Robust video-based object recognition integrating highly redundant cues for indexing and verification

Abstract

The presented approach integrates proven techniques and original approaches to one robust and fast 3D model-based recognition system. A speedy recognition is achieved by operating on a stream of filtered 3D sensor features, reconstructed for navigation tasks of the robot, instead of using a separate sensor data processing. Furthermore, simple, inexpensive recognition strategies are applied. Robustness is obtained by integrating complementary recognition strategies: four indexing techniques and two (2D and 3D) matching methods for verification, are completed by a hypothesis promotion, based on feedback of information to the sensor system. All strategies differ in their requirements, reliability, selectivity, and temporal constraints. Hypotheses are integrated using fusion, ruling out, and aging techniques. The approach is evaluated experimentally with varying calibration errors, scene complexity, and sensing conditions.