|
A line scanning ladar can generate detailed three-dimensional images of a scene, so-called point clouds, by emitting individual laser pulses in quick succession in various directions and measuring the time before arrival of return pulses. As a typical mode of operation, the pulses are emitted along horizontal lines, starting from bottom of the field of view, before gradually increasing the elevation angles of subsequent scanning lines. This paper aims to address an inherent problem with object recognition within point clouds acquired by a line scanning ladar. If some of the scene objects are moving, their position will change slightly between each sweep of a horizontal scanning line. This causes the shape of the moving objects to deform in the resulting point cloud. The problem becomes more severe for wide view angles, slow scan speeds and fast moving objects. An object recognition algorithm is proposed that corrects for shape deformations caused by the delay between individual point measurements. In addition, the algorithm is able to estimate the velocity of the recognized object. The algorithm matches observed objects against a 3D model of the object of interest, by optimally aligning them with each other while simultaneously estimating the optimal shape deformation caused by motion during acquisition. If the observed object and 3D model aligns sufficiently well, according to a certain recognition confidence measure, the observed object is regarded as recognized and its velocity is induced from the estimated shape deformation. To solve the underlying optimization problem, the ”Iterative Closest Point” (ICP) algorithm is modified by incorporating an additional substep, where the shape deformation – and thereby the corresponding velocity - is updated incrementally each iteration. Experiments on simulated and real world data indicate that moving objects can be recognized with high confidence and their velocities can be estimated with high accuracy.
|
