Extraction and classification of vehicles in LADAR imagery Hans Christian Palm*, Trym Vegard Haavardsholm, Halvor Ajer, Cathrine Vembre Jensen Norwegian Defence Research Establishment (FFI), P.O. Box 25, N-2027 Kjeller, Norway ABSTRACT The work presented in this paper is based on a dataset recorded with an airborne sensor. It comprises targets like M-60, M-47, M-113, bridge layers, tank retrievers, and trucks in various types of scenes. The background-object segmentation consists of first estimating the ground level everywhere in the scene, and then for each sample simply subtracting the measured height and ground level height. No assumptions concerning flat terrain etc. are made. Samples with height above ground level higher than a certain threshold are clustered by utilizing a straightforward agglomerative clustering algorithm. Around each cluster the bounding box with minimum volume is determined. Based on these bounding boxes, too small as well as too large clusters can easily be removed. However, vehicle-sized clutter will not be removed. Clutter detection is based on estimating the normal vector for a plane approximation around each sample. This approach is based on the fact that the surface normals of a vehicle is more “modulo 90 o ” distributed than clutter. The aim of the classification has been to classify main battle tanks (MBTs) Two types of algorithms have been studied, one based on Dempster Shafer fusion theory, and one model based. Our dataset comprises clusters of 269 vehicles (among them 131 MBTs), and 253 clutter objects (i.e. in practice vehicle- sized bushes). The experiments we have carried out show that the segmentation extracts all vehicles, the clutter detection removes 90% of the clutter, and the classification finds more than 95% of the MBTs as well as removes half of the remaining clutter. Keywords: Outlier rejection, segmentation, detection, clutter rejection, classification, ladar 1. INTRODUCTION LADAR imagery has shown some great capabilities for automatic target recognition (ATR) [1,2,3,4,5,6,7]. The 3D LADAR images provide a lot more information than conventional 2D images, and this can be really useful in many app- lications. E.g. the reliability of detecting land vehicles on a relatively long distance is much higher using a 3D point cloud than using a 2D set of pixels. Thus, LADAR imagery has a large potential in a future battlefield. The most important LADAR data product is a range image that may be seen as an unstructured point cloud in 3D space. Thus, algorithms for analyzing such images can easily be computationally quite heavy. Our aim has been to investigate what can be achieved by using computationally relatively simple algorithms; algorithms which have potential for real- time processing on a portable platform but nevertheless give reliable output. All parts of the processing chain, from pre- processing to classification, have been addressed, and they will be described in section 3. In section 4 we will present results from experiments we have carried out. Finally, we give a summary and conclusion in section 5. However, before we start presenting the algorithms, we will very briefly present the dataset we have used. *Corresponding author: hans-chr.palm @ffi.no; phone +47 63 80 70 00; fax +47 63 80 72 65 Laser Radar Technology and Applications XVIII, edited by Monte D. Turner, Gary W. Kamerman, Proc. of SPIE Vol. 8731, 873102 · © 2013 SPIE · CCC code: 0277-786X/13/$18 · doi: 10.1117/12.2015363 Proc. of SPIE Vol. 8731 873102-1