A Hybrid Vehicle Extraction Approach from Low-Quality LiDAR Data Based on Robust One-Class Support Vector Machine

Vehicle extraction becomes possible as the high-performance airborne light detection and ranging (LiDAR) systems can offer very dense and accurate point cloud, which means the sophisticated objects can be recorded in detail, combined with color information from airborne image, hyperspectral and intensity. However, few studies have investigated in extracting vehicles from LiDAR data only, especially when its quality is low, which is the main difficulty for most LiDAR applications. In this paper, a hybrid approach has been proposed to extract vehicles from low-quality LiDAR data. In order to extract vehicle from low-resolution LiDAR data, a robust one-class support vector machine-minimum covariance determinant (OCSVM-MCD) is proposed based on a multivariate dispersion estimator and weighted strategy. Firstly, the three-dimensional (3D) point dataset is classified into nonterrain and terrain points with progressive morphological filter with a slight improvement. Secondly, nonterrain points are segmented by clustering technique and missing blobs are searched from terrain points. Then, the vehicles are extracted from clustering and searching results by OCSVM-MCD, and a hybrid principle is put forward to improve the extraction result at last. The proposed method has been evaluated with two benchmark datasets from ISPRS, and proved that by the method, most vehicles can be extracted from low-quality LiDAR data with an encouraging result.