Obstacle Detection Using a Facet-Based Representation from 3-D LiDAR Measurements

Abstract

In this paper, we propose an obstacle detection approach that uses a facet-based obstacle representation. The approach has three main steps: ground point detection, clustering of obstacle points, and facet extraction. Measurements from a 64-layer LiDAR are used as input. First, ground points are detected and eliminated in order to select obstacle points and create object instances. To determine the objects, obstacle points are grouped using a channel-based clustering approach. For each object instance, its contour is extracted and, using an RANSAC-based approach, the obstacle facets are selected. For each processing stage, optimizations are proposed in order to obtain a better runtime. For the evaluation, we compare our proposed approach with an existing approach, using the KITTI benchmark dataset. The proposed approach has similar or better results for some obstacle categories but a lower computational complexity.

Keywords: LiDAR point cloud; facet representation; object contour; obstacle detection.

Figure 1

Potential application: (a–c): Facet-based representation used in automatic emergency braking situations (top view visualization). (a): The gray car detects the black car (red bounding box). (b): The gray car brakes to the rear of the black vehicle with the door open (if the car is detected as the red bounding box). (c): The gray car can perform a smoother braking if the black car is represented by its visible facets (each facet with a different color). (d–f): A large articulated vehicle cannot be modelled as an oriented cuboid during cornering maneuvers.

Figure 2

System architecture.

Figure 3

Channel values representation in a point cloud (top view). The red area is where the value for X is bigger than the value for Y for the same point; the white area is where the value for X is smaller than the value for Y.

Figure 4

Side view of a channel. Detected ground points are colored with gray. X/Y means either the x-axis or the y-axis value is used, depending on the specific channel orientation.

Figure 5

Ground detection. (a): Results shown in 3-D view. (b): Results overlay over the corresponding camera image.

Figure 6

(a): Object primitive clusters in a channel. (b): Boundary points of a cluster shown in red.

Figure 7

The creation of the intra-channel clusters for a van is presented. Each cluster is represented by a rectangle (delimited by the most extreme points—red dots) in its channel side view. The clusters have different sizes and are intersecting or are close to the previous channel clusters.

Figure 8

Clustering results. (a): Object clusters in point cloud, with distinct color per cluster. (b): Detected clusters projected on the corresponding camera image.

Figure 8

Clustering results. (a): Object clusters in point cloud, with distinct color per cluster. (b): Detected clusters projected on the corresponding camera image.

Figure 9

(a): Top view of a van. (b): Contour obtained from clustering.

Figure 10

Runtime comparison graph for ground detection methods on 252 scenes.

Figure 11

Runtime comparison graph for clustering methods on 252 scenes.

Figure 12

Close multiple objects clustered as one single object. (a): Image with close multiple objects. (b): Single cluster created—point cloud view (same label for all the points).

Figure 13

Facet detection on a van, tram, cyclist, pedestrian, buildings, and a wall. (a): Objects from the point cloud. (b): Contour of object (top view). (c): Facets (with red) over contour. (d): 3-D facets over objects.

Figure 14

Circular fences. (a): Top view. (b): Perspective view. The detected facets are displayed (white).

Figure 15

Facets extracted from the KITTI bounding box (red) and 3-D output facets (yellow) from our implementation are paired for comparison.

Figure 16

Projection of facets (orange and blue) on the KITTI-extracted face (with red).

Figure 17

Runtime comparison graph for facet detection methods on 252 scenes.

Figure 18

Facet representation on a complex scene.