor. As it would be a huge inconvenience having to train a classifier for every sensor (setup), we set out to find if, and when is a classifier trained on one sensor setup transferable to another sensor setup. We tested two sensors in five different setups. In total, we acquired 2455 range scans. Experiments showed that the classifier trained on noisier sensor data performed better at classification of data coming from other sensor setups. Classifiers trained on less noisy data were shown to be overconfident, and performed poorly on noisy data. Furthermore, experiments showed that classifiers learned on ten times smaller datasets performed as good as classifiers trained on larger datasets. Since AdaBoost is a supervised learning technique, obtaining same classifier efficiency with significantly smaller dataset means less hand labeling of the data for the same results.