Experimental Evaluation of a People Detection Algorithm in Dynamic Environments

People detection is an important capability both for human-robot interaction in service robotics and to dis- tinguish the stable environment from the perturbation due to people motion in localization and mapping tasks. Several techniques have been proposed for different application contexts and sensors. Range data acquired by laser scanners are met- rically accurate and suitable for computationally-inexpensive people detection. Furthermore, laser scans provide a geometric description of local environment that can be combined with the information about dynamic objects. In this paper, a previously proposed method for detecting people legs from laser scans is experimentally evaluated and exploited to improve scan matching by removing dynamic parts corresponding to people. This algorithm splits laser scans into beam segments and classifies each segment. Classifications of simple features are then combined into a boosted classifier with Adaboost. The fundamental assumption of scan matching is that consecutive scans can be aligned with a rigid body transformation, since they represent the same scene. When dynamic elements like human legs are removed from scans, such assumption holds. We also investigate the effectiveness of the proposed people detection algorithm in terms of its ability to generalize across different environments and to support track persistency across scans.