A parts-based method for articulated target recognition in laser radar data

Target recognition in laser radar data is a new and thriving research topic in the area of automatic target recognition (ATR). In this paper, we present a parts-based method for recognizing highly similar articulated ground vehicles. Based on the fact that man-made vehicles are well-structured and consist of a set of planar surfaces, the relationship between the distribution of projected points and the pose of target is revealed and a measure named projection density entropy (PDE) is introduced. Using PDE, we propose a pose estimation method for rigid object, we also introduce a method for target decomposition and part pose estimation for articulated target. Further, we develop two methods for articulated target recognition, i.e., canonical matching based method and fusion based method. Experiments on a dataset containing 1536 views of 16 vehicles show that our proposed PDE method outperforms the existing methods for pose estimation, with high robustness to occlusion and noise. Good results have also been reported for target recognition, with a recognition rate over 99% achieved.