A full-3D voxel-based dynamic obstacle detection for urban scenario using stereo vision

Abstract

Abstract-Autonomous Ground Vehicles designed for dynamic environments require a reliable perception of the real world, in terms of obstacle presence, position and speed. In this paper we present a flexible technique to build, in real time, a dense voxel-based map from a 3D point cloud, able to: 1) discriminate between stationary and moving obstacles; 2) provide an approximation of the detected obstacle's absolute speed using the information of the vehicle's egomotion computed through a visual odometry approach. The point cloud is first sampled into a full 3D map based on voxels to preserve the tridimensional information; egomotion information allows computational efficiency in voxels creation; then voxels are processed using a flood fill approach to segment them into a clusters structure; finally, with the egomotion information, the obtained clusters are labeled as stationary or moving obstacles, and an estimation of their speed is provided.The algorithm runs in real time; it has been tested on one of VisLab's AGVs using a modified SGM-based stereo system as 3D data source.

Perception, Planning, Control, and Coordination for Autonomous Vehicles

“…In [24,27,32,49,50], the authors implemented the RANSAC algorithm to segment the ground plane in the point cloud with the assumption of flat surface. However, as mentioned in [23,51], for non-planar surfaces, such as undulated roads, uphill, downhill, and humps, this model fitting method is not adequate.…”

Bayesian environment representation, prediction, and criticality assessment for driver assistance systems

“…(g) (h) (i) Figure 2.1: Environment representations of various levels of abstractness from most abstract (a) to most general (i): feature map [160] (a), interval map [267] (b), 2D grid map [223] (c), elevation map [61] (d), stixel world [194] (e), multi-level surface map [249] (f), voxel grid [47] (g), mesh [49] (h), raw sensor data [193] (i). Images taken from the mentioned sources.…”

“…Methods for constructing voxel grids can be subdivided into i) truly probabilistic ones, which have a 3D occupancy grid map as a result [108,175,213], and ii) methods that just employ a voxel grid to coarsen the level of detail of 3D point clouds. The latter is, for example, used in [47] as an intermediate step for stereo vision-based obstacle detection and in [123] for the detection of dynamic objects based on 3D warping. Full probabilistic 3D grid maps are computationally expensive to construct and require high bandwidth for transmission even if memory-efficient, hierarchical data structures such as octrees are employed as in [108,213].…”

“…This independence is a consequence of the independence assumption of all vehicle trajectories, so that the probability of the ego vehicle colliding with one object is not influenced by the probability of colliding with another. If this assumption is omitted, p C k (T P ) must be calculated according to 47) with general indicator function…”

Dense Stereo-Based Robust Vertical Road Profile Estimation Using Hough Transform and Dynamic Programming

“…This map is useful for terrain with a single surface layer, but cannot deal with multiple surface layers such as bridges, tunnels, or underpasses. A 3-D grid map is able to overcome this problem by representing terrain using stacks of cubes with occupancy information [18], [19]. This map is often called a voxel map and can be efficiently described by a hierarchical data structure for spatial subdivision such as an octree [20].…”