Laser Radar 3D Target Recognition by Integrating Point Distribution Feature Descriptors and ISS Algorithm

Abstract

This study proposed a laser radar 3D target recognition method that integrates point distribution feature descriptors and intrinsic shape feature algorithms, aiming to solve the problems of insufficient robustness and low computational efficiency in point cloud recognition in complex scenes. This method constructed a multi-level feature representation with strong invariance to rotation, translation, and noise by combining globally aligned spatial distribution descriptors and point distribution feature descriptors. It utilized the intrinsic shape feature algorithm to stably extract feature points, and introduces improved region growth segmentation and Iterative Closest Point (ICP) registration algorithms to enhance recognition performance. Experiments on two public datasets showed that the research method achieved feature extraction accuracies of 93.68% and 91.46%, both of which were superior to existing mainstream descriptors. The proposed descriptor exhibited superior robustness with an SNR of 19.46 dB and an occlusion rate of 82.75% in noisy environments. In addition, the average feature extraction time was 52.37 ms, which improved efficiency by over 47% compared to methods such as rotational projection statistical feature extraction. In two real-world datasets, the proposed method achieved recognition rates of 89.86% and 80.72% when the noise deviation was 0.8 m, and 88.76% and 80.78% under low resolution conditions. Its recognition performance outperformed the compared methods. The results indicate that the research method has excellent performance in accuracy, robustness, and efficiency, providing an effective technical means for 3D object recognition in complex environments. It also has strong practical application value.