Detection of Road Limits Using Gradients of the Accumulated Point Cloud Density

Detection of road curbs and berms is a critical concern for autonomous vehicles and driving assistance systems. The approach proposed in this paper to detect them uses a 4-layer LIDAR placed near to the ground to capture measurements from the road ahead of the car. This arrangement provides a particular point of view that allows the accumulation of points on vertical surfaces on the road as the car moves. Consequently, the point density increases in vertical surfaces and stays limited in horizontal surfaces. A first analysis of the point density allows to distinguish curbs from flat roads, and a second solution based on the gradient of point density not only detects curbs as well but also detects berms due to the transitions of the gradient density. To ease and improve the processing speed, point clouds are flattened to 2D and traditional computer vision gradient and edge detection techniques are used to extract the road limits for a wide range of car velocities. The results were obtained on the ATLASCAR real system, and they show good performance when compared to a manually obtained ground truth.