Problem denition Experiments Fusion approachSegmentation of moving objects Conclusions References

We present a method to segment dynamic objects on point clouds using images and 3D laser data. Per-pixel background classes are adapted online as Gaussian Mixtures independently for each sensor. The learned classes are fused labeling pixels/voxels that belong to either the background, or the dynamic objects We pay special attention in the calibration and synchronization modules to reach accuracy in registration and data association. We show results of people segmentation in indoor scenes using a Velodyne sensor at a high frame-rate. We address the problem of eliminating moving bodies from high acquisition rate 3D range scans using image data. This work is an extension of the work presented in [1]. Range points are classified according to temporally consistent corresponding images. Per pixel class distributions are learned on-line as Gaussian Mixture [2]. We use a fusion approach for range and intensity images inspired by the adaptive Mixture of Local Experts (MLE) architecture[4]. Sensor calibration Sensor synchronization We present a method to segment dynamic objects from 3D poin