Tracking objects in 6D for reconstructing static scenes

Abstract

This paper focuses on two aspects of a human robot interaction scenario: Detection and tracking of moving objects, e.g., persons is necessary for localizing possible interaction partners and reconstruction of the surroundings can be used for navigation purposes and room categorization. Although these processes can be addressed independent from each other, we show that using the available data in exchange enables a more exact reconstruction of the static scene. A 6D data representation consisting of 3D Time-ofFlight (ToF) Sensor data and computed 3D velocities allows segmenting the scene into clusters with consistent velocities. A weak object model is applied to localize and track objects within a particle filter framework. As a consequence, points emerging from moving objects can be neglected during reconstruction. Experiments demonstrate enhanced reconstruction results in comparison to pure bottom-up methods, especially for very short image sequences.

ToF cameras for active vision in robotics

“…Swadzba et al [22] present a scene reconstruction algorithm that discards dynamic objects, like pedestrians, using a static camera in the difficult case of short sequences (2-3 sec.). Motion is recovered via optical flow in the intensity images, and then transferred to the depth image to compute a 3D velocity vector.…”

“…3D at high rate SR2 (depth) May et al [11,12] 3D mapping 3D at high rate/No required Pan-Tilt SR2 (depth) May et al [13] Pose estimation/3D mapping Registered depth-intensity SR3 (depth + intensity) Hedge and Ye [14] Planar feature 3D mapping 3D at high rate/No required Pan-Tilt SR3 Ohno et al [16] 3D mapping 3D at high rate SR2 Stipes et al [17] 3D mapping / Point selection Registered depth-intensity SR3 May et al [18] 3D mapping/SLAM 3D at high rate SR3 Gemeiner et al [19] Corner filtering Registered depth-intensity SR3 (depth + intensity) Thielemann et al [20] Navigation in pipelines 3D allow geometric primitives search SR3 Sheh et al [21] Navigation in hard env. 3D at high rate SR3 + inertial Swadzba et al [22] 3D mapping in dynamic env. 3D at high rate/Registered depth-intensity SR3 (depth + intensity) Acharya et al [23] Safe car parking Improved depth range/3D at high rate Canesta Gallo et al [24] Gortuk et al [25] Object classification (airbag app.)…”

Dynamic 3D scene analysis for acquiring articulated scene models

“…It is also compared to the neglecting of moving pixels M MPIX and last, to M TRACK [12] where only dynamic objects are determined through tracking without background model feedback and no distinction is made between static background and static movable objects. All methods are checked against a ground truth static scene model M GT , which has been taken without any movable object for each sequence.…”

“…Compared to stereo rigs the 3D ToF sensors can deal much better with prominent parts of rooms like walls, floors, and ceilings even if they are not textured. In addition to the 3D point cloud, contour and flow detection in the image plane yields motion information that can be used, e.g., for person tracking [12], [13].…”

Capturing Time-of-Flight data with confidence

“…Time-of-Flight (ToF) cameras have been successfully used for a variety of applications such as Simultaneous Localisation and Mapping (SLAM) [22,25], 3D reconstruction of static scenes [10,34], and object tracking and scene analysis [27,33]. In contrast to stereo vision and triangulation-based scanners, the ToF camera operates from a single viewpoint and does not rely on matching of corresponding features, which greatly increases its robustness in the presence of traditionally difficult scene materials and internal occlusions.…”