Performance evaluation of 3D descriptors for object recognition in construction applications

3D object recognition from field-acquired point cloud data is important for modeling, manipulation, visualization and other post-processing tasks in the construction domain. However, building semantically-rich models from raw point cloud data is a difficult task due to the high volume of unstructured information as well as confounding factors such as noise and occlusion. Although there exist several computational recognition methods available, their performance robustness for construction applications are not well known. Therefore, this research aims to review and evaluate state-of-the-art descriptors for 3D object recognition from raw point clouds for construction applications such as workspace modeling, asset management and worker tracking. The evaluation was carried out using 3D CAD models with known labels as training data and laser-scanned point clouds from construction sites as testing data. The recognition performance was evaluated with respect to varying level of detail, noise level, degree of occlusion, and computation time. Experimental results show that for all evaluated descriptors, increasing the level of detail and decreasing the noise level results in a moderate increase in recognition accuracy whereas reducing occlusion results in a significant increase in recognition accuracy. In addition, experimental results suggest that the key features that distinguish an object can be derived around the 10 mm level and any further increase in the level of detail do not significantly increase the recognition accuracy.